Philosophical Essays. Volume 2 Philosophical Essays, Volume 2: The Philosophical Significance of Language [Course Book ed.] 9781400833184

Citation preview

Philosophical Essays

This page intentionally left blank

PHILOSOPHICAL ESSAYS Volume 2 the philosophical significance of language Scott Soames

princeton university press p r i n c e t o n a n d ox f o r d

Copyright © 2009 by Princeton University Press Published by Princeton University Press, 41 William Street, Princeton, New Jersey 08540 In the United Kingdom: Princeton University Press, 6 Oxford Street, Woodstock, Oxfordshire OX20 1TW All Rights Reserved Library of Congress Cataloging-in-Publication Data Soames, Scott. Philosophical essays : the philosophical significance of language / Scott Soames. p. cm. Includes bibliographical references and index. ISBN 978-0-691-13682-0 (v. 2 : hardcover : alk. paper) — ISBN 978-0-691-13683-7 (v. 2. : pbk. : alk. paper) 1. Language and languages—Philosophy. 2. Linguistics. 3. Semantics. I. Title. P107.S67 2009 410.9—dc22 2008019492 British Library Cataloging-in-Publication Data is available This book has been composed in Sabon Printed on acid-free paper. ∞ press.princeton.edu Printed in the United States of America 10

9 8 7 6 5 4 3 2 1

THESE VOLUMES ARE DEDICATED TO THE PEOPLE I LOVE

My wife, Martha My sons, Greg and Brian and to the memory of my parents Bill and Ruth Soames

This page intentionally left blank

Contents

The Origins of These Essays

ix

Introduction

1

Part One Reference, Propositions, and Propositional Attitudes

31

Essay One Direct Reference, Propositional Attitudes, and Semantic Content

33

Essay Two Why Propositions Can’t Be Sets of Truth-Supporting Circumstances

72

Essay Three Belief and Mental Representation

81

Essay Four Attitudes and Anaphora

111

Part Two Modality

137

Essay Five The Modal Argument: Wide Scope and Rigidified Descriptions

139

Essay Six The Philosophical Significance of the Kripkean Necessary A Posteriori

165

Essay Seven Knowledge of Manifest Natural Kinds

189

Essay Eight Understanding Assertion

211

Essay Nine Ambitious Two-Dimensionalism

243

Essay Ten Actually

277

viii • Contents

Part Three Truth and Vagueness

301

Essay Eleven What Is a Theory of Truth?

303

Essay Twelve Understanding Deflationism

323

Essay Thirteen Higher-Order Vagueness for Partially Defined Predicates

340

Essay Fourteen The Possibility of Partial Definition

362

Part Four Kripke, Wittgenstein, and Following a Rule

383

Essay Fifteen Skepticism about Meaning: Indeterminacy, Normativity, and the Rule-Following Paradox

385

Essay Sixteen Facts, Truth Conditions, and the Skeptical Solution to the Rule-Following Paradox

416

Index

457

The Origins of These Essays

Essay One

Philosophical Topics 15 (1987): 47–87. Reprinted by permission of the University of Arkansas Press.

Essay Two

Journal of Philosophical Logic 37 (2008): 267–76. Reprinted with the kind permission of Springer Science and Business Media.

Essay Three

Philip P. Hanson, ed. Information, Language, and Cognition. Vancouver: University of British Columbia Press, 1990. 217–46. Reprinted by permission of Oxford University Press.

Essay Four

Philosophical Perspectives 8 (1994): 251–72. Reprinted by permission of Blackwell Publishing.

Essay Five

Noûs 32 (1998): 1–22. Reprinted by permission of Blackwell Publishing.

Essay Six

Philosophical Issues 16 (2006): 288–309. Reprinted by permission of Blackwell Publishing.

Essay Seven

Facta Philosophica 6 (2004): 159–81. Reprinted by permission of Peter Lang Publishing.

Essay Eight

Judith Thompson and Alex Byrne, eds. Content and Modality: Themes from the Philosophy of Robert Stalnaker. New York: Oxford University Press, 2006. 222–50. Reprinted by permission of Oxford University Press.

Essay Nine

Matthew Davidson, ed. On Sense and Direct Reference: Readings in the Philosophy of Language. Boston: McGraw-Hill, 2007. 690–718.

Essay Ten

Aristotelian Society Supplementary Volume 81 (2007): 251–77. Reprinted by permission of Blackwell Publishing.

Essay Eleven

Journal of Philosophy 81 (1984): 411–29. Reprinted by permission of the Journal of Philosophy.

Essay Twelve

Philosophical Perspectives 17 (2003): 369–83. Reprinted by permission of Blackwell Publishing.

x • The Origins of These Essays

Essay Thirteen

JC Beall, ed. Liars and Heaps: New Essays on Paradox. Oxford: Clarendon Press; New York: Oxford University Press, 2003. 128–50. Reprinted by permission of Oxford University Press.

Essay Fourteen Richard Dietz and Sebastiano Moruzzi, eds. Cuts and Clouds: Essays on the Nature of Vagueness. Oxford: Oxford University Press, forthcoming. Reprinted by permission of Oxford University Press. Essay Fifteen

Ali A. Kazmi, ed. Meaning and Reference. Canadian Journal of Philosophy, Supplementary vol. 23 (1997): 211–49. Used with the permission of the University of Calgary Press.

Essay Sixteen

Philosophical Perspectives 12 (1998): 313–48. Reprinted by permission of Blackwell Publishing.

Philosophical Essays

This page intentionally left blank

Introduction

The essays in this volume are concerned with four main topics— propositions and attitudes, modality, truth and vagueness, and skepticism about intentionality. The significance of these issues extends well beyond the philosophy of language. In addition to being semantically encoded by sentences, propositions are asserted, believed, and known. Questions about what they are, and how we come to believe or know them—as well as questions about which propositions are expressed by which sentences, and which are asserted by which utterances—are crucial to epistemology and the philosophy of mind, as well as being the touchstone of the systematic study of meaning. The next topic, modality, brings together the study of reference and essence, indexicality and actuality, and the distinction between metaphysical and epistemic possibility. Here, the central issues include the role of evidence in our knowledge of the necessary a posteriori, the metaphysical makeup of possible world-states, and the different ways we acquire knowledge of them. The third topic, truth and vagueness, must be covered by any systematic study of language. But that’s not all. In addition to being properties of expressions, truth and vagueness apply to that which these parts of language express, or designate. Sentences are true when the propositions they express are true, and—to the extent that they are vague—it is often not because it is vague which perfectly precise propositions they express, but because the propositions they clearly express are vague. The same is true of predicates, the vagueness of which is often tied to the vagueness of the properties they express. Even singular terms, and the objects they designate, are not exempt. The idea that, apart from language, the world and its objects are pristinely precise is a metaphysical prejudice. The spatial and the temporal boundaries of many things are vague. More generally, clarity about the nature of truth, and an appreciation of how vagueness limits our ability to give precise answers to certain questions, are required in every area of philosophy. Finally, skepticism about intentionality is not just skepticism about meaning, but also skepticism about belief, and mental content. There is no understanding, or rebutting, one without doing the same for the other. In addition to addressing these topics, the essays that follow will, I hope, illustrate the interpenetration of issues in the philosophy of language with those in other core areas of philosophy. It’s not that philosophy of language

2 • Introduction

is “first philosophy,” providing solutions to heretofore intractable problems in other areas. The old view—so prevalent in the mid–twentieth century—that philosophical problems are simply linguistic confusions to be dissolved by a correct understanding of meaning has, mercifully, fallen by the wayside. What exactly has replaced it is, I think, not yet clear. However, whatever it turns out to be, language will, I am confident, remain important. The discipline provided by systematic logical and semantic frameworks, plus explicit pragmatics of assertion and implicature, is too valuable to be neglected by those working in other areas. Of course, influence also flows the other way. Philosophers of language working, as I do, on the necessary a posteriori and the contingent a priori, the semantics of ‘actually’, and skepticism about intentionality must draw upon discussions in other areas. In some cases, I fear. I have not done enough of this. Still, I hope that these essays, however imperfect, illustrate how the philosophical investigation of language may influence, and be influenced by, investigations in other core areas.

Part 1: Reference, Propositions, and Propositional Attitudes The aim of the essay 1, “Direct Reference, Propositional Attitudes, and Semantic Content” (1987), is to pull apart two conceptions of semantics. According to the first, the central task of semantics is to tell us “what sentences say” (relative to contexts in which they are used).1 According to the second, the task is to specify the truth conditions of sentences (relative to contexts). On the first conception, the meaning of S is a function from contexts to “what is said by S” in those contexts. On the second, it is a function from contexts to “the truth conditions of S,” as used there. When the essay was written, it was common to endorse both conceptions—implicitly embracing the idea that “what is said by S” (in a context C) can be identified with the truth conditions of S (in C). Against this, I argued that there is no reasonable understanding of truth conditions on which this identification is sustainable. Sentences do 1 The phrase “what sentences say” is, of course, a barbarism. Strictly speaking, sentences don’t say (assert) anything; speakers do. The idea behind the barbarism is, roughly, that what a speaker typically asserts by uttering S in C is ( just) the semantic content of S in C. As readers of volume 1 of this collection know, this (overly simple) conception of the relationship between semantic content and assertion is one I don’t accept. At the time the essay appeared, however, it was the dominant, though typically unspoken, conception of that relationship. For present purposes, this is not an issue, since the essay’s argument about what is, and what is not, the semantic content of S, in the sense of what its semantics systematically contributes to what is said by normal utterances of S, is independent of the precise relationship between semantic content and assertion.

Introduction • 3

have semantic contents (relative to contexts), and these contents are systematically related to what speakers standardly assert by uttering them.2 In addition, the semantic contents of many sentences, as well as what they are used to assert, have truth conditions. But neither these contents, nor the asserted propositions, can be identified with sets of truth conditions. For any such set, there are many semantic and assertive contents which, though true in precisely those conditions, differ sharply from one another. To establish this, I formulated an abstract conception of truth conditions in which different conceptions of truth-supporting circumstances (at which sentences are evaluated for truth) can be arranged on a continuum from the coarsest-grained (complete, consistent, metaphysically possible world-states) to the finest-grained (partial, and sometimes inconsistent, abstract situations). Next to be abstracted were rules, specifying the truth conditions of complex sentences in terms of the truth conditional contributions of their parts—for example, rules specifying the circumstances supporting the truth of a conjunction (disjunction) as the intersection (union) of those supporting the truth of the conjuncts (disjuncts). Finally, three assumptions were articulated—one treating A believes/asserts that S as reporting an agent as believing/asserting the semantic content of S (relative to a context and assignment), one taking these attitude verbs to distribute over conjunction, and one treating names, indexicals, and variables as directly referential—that is, as having semantic contents identical with their referents (relative to contexts and assignments). It is shown that if these assumptions are correct, then the semantic content of a sentence cannot be identified with the set of circumstances supporting its truth—no matter how fine-grained these are taken to be. The first examples used to establish this were attitude ascriptions containing proper names in their content clauses. However, since the directly referentiality of names was still contentious, analogous examples involving indexicals, or variables bound from outside the clauses, were shown to work equally well. Since it is hard to deny that at least some of these are directly referential, the strategy of blocking the result by denying direct reference was rejected. A similar fate befell another strategy for avoiding the result—namely, denying that attitude ascriptions report attitudes toward the semantic contents of their complement clauses. Against this, it was argued that any relational semantics of such ascriptions compatible with obvious linguistic facts will require one to distinguish the semantic contents of sentences from sets of circumstances in which they are true. 2 My view of this relationship is set out in chapters 3 and 4 of Soames (2002), and modified in essays 9, 10, 11, and 14 of volume 1 of this collection.

4 • Introduction

With this negative result in place, the essay articulates a positive account of semantic content, and the objects of propositional attitudes. The account divides into two parts—a metaphysical account of the facts in virtue of which attitude ascriptions are true, and a semantic account of the contents of sentences, appropriate to serve as objects of the attitudes. The former sees sentences and mental states as content-bearing vehicles of assertion and belief. The latter identifies semantic contents of sentences with structured, Russellian propositions. The idea is illustrated using a two-stage semantic theory for a simple, first-order language with quantifiers, lambda abstraction, and propositional attitude verbs. Stage 1 is a recursive assignment of structured propositions to formulas relative to contexts and assignments. Stage 2 is a specification of the truth conditions of structured propositions at truth-supporting circumstances. The attitude problems that defeated strict truth-conditional accounts (without structured propositions) are avoided. However, the substitution problems posed by Frege’s puzzle for directly referential terms are not. The suggestions at the end of the essay for defusing these problems should be seen as initial ideas, to be supplemented by later work— including Beyond Rigidity, as well as essays 9, 10, and 11 of volume 1. Essay 2, “Why Propositions Can’t Be Sets of Truth-Supporting Circumstances” (2008), rebuts Walter Edelberg’s (1994) objection to the main negative argument in essay 1. According to that argument, any truthconditional theory (satisfying certain assumptions) that, correctly, makes (1) and (2) true, must, incorrectly, characterize (3) as also being true. (1) Hesperus is Phosphorus. (2) The ancients believed that Hesperus was the brightest body seen in the evening sky and Phosphorus was the brightest body seen in the morning sky. (3) So, the ancients believed that there was an object that was both the brightest body seen in the evening sky and the brightest body seen in the morning sky. Since these theories fail to capture the truth conditions of (1)–(3), they must be rejected. Edelberg objects that the argument rests on a false claim—namely that the truth-conditional theories characterize all instances of (3′) as being model-theoretically entailed by the corresponding instances of (1′) and (2′). (1′) a = b (2′) c believes that (Fa and Gb) (3′) c believes that ∃ x (Fx and Gx) He is right in maintaining that this entailment claim is incorrect. Theories in which truth-supporting circumstances are allowed to be metaphysically

Introduction • 5

impossible can make (1′) and (2′) true, and (3′) false, relative to certain impossible circumstances in which distinct objects a and b are identified. However, he is wrong in taking my argument to make this false claim. Rather, it claims (i) that in order to be correct, a semantic theory must assign Venus as referent of both ‘Hesperus’ and ‘Phosphorus’, while assigning truth conditions to (1) and (2) that make them turn out (actually) to be true, and (ii) that any truth-conditional theory (of the relevant kind) that does this will wrongly characterize the semantic content of (3) as being a truth-conditional consequence of the semantic contents of (1) and (2). In short, Edelberg’s objection confuses model-theoretic consequence with truth-conditional consequence. Let T be a truth-conditional theory of the relevant sort. A sentence S* is a model-theoretic consequence of a set S of sentences, according to T, iff for every model M conforming to T, and every context C and circumstance E of M, if all the sentences in S are true in M with respect to C and E, then so is S*. Let T be a truth-conditional theory of the relevant sort, with intended model M. The semantic content of a sentence (or formula) S*, relative to a context C of M and assignment A of values to variables, is a truth-conditional consequence of the semantic contents of a set S of sentences (or formulas), relative to C and A, iff for every circumstance E of M, if all members of S are true in M with respect to C, A, and E, then so is S*. The upshot is a cautionary tale on the perils of confusing these two, and a discourse on certain respects in which any semantics that aspires to be a theory of meaning must go beyond a theory of the truth of sentences relative to arbitrary models. Essay 3, “Belief and Mental Representation” (1990), explores the way in which a semantic theory of attitude ascriptions incorporating structured propositions as objects of the attitudes, bears on prominent computational models of belief in the philosophy of mind. My target is a series of books and papers in which Jerry Fodor argues that belief is a relation born by an agent to an internal representation in the agent’s “language of thought.” According to Fodor, the (truth conditional) content of the representation specifies the way the world is represented to be, while its syntax is the basis for computational operations responsible for the belief’s causal powers. Fodor’s goal is to vindicate our ordinary belief/desire explanations of behavior, while also providing cognitive psychology with a properly scientific notion of belief. The aim of my essay— which neither endorses nor opposes Fodor’s general program—is to show that his best hope of success lies in incorporating the insights of contemporary semantic treatments of indexicals and propositional attitudes.

6 • Introduction

The early part of the essay explicates Fodor’s position, resolves crucial ambiguities in his formulations, and eliminates their most implausible features. This leads to a reconstruction according to which ‘believe’ expresses a two-place relation that holds between an agent A and a structured proposition p in virtue of A’s standing in a certain cognitive relation to an internal syntactic representation R that expresses p. Since the structure of p encodes much of the syntax of R—which, in turn, is responsible for the causal power of the belief—the representational content of the belief and its causal efficacy are brought into line. Although the alignment is not perfect, it is, I argue, close enough to vindicate our ordinary belief/desire explanations. Misalignment occurs when syntactically identical representations expressing the same proposition contain different directly referential terms. In these cases, what one does may be sensitive not just to the contents of one’s beliefs and desires, but also to the particular representations that are in play. As a result, claims of the form (4) cannot expected to be true, exceptionless, universal generalizations. (4) If x believes that A is an action that x can perform, and if x believes that performing A will bring it about that Q, and if x wants it to be the case that Q, then x will act in a way intended to be a performance of A. Nevertheless, there is no cause for alarm. Though principles like (4) are important parts of our “folk psychology,” they are so only when supplemented by ceteris paribus clauses. Most of the explanations we give of individual actions are perfectly compatible with taking them in this way, and many are counterfactual supporting, despite the fact that (4) may fail in distant possible world-states in which the same beliefs (desires) are held in significantly different ways. As for scientific psychology, the need (sometimes) to include clauses not just about what is believed (desired), but also about how the beliefs (desires) are held, is neither a threat to, nor a radical departure from, our ordinary conception of the mental. If this is right, then my reconstruction of Fodor’s computational model preserves its most plausible features, while avoiding its most implausible ones. Among the latter is his startling claim that “the language of thought” is entirely innate, and that natural language expressions are learned by connecting them with already understood mentalese counterparts. This is no part of my reconstruction, which allows one’s internal system of mental representations to include elements of the natural language one speaks—which are, of course, not innate. Because of this, some beliefs may depend on the contents of internal representations, which in turn may depend on the contents of expressions in a public language. This fits well with linguistic accounts that attribute the contents

Introduction • 7

of names and natural kind terms to causal-historical connections relating a speaker’s use of terms to other speakers, and ultimately to objects in the world. As I see it, social processes like these are crucially involved in determining the contents of some expressions, which in turn play a role in determining the contents of some mental states. Thus, I reject Fodor’s reductionist view in which the intentionality of natural language is reduced to the intentionality of propositional attitudes, which is then to be reduced to the intentionality of “the language of thought.” The appendix to essay 3 connects these issues to his views about “narrow psychological content,” and its alleged role in scientific psychology. For Fodor, the notion of content needed for the latter must be individuated in terms of the causal powers of mental states, which in turn supervene on brain-states. Since genuine “broad” semantic content is not individuated in this way, and does not so supervene, a notion of “narrow” content is invoked. Roughly put, he takes the narrow contents of mentalese expressions (assigned to brain-states) to be functions from world-states to the “broad” semantic contents these expressions would have if embedded in those world-states. However, there is less here than meets the eye. In order for the parts of Fodor’s system to fit together, expressions in the language of thought must be isomorphic to their narrow contents. For each such expression there must be a content, and for each content an expression. Thus appeal to narrow contents doesn’t allow one to accomplish anything that can’t be accomplished by appealing to expressions. A better way of looking at the two is to see narrow content as that which individuates expressions in the internal representational system that Fodor postulates. Essay 4, “Attitudes and Anaphora” (1992), uses facts about attitude ascriptions to investigate how sentences containing pronouns anaphoric on singular-term antecedents are understood. The aim is to illustrate the use of hyperintensional contexts to distinguish different, but intensionally equivalent, linguistic analyses. The argument rests on semantic assumptions about structured propositions and attitude ascriptions, a standard account of quantification with anaphoric pronouns functioning as variables bound by c-commanding quantifiers, and pragmatic assumptions about what it is to assert and believe bare, Russellian propositions. These assumptions are used to refute analyses according to which pronouns anaphoric on singular-term antecedents inherit the semantic contents of their antecedents (no matter whether these contents are Millian or Fregean). In place of these failed analyses, it is suggested that anaphoric pronouns with c-commanding, singular-term antecedents function as variables bound by a lambda abstraction operator introduced by the anaphoric relation itself. There is, however, a problem. While this analysis yields satisfying results in cases to which it applies, anaphoric relations in which the pronoun isn’t

8 • Introduction

c-commanded by its singular-term antecedent seem to call for the same treatment. What makes this problematic is that accounts of variable binding in natural language are standardly restricted to cases in which bound occurrences of variables have c-commanding antecedents. Essay 4 suggests that either this restriction must be lifted, or some non-variable-binding analysis must be constructed that yields the desired results in the nonc-commanding cases. Although several such analyses are considered, all are found wanting. Thus, the problem is left unsolved.

Part 2: Modality The essays in this section illustrate the interpenetration of metaphysical views about modality, epistemological views about knowledge and belief, and linguistic views about reference and semantic content. Essay 5, “The Modal Argument: Wide Scope and Rigidified Descriptions” (1998) rebuts two strategies for avoiding Saul Kripke’s celebrated modal argument against descriptivism. According to that argument the meanings (semantic contents) of names cannot be identified with the meanings (contents) of the descriptions speakers associate with them because (i) names are rigid designators, whereas (ii) those descriptions aren’t. One strategy for avoiding the argument is to deny (i) by treating names as special descriptions the behavior of which under modal operators simulates rigidity by obligatorily taking wide scope. The other strategy denies (ii) by identifying names with rigidified descriptions. I argue that both strategies fail. The driving force behind both is the desire to preserve the FregeRussell explanation of the behavior of names in attitude ascriptions, while also accommodating their behavior in modal constructions. In the former, different descriptions associated with coreferential names are used to explain the apparent possibility of substitution failure. In the latter, the wide scope given to these descriptions by the first strategy is invoked to explain the seeming rigidity of names, and the accompanying guarantee of substitution success. This strategy treats modal operators and predicates—which combine with an argument A expressing or designating a proposition—as inherently shifty. When A contains no proper names, the modal element is applied to the proposition associated with A. When A does contain a name, the modal is applied to a different proposition. Although such operators and predicates are perfectly coherent, it is clear that modal operators and predicates in English are not shifty in this way. If they were, the following argument would be understood to have true premises and a false conclusion when n meant the G, but G didn’t express an essential property of the referent of n.

Introduction • 9

The proposition that if n exists, then n is G = the proposition that if the G exists, then the G is G P2. The proposition that if the G exists, then the G is G is a necessary truth C. Therefore, the proposition that if n exists, then n is G is a necessary truth P1.



Since, in reality, arguments such as this one are recognized as valid, the wide-scope analysis gives the wrong semantics for English. The second strategy for saving descriptivism treats names as rigidified descriptions. This analysis comes to grief over the interaction of modals with attitude ascriptions. Consider, for example, the claim that ‘Aristotle’ is synonymous with some rigidified description ‘the individual who actually was so-and-so’.3 If this were so, then it would be impossible to believe that Aristotle was a great logician without believing that the individual who actually was so-and-so was a great logician. But that can’t be right. Since to believe that the individual who actually was so-and-so was a great logician is to believe, of the actual worldstate @, that the individual who was so-and-so in @ was a great logician, it is impossible to believe that the individual who actually was soand-so was a great logician without believing something about @. By contrast, some agents in merely possible world-states believe that Aristotle was a great logician, without having beliefs about @. Since it is possible to believe that Aristotle was a great logician without believing that the individual who was so-and-so in @ was a great logician, ‘Aristotle’ isn’t synonymous with ‘the individual who actually was so-andso’. The point generalizes to all names and ‘actually’ rigidified descriptions.4 Thus, Kripke’s modal argument remains intact. Essay 6, “The Philosophical Significance of the Kripkean Necessary A Posteriori” (2006), identifies and assesses two Kripkean routes to the necessary a posteriori. According to the first, successful, route, these truths attribute properties to objects or kinds that, though essential to them, can be known to be possessed by them only empirically. This leads to a sharp distinction between metaphysical and epistemic possibility. When p is a 3 Rigidifying descriptions with David Kaplan’s ‘dthat’ operator is not an attractive option for the descriptivist. Since the content of dthat (the D) is the object denoted by the D, sentences in which coreferential, ‘dthat’-rigidified descriptions are substituted for one another express the same proposition. This undercuts the chief motivation for descriptivism by depriving the analysis of names of the Frege-Russell account of their apparent behavior in attitude ascriptions. For more limitations on ‘dthat’, and on its use in analyses of names, see Soames (2005, 320–22). Further relevant discussion can be found in Soames (2003, 2:414–16). 4 A slightly expanded and updated version of this argument is given in Soames (2002, chap. 2).

10 • Introduction

necessary, a posteriori truth attributing an essential property to something (conditional on its existing), the falsity of p is perfectly conceivable, even though it is metaphysically impossible. Thus, world-states in which p is false are epistemically, but not metaphysically, possible. World-states can be identified with properties. Metaphysically possible states are maximally complete ways the real concrete universe could have been. Epistemically possible states are maximally complete ways the universe can coherently be conceived to be, which it cannot be known a priori not to be. Just as there are properties that ordinary objects could have had, and others they couldn’t have had, so there are maximal properties that the universe could have had, and others it couldn’t have had. Just as some properties that objects couldn’t have had are properties that one can conceive them as having, and that one cannot know a priori they don’t have, so some maximal properties that the universe couldn’t have had are properties that one can conceive it as having, and that one cannot know a priori it doesn’t have. These world-states are epistemically, but not metaphysically, possible. The reason empirical evidence is needed for knowledge of Kripkean necessary a posteriori truths is to rule out metaphysically impossible, but epistemically possible, world-states in which they are false. On this picture, conceivability is a useful, but fallible, guide to possibility. It is fallible because before we know much about what is actual, there are many epistemically possible states that appear to be genuinely possible, and so remain candidates for being metaphysically possible. The more we learn about the world, the more we whittle down this field of candidates, and the better able we are to identify the scope of genuine possibility. In short, our guide to metaphysical possibility is conceivability plus actuality.5 Kripke’s revival of essentialism, together with his antidescriptivist semantics, also gave impetus to externalism about linguistic and mental content, while turning analytic methodology away from its previous overreliance on conceptual and linguistic analysis. However, this wasn’t the whole of his legacy. Unfortunately, these positive developments were threatened by a confusion embedded in his second route to the necessary a posteriori. Although nearly all familiar Kripkean examples of such truths can be reached by the first, essentialist, route, identity sentences involving linguistically simple names or natural kind terms cannot. For these, Kripke appeals to a second route, which he (implicitly) generalizes to all his examples. On this route, the sharp distinction between epistemic and metaphysical possibility is lost, as are the deep lessons for philosophy, including its liberation from the confines of linguistic and conceptual analysis. 5 For further discussion see my “Kripke on Epistemic and Metaphysical Possibility” (Soames n.d.), as well as Soames (2005, 196–209).

Introduction • 11

According to the second route, instances of the necessary a posteriori express necessary propositions about individuals or kinds. However, the reason empirical evidence is required to know these propositions is not to rule out possible world-states in which they are false. Rather, evidence is needed to rule out the falsity of certain contingent—descriptive or metalinguistic—propositions the truth of which we rely on in coming to believe the necessary truths. Essay 6 makes the principles connecting these two sets of truths explicit, and uses Kripke’s own examples to rebut them—thereby undermining his second route to the necessary a posteriori, and restoring the philosophical significance of his first. Kripke’s discussion of the mind-body identity thesis illustrates the perils of not distinguishing the two routes. His discussion centers on two claims.6 (5) Heat = mean molecular kinetic energy (6) Pain = C-fiber stimulation Initially, both appear to be contingent. However, Kripke argues, in the case of (5) this impression is illusory. Suppose that ‘heat’ and ‘mean molecular kinetic energy’ are rigid designators—that is, that the state which is heat couldn’t have existed without being heat, and similarly for the state of having such-and-such mean molecular kinetic energy. Then (5) will be necessary, if true. Since it is true, it is also necessary. What, then, is responsible for the illusion that it isn’t? For Kripke, the answer has to do with how we identify, or fix the referent of, ‘heat’. Since our primary means of identifying heat is by the sensations it causes, he imagines us relying on the description ‘the cause of such-and-such sensations’ to identify it. The illusion that (5) is contingent comes from mistaking this (nonrigid) identifying description for a synonym of ‘heat’, thereby confusing the necessary truth expressed by (5) with the contingent truth expressed by (5*). (5*) The cause of such-and-such sensations in us = mean molecular kinetic energy. One who is confused in this way wrongly takes genuinely possible worldstates at which kinetic energy exists without the usual accompanying sensations to be world-states with kinetic energy but no heat. Hence, the illusion of contingency. Kripke finds the situation with (6) to be different. For the sake of argument, we assume, as before, that the terms flanking ‘=’ are rigid, in which case (6) is necessary, if true. In this case, he argues, there is no way of dismissing the initial impression of contingency. With (5), the illusion 6 For simplicity, we treat these as strict identities, with singular terms designating kinds (rather than instances of the kinds) flanking the identity signs.

12 • Introduction

of contingency is, he thinks, due to the fact that we rely on our sensations to detect the phenomenon that causes them, which is what we use ‘heat’ to talk about. With (6), the sensation is the very thing we use ‘pain’ to talk about. We don’t say to ourselves: What a horrible sensation! Let’s use the word ‘pain’ to talk about whatever causes it. Instead, we use the word to designate the sensation itself. Thus, whereas one can dispel the illusion that (5) is contingent by distinguishing between heat and the sensation used to detect it, one can’t dispel the impression of contingency of (6) that way. Since Kripke didn’t see any other way to do so, he concluded that (6) must really be contingent, if it is true at all. But, it can’t be contingent, since its terms are rigid. Thus, he concludes, isn’t true. The argument contains several points of contention. However, its most glaring weakness is its neglect of other explanations of the impression of contingency—most notably, the distinction between epistemic and metaphysical possibility. Think again about (5). Although Kripke identifies one source of the illusion of its contingency, he neglects another. Imagine a person who does not take ‘heat’ to be synonymous with any phrase about our sensations—on the grounds that the existence of heat doesn’t depend on there being sensations to detect it. Such a person might still regard (5) as contingent, because it is conceivable for heat to be something other than the motion of molecules. After all, the person might reason, it was an empirical discovery that how hot something is depends on how fast its molecules are moving. Since empirical evidence was needed to rule out possibilities in which heat is something other than molecular motion, it must be possible for it to be something else. Hence, the person may conclude, (5) must be contingent. The mistake here is that of confusing epistemic possibility with metaphysical possibility, which is a second source of the illusion that (5) is contingent. But now the flaw in Kripke’s argument against (6) is obvious. Having dismissed only one of two sources of the “illusion” that it is contingent, if true (supposing that the relevant terms really are rigid), he is not entitled to his conclusion, unless he can dismiss the other source as well. This flawed argument is part of a larger problem. The real danger lies in losing the distinction between epistemic and metaphysical possibility, and in wrongly characterizing the relationship between conceivability and possibility. The error to be avoided is in supposing that whenever something genuinely impossible seems to be conceivable, it is always because we are confused about what we are conceiving—in the way in which Kripke imagines one to be confused who takes the conceivability of molecular motion without heat sensations to be the conceivability of molecular motion without heat. The Coherent Conceivability Thesis, enshrining this error, is the price of relying on his faux second route to the necessary a posteriori. According to this thesis, if we can coherently conceive—without confusion of

Introduction • 13

the sort indicated by (5/5*)—of a world-state in which p is true (false), then there are genuine (metaphysically) possible world-states in which p is true (false). Instances of the necessary a posteriori in which I attribute essential properties to myself, using the first-person, singular pronoun, illustrate the falsity of this thesis. Since I identify the referent of my use of ‘I’ directly— without detour through identifying descriptions—confusion of the sort provided by Kripke’s heat example doesn’t arise. Yet the necessary truths expressed are easily conceived to be false. Since in these cases there is no confusion about what we are conceiving, the metaphysical impossibility of what is conceived refutes the Coherent Conceivability Thesis. Essay 6 closes with a sketch of how the missteps in Kripke’s second route to the necessary a posteriori, and his discussion of mind-body identity, are systematized and made worse in David Chalmers’s philosophically ambitious system of two-dimensional modal semantics. I close by noting how the 2D version of Kripke’s uncharacteristic error serves a larger historical agenda of attempting to reinstate the linguistic analysis of the modalities that his genuine insights showed us how to replace, and, in so doing, of returning philosophy to the confining orthodoxy of conceptual analysis, from which we thought we had escaped. Essay 7, “Knowledge of Manifest Kinds” (2004), investigates the linguistic and epistemological underpinnings of examples of the necessary a posteriori involving natural terms governed by reference-fixing intentions of the following sort: (7) a. ‘Green’ is to designate the property of surfaces that causes (nearly) all members of a certain class of paradigmatic samples to appear similar to us (and different from certain other samples). b. ‘Water’ is to designate the property shared by (nearly) all members of a certain class of paradigmatic samples (rain drops, puddles, lakes, etc.) that explains their most salient characteristics (their boiling and freezing points, their clarity, potability, etc.). The properties mentioned here are individuated by their possible instances, so that necessarily equivalent properties are identical. Like other abstract objects, they are capable of having different instances at different world-states. However, it is also natural to think of them as parts of the world that exist, and are known by us, through their instances. These worldly properties are natural kinds. For example, the kind water turns out to be the property of being made up of molecules containing two hydrogen atoms and one oxygen atom. However, ‘water’ is not synonymous with this codesignative phrase. Since simple natural kind terms are directly referential, the property designated by ‘water’ is also its meaning

14 • Introduction

(semantic content). By contrast, the meaning of the phrase is a structured complex the constituents of which are the meanings of its syntactically significant parts. Since this complex content determines the same worldly property at all possible world-states, the phrase rigidly designates the kind rigidly designated by ‘water’. Thus (8) is a necessary truth, which expresses a proposition different from those expressed by (9a) and (9b).7 (8) For all x, x is water iff x is made up of molecules containing two hydrogen atoms and one oxygen atom. (9) a. For all x, x is water iff x is water. b. For all x, x is made up of molecules containing two hydrogen atoms and one oxygen atom iff x is made up of molecules containing two hydrogen atoms and one oxygen atom. The knowledge expressed by (8) is de re knowledge of the kind water. Just as our de re knowledge of individuals standardly depends either on our own acquaintance with them, or on the acquaintance of others who pass important parts of their knowledge on to us, so our de re knowledge of manifest kinds standardly depends either on our own acquaintance with them, or on the acquaintance of others who transmit their knowledge to us. Since acquaintance with these kinds is achieved by acquaintance with their instances, this means that the knowledge expressed by (8) rests ultimately on acquaintance with instances of water, and knowledge, of those instances, that they have a certain molecular structure. This knowledge can be had only empirically. Thus, the proposition expressed by (8) is knowable only a posteriori. The same is true of innocuousseeming examples involving descriptions used to introduce, or semantically fix the reference of, natural kind terms. As I explain in the essay, (10a) is knowable only a posteriori (where W is the set of paradigmatic samples used in the stipulation, (7b), fixing the reference of ‘water’), for the same sort of reason that (10b) is. (10) a. All, or nearly all, members of W are water, if they are members of any one natural kind (of the relevant sort). b. He [said pointing at the man standing in front of me] is standing in front of me, if any (unique) man is standing in front of me. Essay 8, “Understanding Assertion” (2006), discusses Robert Stalnaker’s (1978) model of how the assertive contents of utterances are determined by 7 For more on the distinction between properties in the worldly sense, in which they can be identified with natural kinds, and properties in the linguistic sense, in which they are often taken to be meanings of complex predicates, see my “What Are Natural Kinds?” (Soames n.d.).

Introduction • 15

the interaction of the semantics of the sentences uttered with the presuppositions of conversational participants, obvious features of the context, and the pragmatics of assertion. The model’s chief positive lesson is that even when metaphorical and other special uses of language are put aside, there is often a substantial gap between what one asserts and the semantic content of the sentence one uses to assert it. Stalnaker’s linguistic task is to articulate principles capable of filling this gap. His philosophical task is to use those principles to reconcile Kripkean instances of the necessary a posteriori with his treatment of propositions as functions from possible world-states to truth-values, and his restriction of the epistemically possible to the metaphysically possible. My aim in the essay is to explain how the model is supposed to work, and why, in the end, it doesn’t—while distinguishing what can be salvaged from what can’t. For Stalnaker, the aim of discourse is to distinguish the actual worldstate from other world-states compatible with everything previously assumed or established. The function of assertion is to narrow down this set of actual-world-state candidates by eliminating those incompatible with what is asserted. When a sentence S semantically expresses a necessary truth, assertion of that truth would fail to eliminate any worldstates, and so be pointless. Thus, Stalnaker thinks, uttering S results in the assertion of a different proposition. Typically this is the so-called “diagonal proposition”—the (unique) proposition that is true (false) at a candidate-state w iff the proposition S would express were w to be actual would be true (false). This is the deflationary core of what was to become the 2D account of instances of the necessary a posteriori. In my essay, I argue that Stalnaker’s deflationary 2D explanation is defeated by the inability of his model to accommodate de re beliefs and assertions. If these are allowed, the world-states needed for the assignment of diagonal propositions to utterances predicating essential properties of individuals or kinds turn out to be either metaphysically impossible, or incompatible with what has already been assumed or established in the conversation. Both alternatives violate central tenets of the model. In the end, there is no satisfactory solution. Although the range of counterexamples can be reduced by countenancing epistemically possible worldstates that aren’t metaphysically possible, the basic problem can still be re-created. Nor is it feasible to exclude de re beliefs and assertions altogether. Although doing so might render world-states needed for suitable diagonal propositions compatible with the non–de re beliefs and assumptions of conversational participants, such an exclusion would mischaracterize the information carried by utterances. Nor can such an exclusion be internally justified, since the de re knowledge of world-states presupposed by the model rests on the very de re knowledge of individuals and kinds that would have to be excluded.

16 • Introduction

Thus, the basic structure of the model must be given up. Instead of taking world-states to be basic and propositions to be sets of such states, we should take propositions and their constituents to be basic, and think of world-states as deriving from them. Instead of restricting epistemic possibility to metaphysical possibility, we should recognize the former as outstripping the latter. Instead of aspiring to a deflationary account of the necessary a posteriori, we should embrace the metaphysically robust Kripkean account. Finally, we should recast the rules of Stalnaker’s discourse model in a way that reflects all this, while preserving his insight that pragmatics has an important role to play in filling the gap between semantic content and assertion. The essay closes with some steps in this direction. In essay 9, “Ambitious Two-Dimensionalism” (2007), I offer a broad overview of attempts by opponents of the antidescriptivist revolution of the 1970s to use the technical apparatus of two-dimensional modal logic to reinstate descriptivism in semantics, conceptualism in our understanding of the modalities, and linguistic analysis as the core of our general philosophical methodology. To that end, I articulate and criticize four versions of this ambitious philosophical and linguistic program— Robert Stalnaker’s pragmatic version from the late 1970s, a strong semantic version suggested in the mid-1990s by Frank Jackson and David Chalmers, a weak semantic version that is a natural retreat from the strong version, and a hybrid version suggested by Chalmers in 2002.8 I argue that all these systems fail, and that the central ideas motivating them are incorrect. That said, an interpretive caveat must be noted. The main texts used in arriving at what I call “strong” and “weak” two-dimensionalism— Chalmers’s The Conscious Mind, and Jackson’s From Ethics to Metaphysics—do not explicitly state or unequivocally endorse either system. However, that is only because they don’t explicitly state or unequivocally endorse any system. The main area of inexplicitness concerns the semantic analysis of knowledge, and other attitude ascriptions. Although Chalmers and Jackson offer no precise analyses, the contentious conclusions they draw about the necessary a posteriori require 2D analyses of these constructions—about which their texts contain many suggestive hints.9 My strong and weak 2D systems make these analyses explicit. However, since the analyses are refuted, the 2D conclusions about the necessary a posteriori are left unsupported. Since Chalmers and Jackson don’t explicitly endorse the analyses, strictly speaking, their texts have 8 See Chalmers (1996); Jackson (1998); and Chalmers (2002). The ambitious philosophical motivations of Chalmers and Jackson were shared by Stalnaker only in part. 9 See Soames (2005) for textual documentation.

Introduction • 17

not been refuted—but only because their key arguments were insufficiently explicit to support any far-reaching conclusions. Essay 10, “Actually” (2007), presents a theory of the metaphysics and epistemology of actuality and possibility, and the language we use to talk about them. World-states—which are consistent, maximally informative properties attributed to the universe—include the actual world-state, which is instantiated, metaphysically possible states, which could have been instantiated, and epistemically possible world-states, which cannot be known a priori not to be instantiated. The contents of these properties are knowable a priori. However, it is argued, empirical knowledge of the actual world-state also arises when it is presented to us indexically. This duality is the key to understanding the actuality operator, and to solving important puzzles about the necessary a posteriori and the contingent a priori. Whenever S expresses a contingent truth p, Actually S expresses the necessary truth that p is true at @.10 However, since Actually S is trivially inferable from S, and since the proposition it expresses often doesn’t seem knowable in any other way, it has seemed to be knowable only a posteriori, whenever p is. But this is problematic. If p is true at @, then the proposition that p is true at @ is true, not just at every metaphysically possible world-state, but at every epistemically possible state as well. Why, then, if there are no possible world-states at which this proposition is false, is empirical evidence required to know it? I argue that it isn’t. World-states are properties of making certain world-describing sets of propositions true. Imagine, then, a tiny universe consisting of two blocks side by side, with a third on top. This worldstate, Tiny, is the property of containing blocks 1 and 2 side by side, with block 3 on top. We can know, just by thinking about this property, that if it were instantiated, then block 3 would be sitting on blocks 1 and 2. So, when p is the proposition that block 3 is sitting on those blocks, it is knowable a priori that p is true at Tiny. This point generalizes to realworld inquiries in which the relevant world-states are finitely specifiable. For every such state w, and every proposition p the truth of which is calculable from the basic propositions defining w, the proposition that p is true at w is knowable a priori. This result applies to the actual worldstate (relative to an inquiry), as much to any other. Thus, the propositions expressed by uses of Actually S are often knowable a priori, even when they are not, in fact, known a priori. Since the actual world-state, @, relative to many inquiries, will be much more complex than Tiny, we may not be able grasp it in the nondemonstrative way we grasp Tiny—in which case we won’t be able to calculate the truth-values of propositions 10

‘@’ designates the actual world-state—that is, the world-state of our present context.

18 • Introduction

from a complete specification of @. In such cases, our only practical way of learning that p is true at @ is by inferring it from p. So, when p is a posteriori, our knowledge of the proposition expressed by Actually S may be a posteriori, even though what we know can, in principle, also be known in another way. The proposition that p is true at @ is entertainable in two radically different ways. One way—which, as a practical matter, may exceed our cognitive abilities—involves grasping the propositional content of @. One who entertains the proposition in this way can know it a priori—by deriving p from the propositions that define @. But when @ is presented in this way, there is no way of knowing that it is instantiated. Hence, when one entertains the proposition that p is true at @ in a way that allows one to know it a priori, there is nothing in one’s knowledge that allows one to infer p from it. The second, indexical, way of entertaining the proposition that p is true at @—which is how it is presented using the actuality operator—doesn’t involve grasping the full propositional content of @. When presented with the proposition in this way, we can’t determine it to be true a priori, though we can move a priori from it to p, and vice versa. Since on neither way of knowing that p is true at @ is there a way of establishing p a priori, p is knowable only a posteriori. The contingent a priori gives rise to a related puzzle. I have argued that the function of empirical evidence needed to know Kripkean examples of the necessary a posteriori is to rule out epistemically possible, but metaphysically impossible, world-states in which they are false. This may seem to suggest (11). (11) If p is false at some epistemically possible world-state, then p isn’t a priori. So, if p is a priori, then p is true at every epistemically possible world-state. But then, if p is contingent a priori, it will follow that p is true at all epistemically possible world-states, while being false at some metaphysically possible state. So, if (11) is true, some metaphysically possible worldstates are epistemically impossible. This, I argue, is incorrect. The puzzle can’t be solved by denying the contingent a priori—since, as I show, when S is a contingent truth anyone who, at @, knows the a priori truth expressed by S iff S is in a position to derive the contingent S iff actually S by steps that can be known a priori to be truth preserving. If this is right, then the only way to solve the puzzle is by denying (11). Although the details of the argument are intricate, the basic point is simple. The proposition that p iff p is true at @ is knowable a priori by agents evaluating it at @—when it is presented to them by S iff actually S—because they know that the world-state @ it is

Introduction • 19

about is the world-state at which they are evaluating it. The fact that it fails to be true when evaluated at world-states distinct from @ is irrelevant. No empirical evidence is required to rule out such states because the agents know in advance that they are not evaluating it there. Thus, (11) is false. So is the closely related principle (11+) (11+) If it can be known a priori both (i) that p is false, and (ii) that p is true at w, then it can be known a priori that (iii) w is not instantiated (and so is epistemically impossible). The falsity of this principle is due to the fact that a priori consequences of propositions that are knowable a priori are sometimes not themselves knowable a priori. In this case, the proposition that it is false that (~p and p is true at @)—playing the role of (i) in (11+)—can be known by us a priori, when @ is presented indexically—as it is in It is false that (~S & actually S). By contrast, the proposition that the proposition that (~p and p is true at @) is true at w—playing the role of (ii) of (11+)—can be known a priori only when it is known a priori that p is true at @. This requires @ to be presented nonindexically, in terms of its propositional content. Since there is no way of merging the a priori routes to (i) and (ii) into a single a priori route to (iii), an agent can’t derive (iii) from a priori knowledge of (i) and (ii), and (11+) fails. Having disposed of puzzles about the necessary a posteriori and the contingent a priori, I turn to a puzzle about language. On the one hand, prefixing the actuality operator to a contingent, a posteriori truth results in a dramatic change in semantic content (from contingent to necessary and from a posteriori to a priori). On the other hand, adding it in ordinary conversation often seems to be a purely rhetorical move, with no effect on what is asserted. For example, the difference between uttering, “Actually, I live in California,” and “I live in California” lies not in what is asserted (which is the same), but only in the suggestion, carried by the first utterance, that my assertion may be unexpected. My explanation of this follows naturally from the indexical semantics of the actuality operator that gives it its modal, epistemic, and semantic punch. The essay ends with a discussion of broader linguistic, epistemic, and metaphysical challenges.

Part 3: Truth and Vagueness Essay 11, “What Is a Theory of Truth?” (1984), explains, and gives a limited defense of, the philosophical significance of Tarski’s theory/definition

20 • Introduction

of truth. The defense depends on taking the notion he defined as a replacement for our ordinary truth predicate that neither captures the meaning of ‘true’, nor provides a notion suitable for explicating meaning, or advancing contentious philosophical theses. The virtues of Tarski-truth are (i) that it avoids the liar paradox, (ii) that it provides a precisely defined basis for mathematically serviceable notions such as arithmetical truth, definability, logical truth, and logical consequence, (iii) that it uses semantic ascent to enrich the expressive power of metatheoretical investigations, and (iv) that it is transparently deflationary—thereby removing the dangerous temptation to read contentious epistemological and metaphysical views into what should be a philosophically neutral notion. The price to be paid for this progress is the frank recognition of the irrelevance of Tarski-truth to reductionist programs like physicalism, to the project of giving theories of meaning for interpreted languages, and to the practice of providing interpretations for uninterpreted languages by stipulating the truth-conditions of their sentences. The key point is that statements of the Tarski-truth conditions of sentences provide no information about their meanings. For example, when ‘T-in-L’ is a Tarskian truth predicate, (12a) and (12b) are interderivable using elementary truths of logic, set theory, and the syntax of L. (12) a. ‘Ws’ is T in L iff snow is white. b. Snow is white iff snow is white. Since knowing the latter provides no information about meaning of ‘Ws’, knowing the former doesn’t either. Although this makes Tarski-truth useless for Davidson’s project using truth to explain meaning, for Tarski it is the inevitable by-product of something good. Since the only commitments carried by S is T-in-L not carried by S itself are those involving syntax and elementary set theory, we are guaranteed that if L is unproblematic, and syntax and set theory are too, then introducing ‘T-in-L’ can’t cause problems. The defense of Tarski on this point is that Davidsonian semantic theories are not what we want anyway. Thus, the incompatibility of Tarski-truth with these theories of meaning is not, in itself, a black mark against Tarski. A related defense is given against Hartry Field’s criticism that Tarski’s theory doesn’t do enough to reduce semantic facts to physical facts. The defense is that such a reduction is irrelevant to Tarski’s project, which doesn’t do anything to advance it. If one distinguishes genuine semantic facts (having to do with meaning) from facts stateable using Tarski-truth, and related notions, then—except for his embrace of abstract objects (sets, expression types, etc.)—the latter pose no problem

Introduction • 21

for physicalism.11 What about genuine facts about the meanings of the expressions uttered by speakers? In the essay, I play with the idea that if we take meanings to be essential properties of expressions, then we might be able to give a purely formal semantic account— linking expressions with entities they express or designate—while confining philosophically and empirically significant issues to questions about which abstractly characterized expressions speakers should be regarded as using. I am not convinced that this is the best way of thinking of the matter, only that it is a defensible one for formalists about meaning to adopt. The essay closes with an understated admission of the limitations of my defense of Tarski. I don’t think that Tarski-truth is sufficient for all theoretical purposes for which we need a truth predicate. Although I don’t accept Davidsonian-style theories of truth and meaning, I believe we do need non-Tarskian notions of truth in semantics that are much closer to our ordinary notion than any he provided. I also believe that more work on truth has yielded, and will continue to yield, better accounts of the liar. However, I appreciate Tarski’s deflationary perspective, and remain convinced that deflationism is crucial to any adequate understanding of truth.12 Essay 12, “Understanding Deflationism” (2003), is devoted to spelling out what deflationism is. Roughly put, a predicate that applies to all and only the true sentences of a language L is a deflationary truth predicate iff the claim that it applies to a sentence of L is a trivial, necessary, and a priori consequence of the claim made by the sentence itself, and vice versa. In this sense, a Tarskian truth predicate for a restricted fragment E of English for which it can be defined is deflationary. However, the one-place predicate ‘is a true sentence of E’, formed by filling the second argument place of the relational predicate ‘is a true sentence of’ that occurs in ordinary English is not. For some purposes, the deflationary truth predicate for E is sufficient; for others, the nondeflationary truth predicate—which can be defined in terms of propositional truth—is required. This raises the 11 Tarski was not as clear as one might have hoped about the distinction between genuine semantics and semantics based on his notion of truth. In the essay, I note in passing that he recognized that his notion of truth-in-L can’t be used to give the meanings of the truth-functional connectives in L. However, I fail to note that other comments he makes (Tarski 1944, 1969) indicate an insufficient appreciation of the chasm separating his truth predicate from the ordinary English predicate, and of the related distinction between genuine facts about meaning and facts stateable using his truth predicate. For discussion see Soames (1999b, 238–44); also Soames (1995). 12 See Soames (1999b, chaps. 4, 6, and 8 ); also essays 5, 7, and 8 of volume 1 of this collection.

22 • Introduction

issue of truth for propositions. To say that propositional truth is deflationary is, I suggest, to say (i) that p and the proposition that p is true are trivial, necessary, and a priori consequences of one another, and (ii) that any warrant for believing, assuming, denying, or doubting one is warrant for taking the same attitude toward the other. In this sense, I argue, propositional truth is deflationary. Although this conclusion is, I think, both true and informative, it doesn’t provide an explanation of what it is to understand the predicate ‘true’, when applied to propositions. Since I appeal to necessary and a priori consequence—which are themselves defined in terms of truth—my claim that propositional truth is deflationary doesn’t explain the truth predicate in a way that can be grasped by someone who doesn’t already understand it. Thus, it’s not an analysis of what truth is, or what ‘true’ means. It is simply a philosophically important truth about truth. In giving my favorable assessment of deflationism, I identify its chief philosophical opponents as nonfactualists. It is these philosophers—who deny that warrant for p is warrant for the claim that p is true (rather than those who merely add contentious doctrines to deflationism connecting truth to what we can know, or what it is useful for us to believe)—who are the real inflationists about truth. Nonfactualists accept (believe, assert) certain propositions p (about one or another philosophically contentious domain) while rejecting (disbelieving, denying) corresponding claims to the effect that p is true. Interestingly, these philosophers don’t deny that the claim that p is true is a necessary consequence of p. Instead, they maintain, this only shows that they are required to accept the truth of the claim that p is true if they accept the truth of p—which they don’t. In so doing, they reject the idea that one who accepts p, while recognizing that q is a consequence of p, is thereby committed to q. I argue that this rejection is self-defeating. It also misses the explanatory power of deflationism. We are interested in the relation of logical consequence among sentences, and necessary consequence among propositions, because we take them to be connected to the argumentative commitments we take up when we accept a set of premises. The reason why a definition of consequence in terms of guaranteed truth-preservation serves this interest is that we take the central deflationist point for granted: namely that acceptance of S (and the proposition it expresses) carries with it a commitment to the truth of S (and the proposition it expresses), and vice versa. If we didn’t take this for granted, then recognizing an argument to be truth preserving would not give us guidance about what acceptance of its premises committed us to—which, of course, it does. Thus, deflationism about truth is central to our practice of using logical and necessary consequence to track our argumentative commitments.

Introduction • 23

Essay 13, “Higher-Order Vagueness for Partially Defined Predicates” (2003), explains what the higher-order predicate ‘is determinately red’ has in common with the partially defined predicate ‘is red’.13 To say that the latter is partially defined is to say that the rules governing its application provide sufficient conditions for it to apply to an object, and sufficient conditions for it not to apply, but no conditions that are both individually sufficient and disjunctively necessary for it to apply, or not to apply. This results in a three-way classification of objects. Those to which the rules of the language, plus the underlying nonlinguistic facts, determine that ‘is red’ applies are in the determinate-extension of the predicate. Those to which the rules plus facts determine that it doesn’t apply are in its determinate antiextension. Those for which the predicate is undefined, because the rules plus facts determine neither result, are excluded from both classes. Next consider the predicate ‘is determinately red’, which applies to objects in the determinate-extension of ‘is red’, and does not apply to objects not in the determinate-extension. If every object is either in that determinate-extension or not, then ‘is determinately red’ is totally defined, and so, on my account, is not vague. That seems wrong. Not only is there no sharp and precise line dividing the objects to which ‘is red’ applies from those to which it doesn’t, there also seems to be no sharp and precise line dividing the objects for which it is determined, by the rules of the language and the underlying facts, that ‘is red’ applies from those for which this is not determined. Thus, there is reason to resist the claim that ‘is determinately red’ is totally defined. For some putative rules it is indeterminate whether or not they are rules of the language governing ‘is red’. Let R be the class of such putative rules. For certain objects o, the question Q. Is the claim that ‘is red’ applies to o a necessary consequence of the rules of the language governing the predicate, plus the underlying facts?

13 ‘Is red’ is also context sensitive. Its default determinate-extension is the set of things to which the rules of the language, together with underlying nonlinguistic facts, determine that it applies. Its default determinate antiextension is the set of things to which the rules of the language plus underlying facts determine that it doesn’t apply. For all objects o, ‘is red’ is undefined for o just in case o is in neither of these classes. Since these classes don’t exhaust all cases, speakers have the discretion of adjusting the extension and antiextension so as to include initially undefined cases. Often they do this by explicitly predicating ‘is red’ of an object o, or by explicitly denying such a predication. When this happens, and other conversational participants go along, the extension (or antiextension) of the predicate in the context is adjusted so as to include o, plus all objects that bear a certain relation of similarity to o.

24 • Introduction

can be answered only by assuming that certain members of R are rules of the language governing ‘is red’, or by assuming that they aren’t. Since neither of these assumptions can be established, there is no possible justification for accepting them. Thus, we should reject both the claim that these objects are determinately red, and the claim that they are not. Like the application of ‘is red’, the application of ‘is determinately red’ is subject to a range of vagueness and indeterminacy. Of course, this range is smaller in the latter case than in the former. Can it be further reduced by additional iterations of the determinately operator? The range of indeterminacy for ‘is determinately red’ is smaller than the range of indeterminacy for ‘is red’. Is the range of indeterminacy for ‘is determinately, determinately red’ still smaller? The answer turns out to be ‘no’. Iterating ‘determinately’ has no effect. The predicate ‘is determinately red’ and ‘is determinately, determinately red’ are equivalent. They are weakly partial, in a sense analogous to the sense in which ‘is red’ is partial. It is important not to be misled by this result. Nothing in the semantics of partial, or weakly partial, predicates shows that the indeterminacy to which they are prone can’t be eliminated. Once we understand them, we can always introduce artificially defined substitutes that are totally defined. By the same token, nothing in the semantics of partial, or weakly partial, predicates shows that they don’t impose sharp and precise lines distinguishing different categories of objects. They do. However, the resulting fine-grained classifications are less worrisome than they have sometimes been thought to be. They do not, for example, pose the threat to our notion of linguistic competence posed by a sharp and precise division between the objects that an ordinary vague predicate is determinately true of and those it is determinately not true of. The distinction between truth and untruth is important to us; and the norms of language use presuppose that we are able (given suitable facts) to track it. The same cannot be said of the distinction between (i) statements that are true because they are determined to be so by the rules of one’s language, together with nonlinguistic facts, and (ii) statements for which there is no saying whether they are true for that reason, or true because speakers have made them true by exercising a minimal amount of discretion in adjusting the boundaries of the partial, context-sensitive, predicates employed. If my model of vague predicates is correct, there may be a sharp distinction between (i) and (ii). However this distinction, unlike the one between truth and untruth, is highly theoretical, and need not be reliably tracked by competent speakers. Thus, the fact that they are typically oblivious to it is not paradoxical. Essay 14, “The Possibility of Partial Definition,” defends this analysis of vague predicates against an objection due to Michael Dummett (1978)

Introduction • 25

and Michael Glanzberg (2003). The objection rests on the view that an assertion of a proposition p is correct (in the sense of satisfying the intrinsic norm of assertion) just in case p is true, and incorrect just in case p is not true. If this conception of assertion is correct, then the impossibility of partially defined predicates follows from (i) and (ii), each of which is integral to the theory of partial definition (in my sense). (i) Assertions of propositions expressed by sentences attributing partially defined predicates to objects for which they are undefined violate the norm of assertion, and so are incorrect. (ii) Truth is undefined for the propositions in (i)—in which case, the proper response is to reject both the claim that they are true, and the claim that they are not true. The reply to this objection is that although (i) and (ii) are correct, the Dummett-Glanzberg conception of assertion is not. Instead, it should be replaced by Timothy Williamson’s (1996) conception, according to which an assertion of a proposition p is correct (satisfies the norm of assertion) just in case the agent knows p, and incorrect just in case the agent doesn’t know p. After giving reasons for preferring this account, I show how the incorrectness of asserting undefined propositions follows from they fact that such propositions are, in principle, unknowable. Thus, the DummettGlanzberg argument against the possibility of partially defined predicates is defeated. Intuitively, this is the right result. No matter how attractive, or unattractive, partially defined predicates may be for various purposes, surely it is at least possible to introduce and use them in certain situations. I flesh out this possibility by elaborating an imaginary case in which the color terms of an isolated community living in a restricted environment are introduced by meaning-giving stipulations that seem, transparently, to be partial. These predicates are meaningfully used for a time, after which the discovery of previously unencountered hues gives rise to a practice in which speakers allow themselves a range of conversational discretion in applying the predicates to items for which they were initially undefined. At this point the predicates become context-sensitive, as well as partially defined. After describing the semantic properties of these predicates, I bring the example closer to home by allowing the initially clear and precise specification of the rules of the language governing them to become muddied— with a consequent blurring of the boundary between what is in their default determinate-extensions, and what isn’t. This results in the kind of indeterminacy explored in essay 13. As with vague predicates of English, so with the context-sensitive, partially defined predicates in my imagined language, there is a limited range of cases in which it is indeterminate

26 • Introduction

whether the claim that a predicate applies to an object is true because the rules of the language, plus facts about the object, determine it to be so, or whether it is true because speakers have exercised minimal discretion in adjusting the boundaries of the predicate to make it apply to the object. What precisely is the range of this indeterminacy? I am now inclined to think the Williamsonian view—that, somehow, the question has a definite answer, even though we can’t find it—is correct. If so, then context sensitivity and partial definition don’t tell the whole story about vague language. I do, however, continue to believe that they are part of the story. The essay therefore concludes with two arguments—one identifying a theoretical advantage in recognizing vague predicates as partially defined, given that they are context sensitive, and the other showing how, given partial definition, context sensitivity can be used to explain our differing reactions to violations of different “laws” of classical, two-valued logic.

Part 4: Kripke, Wittgenstein, and Following a Rule Essays 15 and 16—“Skepticism about Meaning: Indeterminacy, Normativity, and the Rule-Following Paradox” (1998) and “Facts, Truth Conditions, and the Skeptical Solution to the Rule-Following Paradox” (1998)—are commentaries on Kripke’s (1982) work on Wittgenstein’s rule-following paradox, read, not as an interpretation of Wittgenstein, but as the development of a novel skeptical position about intentionality. The position has two parts. The first is a skeptical argument for the paradoxical conclusion that there are no facts about meaning, or representational content. The second is a “skeptical defense” that attempts to save our practice of ascribing meanings, and contents, by stripping it of the presumption that such ascriptions purport to state facts. In essay 15, I argue that Kripke’s skeptical argument is guilty of an equivocation similar to one in Quine’s argument for the indeterminacy of translation.14 In Kripke’s case, the equivocation centers on the normativity requirement he places on attempts to specify the putative fact that determines addition as the meaning of ‘+’. I argue that two different formulations of this requirement—NE and NM—are extractable from his discussion. NE

If the fact that F determined that (in the past) one meant addition by ‘+’, then knowing that F should provide one with a sufficient basis for concluding that ‘125’ is the correct answer to the

14 I discuss Quine’s arguments, and equivocation, in further detail in Soames (1999a; 2003, vol. 2, chap. 10; 2007).

Introduction • 27

NM

question “What is 68 + 57?”, and so is the answer one ought to give, provided one intends to use ‘+’ with the same meaning it had in the past. (Similarly for other calculations) If the fact that F determined that (in the past) one meant addition by ‘+’, then one means addition by ‘+’ at any possible world-state at which it is true that F; hence ‘125’ is the correct answer, and so the answer one ought to give, to the question “What is 68 + 57?”, in such a world-state. (Similarly for other calculations)

NE is plausible, provided that the determination relation in question is given by DetE. DetE P determines Q only if, given P, one can demonstrate Q without appealing to any other empirical facts—i.e., only if Q is an a priori consequence of P. Read in this way, the effect of NE is to restrict potential meaning-of-‘+’determining facts that F to those knowledge of which would, in principle, allow one to demonstrate that (in the past) one meant addition by ‘+’. Kripke argues that whatever one’s past dispositions to calculate using ‘+’ may have been, they do not satisfy this restriction. Since (i) it is plausible to suppose that a similar conclusion can be reached for any nonintentional fact that F one might cite, and (ii) intentional facts have previously been dismissed as possible meaning-determining facts (on the grounds that they are subject to the same skeptical argument), Kripke’s skeptic reaches the conclusion that no facts determine the meaning of ‘+’. Unfortunately for the skeptic, what the argument establishes is not particularly skeptical, or paradoxical. The basic point—that truths about meaning are not a priori consequences of nonintentional truths—is philosophically interesting, but not paradoxical. In contrast to the claim that statements about meaning are not necessary consequences of nonintentional truths, it does not threaten the very existence of truths about meaning. This is where NM comes in. When the normativity requirement is understood in this way, it is plausible, provided that the determination relation in question is given by DetM. DetM

P determines Q only if Q is an necessary consequence of P.

Read in this way, NM restricts potential meaning-of-‘+’-determining facts to those which have the claim that (in the past) one meant addition by ‘+’ as a necessary consequence. If the skeptic can show that no nonintentional facts satisfy this restriction, then—since, surely, facts about what one means can’t simply float free of underlying nonintentional facts about one’s history, environment, linguistic behavior, and brain-states—the skeptic may plausibly maintain that there are no facts whatsoever about

28 • Introduction

what one means by ‘+’. However, now he faces a different problem. Although many candidates for the nonintentional fact determining addition as the meaning of ‘+’—including some simple dispositional facts—fail to satisfy NM, it is far from obvious that this is true for all such candidates. On the contrary, it is highly plausible to suppose that some appropriately global candidates do satisfy the requirement. Since the Kripkean skeptic does nothing to undermine this supposition, his argument fails. In essay 16, I turn my attention to Kripke’s so-called skeptical solution. What makes the position skeptical is that it purports to accept the conclusion that there are no facts about meaning; what makes it a solution is that it nevertheless defends the correctness of ascriptions like ‘Soames means addition by ‘+”. I argue that there are two interpretations of the position, corresponding to two interpretations of what is meant by ‘facts’. According to the first, minimal, interpretation, all instances of the schema, (13), are embraced. (13) It is fact that S iff it is true that S iff the proposition that S is true iff S On this interpretation, the skeptical solution claims that meaning ascriptions like (14) don’t express propositions, or purport to state facts, but rather have some other, nondescriptive, kind of meaning. (14) In the past, I meant addition by ‘+’. I argue that, when understood in this way, the skeptical solution is indefensible, and ultimately self-defeating—in addition to being unsuitable for certain key argumentative purposes of Kripke’s Wittgenstein. According to the second interpretation, what is denied is not that there are meaning facts in the minimal sense of (13), but that there are facts about meaning, truth, and reference the grasp of which by speakers plays a central role in explaining how words come, initially, to be endowed with meaning, and how, later, they are understood. The target here is the classical truth-conditional conception of meaning. According to it, all words (and sentences) are endowed with meaning by conceptually prior intentions to use them with certain reference (and truth) conditions. Once they have acquired their meanings, one who understands them does so in virtue of recognizing these conditions. I argue that there is an important element of truth in the second interpretation of the skeptical solution, which disputes this explanatory picture. For some words and sentences, one does not understand them because one knows their reference and truth conditions. Instead, one knows those conditions, in part, because one understands them—where understanding crucially involves not just one’s private semantic beliefs, but also one’s position in a broader linguistic community.

Introduction • 29

There is much in the argument I give for this that fits Kripke’s text. However, there are also important differences. Whereas the critique of the classical truth-conditional conception he offers is global—purporting to show that no expressions conform to it—mine is local, establishing only that some don’t. In fact, my critique doesn’t apply to Kripke’s paradigmatic example of ‘+’. This points to a larger interpretive problem. To argue effectively against the classical conception one must show that one class of intentional facts—which includes semantic beliefs and intentions—is not always conceptually prior to, and part of the explanation of, a different class of intentional facts—which includes understanding expressions and using them to mean certain things. Kripke’s Wittgenstein doesn’t argue in this way. Instead, he looks for a nonintentional basis of all intentional facts, and suggests that there is none. Because of this, his skeptical argument can’t be seen as a reductio ad absurdum of the classical truth-conditional conception of meaning. Thus, in the end, Kripke’s text leaves us in a quandary. The most natural interpretation of the skeptical argument makes no room for a reasonable version of the skeptical solution, while the most reasonable version of the skeptical solution can’t be seen as a response to the most straightforward reconstruction of the skeptical argument. Though hybrid interpretations are possible, none is fully satisfying. The reason for this, I suggest, is that there is no single, coherent line of argument running through the entire text. My recommendation is to give up the illusion of a single unifying interpretation, to distinguish different illuminating partial interpretations, and to identify what there is to learn from each. That, at any rate, is what I offer.

References Chalmers, David. 1996. The Conscious Mind. New York: Oxford University Press. ———. 2002. “Components of Content.” In The Philosophy of Mind: Classical and Contemporary Readings, ed. David Chalmers, 608–33. New York: Oxford University Press. Dummett, Michael. 1978. “Truth.” In Truth and Other Enigmas, 1–24. Cambridge: Harvard University Press. Originally published in Proceedings of the Aristotelian Society 59 (1959): 141–62. Edelberg, Walter. 1994. “Propositions, Circumstances, and Objects.” Journal of Philosophical Logic 23:1–34. Glanzberg, Michael. 2003. “Against Truth-Value Gaps.” In Liars and Heaps: New Essays on Paradox, ed. JC Beall, 151–94. Oxford: Clarendon Press; New York: Oxford University Press. Jackson, Frank. 1998. From Metaphysics to Ethics. Oxford: Clarendon Press.

30 • Introduction Kripke, Saul A. 1982. Wittgenstein: Rules and Private Language. Cambridge: Harvard University Press. Soames, Scott. 1995. “T-Sentences.” In Modality, Morality, and Belief: Essays in Honor of Ruth Barcan Marcus, ed. Walter Sinnot-Armstrong in collaboration with Diana Raffman and Nicholas Asher, 250–70. Cambridge: Cambridge University Press. ———. 1999a. “The Indeterminacy of Translation and the Inscrutability of Reference.” Canadian Journal of Philosophy 29:321–70. ———. 1999b. Understanding Truth. New York: Oxford University Press. ———. 2002. Beyond Rigidity: The Unfinished Semantic Agenda of “Naming and Necessity.” New York: Oxford University Press. ———. 2003. Philosophical Analysis in the Twentieth Century. 2 vols. Princeton: Princeton University Press. ———. 2005. Reference and Description: The Case against Two-Dimensionalism. Princeton: Princeton University Press. ———. 2007. “What We Know Now That We Didn’t Know Then.” Philosophical Studies 135:461–78. ———. n.d. “Kripke on Epistemic and Metaphysical Possibility: Two Routes to the Necessary Aposteriori.” In Saul Kripke, ed. Alan Berger. Cambridge: Cambridge University Press, forthcoming. ———. n.d. “What Are Natural Kinds?” Forthcoming in Philosophical Topics. Stalnaker, Robert. 1978. “Assertion.” Syntax and Semantics 9:315–32. Tarski, Alfred. 1944. “The Semantic Conception of Truth and the Foundations of Semantics.” Philosophy and Phenomenological Research 4:341–75. ———. 1969. “Truth and Proof.” Scientific American, June, 64–77. Williamson, Timothy. 1996. “Knowing and Asserting.” Philosophical Review 105:489–523.

PA RT O N E

Reference, Propositions, and Propositional Attitudes

This page intentionally left blank

ESSAY ONE

Direct Reference, Propositional Attitudes, and Semantic Content

What do we want from a semantic theory? A plausible answer is that we want it to tell us what sentences say. More precisely, we want it to tell us what sentences say relative to various contexts of utterance. This leads to the view that the meaning of a sentence is a function from contexts of utterance to what is said by the sentence in those contexts. Call this the propositional attitude conception of semantics. Another semantic picture that has enjoyed considerable popularity is the truth-conditional conception. According to it, the job of a semantic theory is to tell us what the truth conditions of sentences are. On this view, the meaning of a sentence can be thought of as a function from contexts of utterance to truth conditions of the sentence as used in those contexts. Suppose now that we put the propositional attitude and the truthconditional conceptions together. If we do this, it is virtually irresistible to conclude that what is said by a sentence in a context consists in its truth conditions relative to the context. But what are truth conditions? One natural idea, embraced by the ruling semantic paradigm, is that the truth conditions of a sentence, relative to a context, are the metaphysically possible worlds in which the sentence, as used in the context, is true. Such truth conditions can be specified by a recursive characterization of This essay grew out of work originating in my critique, “Lost Innocence,” of Situations and Attitudes, by Jon Barwise and John Perry. It was written in 1983–84 while on leave from Princeton University on the Class of 1936 Bicentennial Preceptorship, and while a guest, first, of the Syntax Research Center at the University of California, Santa Cruz, and, later, of the University of Washington philosophy department. Portions of it provided the basis for talks at the University of California at Berkeley, Riverside, and Santa Cruz (1983–84); the University of Illinois (1985); North Carolina State University (1985); the University of Pennsylvania (1985); Princeton University (1984); Simon Fraser University (1983); Stanford University (1984); and the Pacific Division Meetings of the American Philosophical Association (1985). A shortened version of the essay, adapted from the APA talk appears in Almog, Perry, and Wettstein (1989). I have benefited considerably in the development of several important points from extensive discussion and correspondence with Joseph Almog, David Kaplan, and Nathan Salmon. I have also profited from discussion with Ali Akhtar Kazmi, Julius Moravcsik, and Mark Richard.

34 • Essay One

truth relative to a context and a world. This characterization implicitly associates with each sentence a function representing its meaning. The value of the function at any context as argument is the set of metaphysically possible worlds in which the sentence, as used in the context, is true. It is this that is identified with what is said by the sentence in the context, when the propositional attitude conception of semantics is combined with this version of the truth-conditional conception. This identification is, of course, highly problematic. The first difficulty one notices is that if S and S′ are necessarily equivalent relative to a context, then they are characterized as saying the same thing, relative to the context. However, it is highly counterintuitive to hold that all necessary truths say the same thing, that the conjunction of a sentence with any necessary consequence of it says the same thing as the sentence itself, and so on. A plausible pragmatic principle extends this difficulty to the propositional attitudes of speakers. (1) A sincere, reflective, competent speaker who assertively utters S in a context C says (or asserts), perhaps among other things, what S says in C. This principle reflects an incipient relational analysis of the attitude of saying, or asserting—an analysis that sees it as a relation between speakers and things which serve as the semantic contents of sentences. Once this analysis is accepted, it is a short step to view propositional attitude reports in accord with (2) and (3). (2) An individual i satisfies x says (asserts) that S relative to a context C iff i stands in a certain relation R to the semantic content of S in C. (3) An individual i satisfies x v’s that S (where v = ‘believes’, ‘knows’, ‘proves’, ‘expects’, etc.) relative to a context C iff i stands in a certain relation R′ to the semantic content of S in C. But now our difficulties are surely unmanageable. Let us characteriz distribution over conjunction and closure under necessary consequence a follows: Distribution over Conjunction If an individual i satisfies x v’s that P&Q relative to C, then i satisfies  x v’s that P and x v’s that Q relative to C. (For example, anyone who asserts that P&Q asserts that P and asserts that Q.) Closure under Necessary Consequence If an individual i satisfies x v’s that P relative to C, and if every possible world in which P is true relative to C is a possible world in which

Direct Reference • 35

Q is true relative to C, then i satisfies x v’s that Q relative to C. (For example, anyone who asserts that P asserts everything that necessarily follows from P.) The second main difficulty with our combined truth-conditional and propositional attitude conception of semantics is that it equates distribution of a propositional attitude verb over conjunction with closure of the attitude under necessary consequence. For if Q is a necessary consequence of P, then the set of metaphysically possible worlds in which  P&Q is true is the same as the set of worlds in which P is true. Given the identification of truth conditions with semantic content, this means that their semantic contents are the same. But then, a relational semantics of propositional attitude reports together with distribution over conjunction will yield closure under necessary consequence. The problem is that for many propositional attitude verbs distribution over conjunction is a fact whereas closure under necessary consequence is not. My four year old son Greg has said many things, and whenever he says that P&Q he says that P and he says that Q. However, there are lots of necessary consequences of things he has said that he has left unasserted, for example that 2 to the ninth = 512, that first order logic is complete but undecidable, and that stones are made up of molecules. A third difficulty with our semantic conception takes this problem one step further. The same considerations that lead to the view that beliefs and assertions are closed under necessary consequence lead to the view that no one has ever believed or asserted anything that couldn’t have been true (in any metaphysically possible world). Since every Q is a necessary consequence of an impossible P, anyone who believes or asserts what P expresses believes or asserts everything. And surely, no one ever has, or could have, done that. The semantic assumptions that lead to these difficulties can be summarized as follows: A1a. The semantic content of a sentence (relative to a context) is the collection of circumstances supporting its truth (as used in the context). A1b. The collection of circumstances supporting the truth of a sentence (as it is used in a context) = the set of metaphysically possible worlds in which it is true (relative to the context). A2. Propositional attitude sentences report relations to the semantic contents of their complements—i.e., an individual i satisfies  x v’s that S (relative to a context C) iff i bears R to the semantic content of S (relative to C).

36 • Essay One

A3. Many propositional attitude verbs, including ‘say’, ‘assert’, ‘believe’, ‘know’, and ‘prove’ distribute over conjunction. Since these assumptions lead to unacceptable results, one or more of them must be rejected. The crucial assumptions are A1 and A2, which, in turn, are direct descendants of the two conceptions of semantics mentioned earlier. A1 (a and b) represent the truth-conditional conception, with metaphysically possible worlds taken as truth conditions. A2 represents the propositional attitude conception, with the relational analysis of ‘say’, and ‘assert’ extended to propositional attitude reports generally. The need to give up one or the other of these assumptions makes it necessary to rethink the fundamental issues underlying these semantic conceptions. I will focus on the truth-conditional conception. Much of the support it has enjoyed comes from the familiarity of the possible worlds machinery plus the fact that the semantic content of a sentence (relative to a context) should determine the possible worlds in which it is true. However, there is a big difference between admitting that semantic content determines such truth conditions and claiming that it should be identified with them. What we need is some conception of semantics in which the content of a sentence determines, but is not determined by, the metaphysically possible worlds in which it is true. There are two main ways in which such a conception might be developed. One way is to retain the basic assumption, A1a, of the truthconditional conception, while rejecting the characterization of truth conditions, or truth-supporting circumstances, as metaphysically possible worlds. The idea is to try and find some more finely grained circumstances that will distinguish among sentences true in the same worlds. The second way in which an appropriate semantic account might be developed is to give up A1a, thereby abandoning the fundamental tenet of the truth-conditional conception. In its place, one might substitute a conception of semantic contents as complex objects that encode much of the structure of the sentences that express them, and that determine sets of truth-supporting circumstances, without being identified with them. In what follows, I will argue for the second approach. The heart of my argument involves the interaction of propositional attitudes with the phenomenon of direct reference. Let us say that a singular term is directly referential if its semantic content relative to a context (and assignment of values to variables) is its referent relative to the context (and assignment). Variables are the paradigm examples of such terms. In recent years, a number of arguments have been given for treating names and

Direct Reference • 37

indexicals as directly referential as well. Later, I will show how this view can be defended against certain objections based on the behavior of such terms in propositional attitude ascriptions. To begin with, however, I wish to note the destructive consequences it has when added as a fourth assumption to A1–A3. A4. Names, indexicals, and variables are directly referential. This expanded set of assumptions has a number of clearly unacceptable consequences. Suppose, for example, that Mary assertively utters (4a) while pointing at me. On the assumptions we are considering, she cannot correctly be reported to believe, or to have said, that I am David Kaplan.1 (4) a. He is David Kaplan. (Said pointing at Scott.) b. Mary says (believes) that he (Scott) is David Kaplan. The reason for this is that the semantic content of the complement sentence, relative to the context, is taken to be the set of metaphysically possible worlds in which two distinct objects are absolutely identical with one another—that is, the empty set. But then the third difficulty noted above—the impossibility of saying or believing the impossible—comes into play, ruling out the possibility that Mary said or believed what she seemed to say and believe. The same problem arises in a variety of cases, including those in (5). (5) a. John says (believes) that Ruth Marcus is Ruth Barcan’s sister. b. Martin says (believes) that this table is made up of atomic particles with properties P, Q, and R. (Where it is later discovered that nothing made of such particles could be a table.) The significance of these difficulties is not that they mar an otherwise unproblematic account of the attitudes. As we have seen, the conjunction of A1–A3 is problematic in its own right. Nevertheless, the difficulties arising from the addition of A4 are special. I will argue that these difficulties are intractable for theories that identify semantic contents of sentences with sets of truth-supporting circumstances. Although many of the problems encountered in standard, truth-theoretic accounts of the attitudes can be avoided by substituting fine-grained circumstances for metaphysically possible worlds, those posed by names and indexicals cannot. Not only do these problems resist such treatment, they remain even when assumptions A2, A3, and A4 are weakened substantially. 1 I assume here, and in what follows, that the semantic content of the complement sentence in a propositional attitude ascription is compositionally determined from the semantic contents of its parts.

38 • Essay One

In effect, directly referential singular terms can be used to show that semantic contents of sentences (relative to contexts) cannot be sets of truthsupporting circumstances, no matter how fine-grained. The reason for this is that such terms require the introduction of structure into semantic contents. After establishing this, I will consider two different ways in which such structure might be constructed—one based on a modified version of the truth-theoretic approach, the other based on the introduction of structured, Russellian propositions. Although considerations involving directly referential singular terms are insufficient to decide between these alternatives, I shall argue that additional factors favor the Russellian approach. Thus, the end result is an argument for an expanded conception of semantics that includes Russellian propositions as semantic contents of sentences, over and above standard, truth-theoretic intensions and extensions.

2. Let us begin with the strategy of substituting fine-grained truth-supporting circumstances for metaphysically possible worlds. These circumstances can be thought of as arising from the relaxation of certain constraints that hold for such worlds. Taking a cue from Carnap’s notion of a state description, we can describe these constraints in terms of their role in constructing a semantics for a language L. Let D be the set of individuals L is used to talk about, and B be the set of properties expressed by simple predicates of L plus their complements.2 Let us say that a C-description is a set each of whose members consists of an n-place property plus an n-tuple of objects drawn from D (for variable n). A C-description X is complete iff it contains a complete assignment of objects to properties—i.e., iff for every n-place property P in B, and every o1, . . . ,on in D, either [P,o1, . . . ,on] is a member of X or [[− P],o1, . . . ,on] is a member of X, where [− P] is the complement of P. A C-description X is consistent iff no two of its members are negations of one another—i.e., iff for every n-place property P in B, [P, o1, . . . ,on] is a member of X only if [[− P],o1, . . . ,on] is not a member of X. A C-description is metaphysically possible only if it is metaphysically possible for the objects mentioned in the description to ( jointly) instantiate the properties they are paired with in the description. 2 For example, the properties of being human and of not being human are complements of one another. I will assume that every property has a (unique) complement and that P is the complement of Q iff Q is the complement of P.

Direct Reference • 39

For present purposes, truth-supporting circumstances might either be identified with C-descriptions, or be taken to correspond to them. The classifications “complete,” “consistent,” and “metaphysically possible” can then be applied to circumstances. Metaphysically possible worlds are truth-supporting circumstances that are metaphysically possible, complete, and consistent. Suppose the first of these constraints is relaxed, while we retain the second and third. This allows truth-supporting circumstances corresponding to every consistent and complete C-description. Thus, we allow metaphysically impossible circumstances in which Ruth Marcus is Ruth Barcan’s sister, 2 to the ninth is not 512, and I am identical with David Kaplan (‘=’ being treated as a simple, nonlogical predicate in the object language). In effect, we substitute what might be called “logically possible” worlds or circumstances for “metaphysically possible” worlds or circumstances. However, the structure of the semantic theory remains the same as before. It continues to be a recursive characterization of truth relative to a context and circumstance, with the recursive clauses retaining their standard specifications. The semantic content of a sentence relative to a context is identified with the set of circumstances in which it is true. But since these circumstances are more finely grained than metaphysically possible worlds, we no longer have the results that metaphysically equivalent sentences have the same semantic content, that distribution of a propositional attitude verb over conjunction requires closure of the attitude under metaphysically necessary consequence, or that no one can believe or assert the metaphysically impossible. In this way, substitution of A1b′ for A1b might be seen as alleviating the original difficulties with A1–A4. A1b′. The collection of circumstances supporting the truth of a sentence (relative to a context) = the set of logically possible worlds in which it is true (relative to the context). It does, of course, remain true on this view that logically equivalent sentences have the same semantic content, that distribution of a propositional attitude verb over conjunction requires closure of the attitude under logical consequence, and that no one can believe or assert the logically impossible. However, with another weakening of the constraints even these results can be avoided. Suppose we give up the requirement that truth-supporting circumstances be complete. Instead we allow circumstances to correspond to (and, in effect, be exhausted by) any consistent C-description. Such circumstances are more like “logically possible facts” than “logically possible worlds.” For example, one such circumstance may consist entirely of my being human.

40 • Essay One

The introduction of partial circumstances has import for certain logical constructions, most notably negation. In order to make semantic use of partiality, one must distinguish between it not being the case that in C an individual o has the basic property P, and it being the case that in C, o has the property of not being P. The latter is a truth-supporting circumstance for the negation of the atomic sentence that predicates P of o; the former is not. Full-fledged negation, applied to sentences of arbitrary complexity, as well as related constructions like material implication, raise complications that we need not go into. However, other constructions are straightforward. For example, the recursive clauses governing conjunction, disjunction, and existential generalization are exactly those used in standard, truth-theoretic accounts. The semantic content of a sentence relative to a context is, as usual, the set of circumstances supporting its truth, as used in the context. However, since circumstances are partial, the semantic contents of logically equivalent sentences are no longer identified. For example, the content of (6a) is not the same as the content of (6b), because the former includes “facts” that are, so to speak, silent about radioactivity. (6) a. Plymouth Rock is in Massachusetts. b. Plymouth Rock is in Massachusetts & (Plymouth Rock is radioactive v Plymouth Rock isn’t radioactive). This is significant, since, it might be argued, a person lacking the concept of radioactivity might believe that which is expressed by (6a) without believing that which is expressed by (6b). Certainly, it would seem that someone could assert the former without asserting the latter. One way of accounting for this within the framework of A1–A4 is to substitute A1b′′ for A1b′′. A1b′′. The collection of circumstances supporting the truth of a sentence (relative to a context) = the set of logically possible facts that would make it true (as used in the context). This strategy is followed by Jon Barwise and John Perry in their book Situations and Attitudes (1983). However, they take it one step further, allowing truth-supporting circumstances to be inconsistent, as well as incomplete and metaphysically impossible. If one ignores complications involving time, tense, and spatiotemporal location, one can take their “abstract situations” to be arbitrary C-descriptions.3 Allowing these circumstances to be inconsistent, and substituting A1s for A1b′′, makes it possible to correctly 3 The idea of thinking of abstract situations as resulting from relaxing constraints on Carnapian state descriptions was suggested to me by David Kaplan.

Direct Reference • 41

characterize certain agents as believing and asserting contradictions—e.g., as believing and asserting that London is pretty and London is not pretty. A1s. The collection of circumstances supporting the truth of a sentence (relative to a context) = the set of abstract situations which would make it true (as used in the context). Logically complex constructions are characterized along familiar truththeoretic lines. For example, we have: (7) a. The semantic content of a conjunction (relative to a context) is the intersection of the semantic contents of the conjuncts (relative to the context). b. The semantic content of a disjunction (relative to a context) is the union of the semantic contents of the disjuncts (relative to the context). c. The semantic content of an existential generalization, For some x: Fx, (relative to a context) is the set of circumstances E such that for some object o in E, o “is F” in, or relative to, E (and the context).4 d. The semantic content of F[an x: Gx] (relative to a context) is the set of circumstances E such that for some object o in E, o “is G” and o “is F” in, or relative to, E (and the context). e. The semantic content of F[the x: Gx] (relative to a context) is the set of circumstances E such that for exactly one object o in E, o “is G” in, or relative to, E (and the context); and, moreover, o “is F” in, or relative to, E (and the context). The invariance of these principles across different choices of truthsupporting circumstances reflects the fact that no matter what one’s conception of circumstances, the circumstances that make a conjunction true are those that make the conjuncts true; the circumstances that make a disjunction true are those that make either disjunct true; and so on. Indeed we may take the principles in (7) to be partially constitutive of the view that the semantic content of a sentence consists in the circumstances that support its truth. As such, they may be regarded as corollaries of assumption A1a. There is, then, a whole range of possible theories within the standard, truth-conditional framework that adopt the same basic approach to the problems posed by various kinds of propositional attitudes. The central 4 Thus, the content of an existential generalization is a superset of the contents of instances from which it follows. It should be noted that no formal treatment of existential quantification is provided in Barwise and Perry (1983). Nevertheless, (7c) accords well with the leading ideas of that work. (7a), (7b), (7d), and (7e) are explicitly endorsed.

42 • Essay One

idea is to relax the constraints on truth-supporting circumstances. This results in more finely grained semantic contents being attached, in the first instance, to atomic sentences. Logically complex constructions are given the usual recursive treatment, resulting in semantic contents for complex sentences along the lines of (7). This approach can be seen as an attempt to save the truth-conditional conception of semantic content, while countenancing direct reference and continuing to take semantic contents of sentences to be objects of propositional attitudes. Although not without plausibility, it is, I believe, fundamentally flawed. Its chief virtue is its recognition that if assumptions A2, A3, and A4, plus an elementary principle of compositionality,5 are to be retained, then semantic contents must be more fine-grained than sets of metaphysically possible worlds. Its chief error is its failure to recognize that if these assumptions are retained, then no conception of truth-supporting circumstances validating (7) can do the job, no matter how fine-grained.

3. A number of different arguments can be used to show this. For example, consider (8). (8) a. The ancients believed (asserted) that ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Phosphorus. b. The ancients believed (asserted) that ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Hesperus. (From A2, A4, and compositionality in the complement) c. The ancients believed (asserted) that ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Hesperus and for some x, ‘Hesperus’ referred to x and ‘Phosphorus’ referred to x. (From A1a, and A2) d. The ancients believed (asserted) that for some x, ‘Hesperus’ referred to x and ‘Phosphorus’ referred to x. (Where the quantifier is inside the scope of the propositional attitude verb.) (From A3) 5 The compositional principle I will appeal to may be understood as applying to sentences free of quotation and opacity-producing operators.

If S and S' are nonintensional sentences with the same grammatical structure, which differ only in the substitution of constituents with the same semantic contents (w.r.t. their respective contexts and assignments of values to variables), then the semantic contents of S and S' will be the same (w.r.t. those contexts and assignments). This principle is presupposed in standard versions of truth-conditional semantics, and is itself a corollary of assumption A1a.

Direct Reference • 43

Since (8d) is tantamount to the claim that the ancients believed and asserted that the terms ‘Hesperus’ and ‘Phosphorus’ were coreferential, it is false. Since (8a) can be regarded as true, at least one of the principles used in going from (a) to (d) must be rejected. The first thing to note is that these principles do not include A1b, A1s, or any other specific characterization of truth-supporting circumstances. The only use made of truth-supporting circumstances was the appeal to (7a) and (7c) in the move from (b) to (c) in the argument. Since these principles are corollaries of A1a, acceptance of the other assumptions in the argument requires rejection of the claim that the semantic content of a sentence (relative to a context) is the set of circumstances supporting its truth (as used in the context). The same point can be made using definite descriptions instead of existential quantification. For example consider (9). (9) a. y believes (asserts) that Hesperus = the x:Fx and Phosphorus = the x:Gx. b. y believes (asserts) that Hesperus = the x:Fx and Hesperus = the x:Gx. (From A2, A4, and compositionality in the complement) c. y believes (asserts) that Hesperus = the x:Fx and Hesperus = the x:Gx and the x:Fx = the x:Gx. (From A1 and A2) d. y believes (asserts) that the x:Fx = the x:Gx. (Where the descriptions are used attributively and are within the scope of the propositional attitude verb.) (From A3) The move from (b) to (c) is justified if every circumstance supporting the truth of the complement of (b) supports the truth of the complement of (c). One gets this if circumstances are metaphysically possible worlds, since any world in which o is identical with o′ and o′′ is a world in which o′ and o′′ are identical. However, there is no need to rest the case on special assumptions about circumstances, or identity. By recasting the example one can make use of the semantics (7e) for definite descriptions to construct an argument that applies to all the theories in section 2. One simply starts with (9a′) instead of (9a). (9a″) y believes (asserts) that Hesperus = the x:Fx and Phosphorus = the x:Gx and the x:Fx = the x:Fx and Hesperus = the x such that Hesperus = x. It follows from (7a) that a circumstance E will support the truth of the complement of (9a′) iff it supports the truth of each of its conjuncts. It follows from (7e) that E will support the truth of the final conjunct iff

44 • Essay One

there is exactly one object o such that Hesperus = o in E. Since Hesperus is Phosphorus, this means that o must be both the unique F-er in E and the unique G-er in E. The third conjunct requires that o = o in E. This guarantees that E will be a member of the semantic content of the complement of (9d). Thus, A2, A3, A4, and a principle of compositionality, allow one to derive (9d) from (9a′), no matter how finely grained one takes truth-supporting circumstances to be.6 Since (9d) may be false even when (9a′) is true, acceptance of A2, A3, A4, and the compositionality principle requires rejection of A1a. A more startling illustration of this conclusion can be constructed using the examples in (10). (10) a. Mark Twain = Herman Melville and Samuel Clemens = Stephen Crane. b. Mark Twain = the x such that Mark Twain = x. (a) is an embarrassment to standard treatments of the attitudes (encompassing A2–A4) in which truth-supporting circumstances are taken to be metaphysically possible worlds. Since its semantic content in such systems is the empty set, everything is a semantic consequence of it. Thus, that which it expresses cannot be believed or asserted. One of the virtues of systems that relax constraints on truth-supporting circumstances is that they avoid this embarrassment. In such systems the semantic content of (10a) is a nonempty set of circumstances in which three distinct individuals are identified. Although such circumstances are metaphysically impossible, they are regarded as semantically legitimate, and hence are available for the construction of semantic contents. Thus, it is perfectly possible, in a system like that of Barwise and Perry (1983), for a person to believe or assert that which is expressed by (10a). Belief or assertion of that which is expressed by (10b) is unproblematic on any account. However, now consider their conjunction, (10c). (10c) Mark Twain = Herman Melville and Samuel Clemens = Stephen Crane and Mark Twain = the x such that Mark Twain = x. In order for a circumstance E to be a member of the semantic content of this sentence, E must be a member of the semantic content of each conjunct. In order for E to be a member of the semantic content of the first two conjuncts, it must be the case that in E Mark Twain is identified with two distinct individuals. But now E cannot be a member of the semantic content of the third conjunct, since, by (7e), that conjunct requires that Mark Twain be identified with only one object. The semantic content of 6 So long as they validate (7a) and (7e). This continues to hold when any two-place relation replaces identity.

Direct Reference • 45

(10c) is, therefore, the empty set. Thus, the problems posed by (10a) for theories embracing the original A1–A4 are reproduced by (10c) for theories that substitute finer-grained truth-supporting circumstances for metaphysically possible worlds.7 Although this example is particularly graphic, the basic difficulty is extremely general. It is repeated in (11), where (b) is derived from (a) using the semantics, (7d), for indefinite descriptions, and in (12), where a similar derivation uses material implication.8 (11)

a. x believes (asserts) that Mark Twain wrote the greatest American novel and Samuel Clemens was an ignorant illiterate. b. x believes (asserts) that an ignorant illiterate wrote the greatest American novel. (Where the indefinite description is attributive and inside the scope of the propositional attitude verb.) (12) a. x believes (asserts) that Mark Twain is F and if Samuel Clemens is F then S. (Where F is any predicate and S is any sentence.) b. x believes (asserts) that S.

The difficulty common to all these cases is, I suggest, not due to special assumptions about particular constructions (existential quantification, definite descriptions, indefinite descriptions, conjunction, material implication, etc.). Rather, the general assumptions A1a, A2, A3, and A4 (plus compositionality in the complements of propositional attitude verbs) are 7 One might, of course, try to avoid this result by tampering with the semantics of definite descriptions. For example, one might try substituting the unlovely (7e') for (7e).

(7e') The semantic content of F[the x: Gx]  (relative to a context) is the set of circumstances E such that there is at least one object o in E which is both an F-er and a G-er in E; and moreover, for any other object o', if o' is a G-er in E, then o = o' and o'= o in E, and, more generally, o and o' have exactly the same properties (and stand in the same relations to the same objects) in E. One drawback of this from the point of view of a system like that of Barwise and Perry (1983) is that it gives up the view that definite descriptions determine partial functions from circumstances to objects that uniquely satisfy the descriptions in those circumstances. Since this feature of definite descriptions is used extensively in Barwise and Perry (1983), it is not clear that Barwise and Perry would be willing to replace (7e) with (7e'). In any case, such a move would do nothing to remove the problem posed by (9a'). 8 The derivation in (12) depends on the assumption that if E supports the truth of P and also supports the truth of P → Q, then E supports the truth of Q. This will hold if truthsupporting circumstances are logically possible worlds and E supports the truth of a material conditional whenever it supports the truth of the consequent or fails to support the truth of the antecedent. When circumstances are allowed to be partial and inconsistent, the situation is no longer straightforward. For example, the system in Barwise and Perry (1983) provides no treatment of conditionals, and so is not subject to the argument based on (12).

46 • Essay One

jointly incompatible with facts about propositional attitudes and propositional attitude ascriptions. In short, we have established (13). (13) If direct reference is legitimate and (some) propositional attitude verbs have a relational semantics (A4 plus A2), then (assuming compositionality and distribution over conjunction) the semantic contents of sentences relative to contexts cannot be sets of truth-supporting circumstances (that validate (7)). This way of putting the matter is, of course, not neutral, since it suggests that the assumption to be rejected is A1a. This suggestion can be supported by showing that the remaining assumptions are both stronger than needed to refute A1a and more plausible than they might initially appear. 4. First consider A4. The arguments in section 3 all involve proper names. Thus, one response to them might be to give up the claim that names are directly referential, thereby blocking substitution of coreferential names in propositional attitude ascriptions. It is important to note that this response is insufficient, since, in each case, the problem can be re-created using other terms. For example, so long as direct reference is retained for demonstratives, A1a, A2, A3, and compositionality will allow one to derive the false (14b) from the potentially true (14a). (14) a. The ancients believed (asserted) that their such-and-such utterance referred to this (pointing in the morning to Venus) and (speaking very slowly) their so-and-so utterance referred to that (pointing in the evening to Venus). b. The ancients believed (asserted) that for some x, their suchand-such utterance referred to x and their so-and-so utterance referred to x. The same point can be made using variables in place of names and indexicals. (15) a. There is a planet x which is seen in the morning sky and a planet y which is seen in the evening sky and the ancients believed that x was seen in the morning sky and y was seen in the evening sky. (∃x: Px&Mx) (∃y: Py&Ey) (a believed that (Mx&Ey)) b. The planet seen in the morning sky is the planet seen in the evening sky. the x:(Px&Mx) = the y:(Py&Ey)

Direct Reference • 47

(15a) is true iff there is an assignment f which assigns a planet seen in the morning sky to ‘x’ and a planet seen in the evening sky to ‘y’ such that the open belief sentence is true with respect to f. From (15b) it follows that the referents of ‘x’ and ‘y’ with respect to f are identical. But now A1–A4 can be applied as before to derive the false (15c–d) from the true (15a–b).9 (15) c. There is a planet x and a planet y such that the ancients believed the following: that x was seen in the morning sky and y was seen in the evening sky and there was something which was (both) seen in the morning sky and seen in the evening sky. (∃x: Px) (∃y: Py) (a believed that ((Mx&Ey) & ∃z(Mz&Ez))). d. The ancients believed that there was something which was (both) seen in the morning sky and seen in the evening sky. a believed that ∃z(Mz&Ez). Thus, if direct reference is the source of the difficulty, it must be banned entirely—for names, indexicals, and variables. But this is implausible; the arguments for it are too strong, and there are too many cases (where the words of the speaker differ systematically from those of the agent of the attitude) in which it is instrumental in capturing clear semantic intuitions. There is, however, another way in which one might try to block the problematic arguments. Each of them relies on assumptions—A2, compositionality, and some version of direct reference—that jointly legitimate the substitution of coreferential terms in propositional attitude ascriptions. It might be thought that such substitution is the source of the problem. As against this, it is worth noting that the difficulty can be re-created without appealing to substitutivity, or the assumptions that give rise to it. Instead of relying on semantic analyses of propositional attitude statements one can invoke principles underlying our practice of reporting propositional attitudes and ascribing them to individuals. Why, for example, do we ascribe to the ancients the belief and assertion that Hesperus

9

Here, A1–A4 must be understood as relativizing semantic content and associated truth conditions to both contexts and assignments of values to variables. A1a. The semantic content of a sentence (relative to a context C and assignment f) is the collection of circumstances supporting its truth (as used in C with respect to f). A2. An individual i satisfies x v’s that S (relative to C and f) iff i bears R to the semantic content of S (relative to C and f). A3. For many propositional attitude verbs (including ‘say’, ‘assert’, and ‘believe’) if i satisfies x v’s that P&Q (relative to C and f), then i satisfies x v’s that P and x v’s that Q (relative to C and f).

48 • Essay One

was visible in the evening, while being reluctant (at least initially) to ascribe to them the belief and assertion that Phosphorus was visible in the evening? Probably because they assertively uttered sentences whose English translation is ‘Hesperus is visible in the evening’, but refused to assertively utter (and indeed dissented from) sentences whose English translation is ‘Phosphorus is visible in the evening’. These examples suggest (if we focus on indexical-free sentences, and ignore complications involving time and tense) the following principles of propositional attitude ascription. (16) a. If a competent speaker x of a language L sincerely and reflectively assents to (or assertively utters) an indexical-free sentence s of L, and if p is a proper English translation of s, then x satisfies y believes that p. (Note that this covers the case in which L = English and s = p.) b. If a sincere, reflective, and competent speaker x of a language L assertively utters an indexical-free sentence s of L, and if p is a proper English translation of s, then x satisfies  y says (asserts) that p. These principles are, of course, modeled after Kripke’s principles of (weak) disquotation and translation.10 With them we can derive the conclusion that Kripke’s bilingual speaker Pierre believes and asserts both that London is pretty and that London is not pretty. The former follows from his sincere and reflective utterance of ‘Londres est jolie’, plus (16) and an elementary truth of translation. The latter follows from his equally sincere and reflective utterance of ‘London is not pretty’, plus either (16) alone, or (16) in conjunction with homophonic translation. It seems to me that these ascriptions to Pierre are correct. It is, of course, striking that Pierre’s beliefs and assertions should be contradictory without his having made any mistake in logic or reasoning. However, this just shows that in certain cases one may be in no position to determine the consistency of one’s statements and beliefs. The point is particularly obvious in the case of what is said or asserted. Imagine Pierre on the telephone talking to a friend in Paris. During the course of the conversation he assertively utters ‘Londres est jolie’. After hanging up the phone he says ‘London is not pretty’ to a visitor who asks his opinion of the city he lives in. What has Pierre said? Clearly, he has said both that London is pretty (to his friend) and that London is not pretty (to the visitor). Now consider a slight extension of the example. Suppose that there are a number of Frenchmen in London in the same linguistic and epistemic 10

Kripke (1979).

Direct Reference • 49

situation as Pierre. When together, they converse with one another in French—standard French plus one addition. Since they are unaware that ‘Londres’ names the city they live in, they use the name ‘London’ for that purpose. One day Pierre assertively utters ‘Londres est jolie et London n’est pas jolie’. I, an English speaker, am asked to report what he said. Since Pierre is competent in his own dialect, I can appeal to (16). Since his dialect is one in which both ‘London’ and ‘Londres’ are properly translated into English as ‘London’, I can report that he said (asserted) that London is pretty and London isn’t pretty. To avoid puzzling my audience, I will, of course, say more than this. However, the initial report is surely correct. In certain cases two words in one language do have the same translation into a second language (e.g., ‘Peking’ and ‘Bejing’ in English); and assertive utterances by normal, competent speakers can be reported in indirect discourse of the second language.11 This fact can be used to reconstruct the arguments of section 3 without appealing to direct reference, compositionality, or substitutivity at all. In the case at hand, we have used (16b) plus a truth of translation to establish (17). (17) Pierre said (asserted) that London is pretty and London is not pretty. To derive (18), (18) Pierre said (asserted) that London is pretty and London is not pretty and for some x, x is pretty and x is not pretty, we need only appeal to corollaries (7a) and (7c) of A1a, plus the following weakened version of A2. A2'.

An individual i satisfies x v’s that S (relative to a context C) iff i bears a certain relation R* to the pair consisting of the content of S (relative to C) and the character of S (i.e., the function from contexts to contents that represents the meaning of S).12

(19) follows from (18) by A3. (19) Pierre said (asserted) that for some x, x is pretty and x is not pretty. 11

See Kripke (1979, 268 and n. 42) for relevant discussion. A2' is a consequence of A2, but not vice versa. If A2 is true, then R* in A2' can be taken to be a relation that an individual bears to a content-character pair, , iff i bears the relation R of A2 to the content y. However, A2' might be true even if substitution of complement sentences with the same content sometimes failed to preserve truth-value, in which case A2 would be false. 12

50 • Essay One

But (19) is false—Pierre didn’t assert the proposition that something is both pretty and not pretty. Thus, we have another reductio of A1a, this time from a considerably weakened set of premises. Similar reductios can be constructed corresponding to each of the arguments in section 3.13 However, the premises are still stronger than they need to be. Although A3 is useful in deriving obviously false conclusions, it is not strictly necessary. (8c), (8c′), (9c), (9c′), (15c), and (18) are all false, and can be derived without A3.14 (8c′) The ancients believed (asserted) that ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Phosphorus and for some x, ‘Hesperus’ referred to x and ‘Phosphorus’ referred to x. (9c′) y believes (asserts) that Hesperus = the x:Fx and Phosphorus = the x:Gx and the x:Fx = the x:Gx. Even A2, and its weakened counterpart A2', may give a misleading impression of strength. As presently formulated, they ignore one possible type of semantic information—to whit, information fixing the referent of a name as a matter of linguistic convention. I suspect that arabic numerals are names that carry such information.15 Some might hold that ‘Hesperus’ and ‘Phosphorus’ are too.16 If they are, then the weakened principle, A2*, will block substitution of one for the other in propositional attitude ascriptions.17 A2*. An individual i satisfies x v’s that S (relative to a context C) iff i bears a certain relation R** to the triple consisting of the content of S (relative to C), the character of S, and an n-tuple of properties [P1, . . . ,Pn], where Pi fixes, as a matter of linguistic convention, the referent of the ith name in S. However, such a move will not block the reductio of A1a. First, not all proper names have conventionally associated reference-fixing properties. Second, as Kripke has shown, variants of the Pierre case can be constructed in which the names ‘London’ and ‘Londres’ are associated with the same properties (provided they are not “purely qualitative”).18 Finally, substitu13

The basis for these reductios is partially prefigured in Kripke (1979, 257–58, 262). In the case of (10), one can use (7e) to derive x believes that S&P from x believes that S, where S is (10c) and P is any sentence at all. See, however, the qualification in note 7. 15 See Richard (1986). 16 See Kripke (1979, n. 43) for relevant discussion. 17 I am indebted to Joseph Almog for suggesting that I reconstruct my argument using this sort of weakening of A2. 18 Kripke (1979, 260–63). If the properties are not required to pick out a unique referent (e.g., if the property of being a famous Roman is the one associated with both ‘Cicero’ and ‘Tully’), then problematic substitution will go through even when the property is “purely qualitative.” See Kripke (1979, n. 9). 14

Direct Reference • 51

tion of one term for another is not always required for refutations of A1a. Suppose, for example, that ‘Hesperus’ and ‘Phosphorus’ share the same object as content and the same constant function from contexts to that object as character, but differ in reference-fixing properties. Although A2* will then block the derivation of (8b) and (8c) from (8a), it will still allow the derivation of (8c′). (The same goes for (15).)19 Results like these suggest that the reductio of A1a cannot be blocked by any plausible weakening of the subsidiary premises used in the original argument. It is true that those premises jointly give rise to some surprising, and initially counterintuitive, results involving substitution in propositional attitude ascriptions. However, the reductio can be re-created (in a variety of ways) even when those results are avoided, or minimized. A final illustration of this point is provided by the following example: Professor McX, looking through the open back door of the faculty lounge, sees Y walking down the hall and says to a visitor, “He (pointing to Y) is a professor in the department.” A few seconds later Y passes by the front door, and McX says, “He, (pointing to Y again) is a graduate student in the department.” Although McX doesn’t realize that he has pointed twice to the same individual, Y, who has overheard the remarks, can correctly say, “McX said both that I am a professor in the department and that I am a graduate student in the department.” Developing the example further, we can have McX conjoin his remarks: (20) Who is in the department? Let me see. He (pointing to Y as he passes the back door) is a professor in the department and (turning) he (pointing to Y as he passes the front door) is a graduate student in the department. 19 There is another respect in which A2 and its weakened counterparts may give a misleading impression of strength. They may suggest that the arguments against A1a rely crucially on assumptions about the semantics of sentences of the form x v’s that S. In fact, such sentences are dispensable. There are two leading ideas behind the various versions of A2. The first is that propositional attitudes like saying, asserting, and believing are relations to things that are said, asserted, and believed. The second is that these things are said, or semantically expressed, by sentences. If these ideas are correct, then the arguments against A1a can be reconstructed— either directly, in terms of what sentences say, or indirectly, using (1) and, if desired, A3' to derive conclusions about what speakers say.

A3'.

If x says (or asserts) that which is said (expressed) by a conjunction in a context C, then x says (or asserts) that which is said (expressed) by each conjunct in C.

Using these principles, one can derive the incorrect conclusion that x has said (or asserted) that which is expressed by For some y Ryy from the premise that x has assertively uttered  R(Hesperus, Phosphorus) or R(Londres, London).

52 • Essay One

On the basis of McX’s remark, Y says: (21) McX said that I am a professor in the department and I am a graduate student in the department. Y’s assertion is unexceptionable. Unlike some other examples we have considered, this one does not require the creation of an unusual situation; it does not involve attributing conflicting statements (or beliefs) to an otherwise rational agent; nor does it raise the suspicion that adherence to otherwise plausible principles forces us to accept a counterintuitive result. Whatever semantic analysis of propositional attitude ascriptions turns out to be correct, Y’s report is one that we want, pretheoretically, to come out true. This is not the case with (22) (where the quantifier is understood as being inside the scope of the propositional attitude verb). (22)

a. Professor McX said (asserted) this: that there is at least one x such that x is a professor in the department and x is a graduate student in the department and I am a professor in the department and I am a graduate student in the department. b. Professor McX said (asserted) that there is at least one x such that x is a professor in the department and x is a graduate student in the department.

These reports are clearly not true.20 If this is correct, then the problem for A1a is obvious. Corollaries (7a) and (7c) of that principle characterize the complements of (21) and (22a) as having the same content (with respect to the context). But then there will be no semantic value (content, character, or reference-fixing properties) differentiating them. As a result, virtually any relational semantics of assertion-ascriptions (e.g., A2, A2', A2*) will assign (21) and (22a) the same truth-value. A3 will then extend this error to (22b). Since these results are unacceptable, while relational treatments of assertion and other attitudes remain plausible, A1a should be rejected.

5. We have just seen that the impossibility result of section 3 can be reproduced using A1a together with sets of auxiliary premises considerably weaker than the original A2, A3, A4, and compositionality. This 20 The same point could be made using other logical constructions—for example, indefinite descriptions—in place of existential quantification in the complement sentence.

Direct Reference • 53

constitutes an important reason for taking that result to be a reductio of the assumption that semantic contents of sentences are sets of truthsupporting circumstances. Another reason is that the supplementary assumptions of the original argument are themselves highly justified. This can be seen by looking at what many regard as the most questionable consequence of those assumptions, namely (23). (23) If i satisfies x v’s that S relative to a context C (and assignment f), and if t and t′ are names, indexicals, or free variables having the same referent relative to C (and f), then i satisfies x v’s that S′ relative to C (and f), where S′ arises from S by substituting one or more occurrences of t′ for occurrences of t.21 Many seem to think that counterexamples to this principle are easy to come by. In the case of belief ascriptions, they tend to be examples in which a competent speaker assents to S and I believe that S, while dissenting from S′ and I believe that S′, even though the latter arise from the former by substitution of names or indexicals with the same referent. Such cases tell against (23) only if assent and dissent are reliable guides to what is, and what is not, believed. However, dissent is not reliable in this way.22 A recent example of Mark Richard’s makes this point quite nicely. Consider A—a man stipulated to be intelligent, rational, a competent speaker of English, etc.—who both sees a woman, across the street, in a phone booth, and is speaking to a woman through the phone. He does not realize that the woman to whom he is speaking—B, to give her a name—is the woman he sees. He perceives her to be in some danger—a run-away steamroller, say, is bearing down upon her phone booth. A waves at the woman; he says nothing into the phone.

21

And where S is free of quotation and related constructions. The other main type of putative counterexample to (23) involves cases in which a competent speaker assents to (translations of) n is F and m is not F in a context in which n and m are coreferential names or indexicals. With (16a) plus translation one gets the result that the agent satisfies x believes that n is F and x believes that m is not F. Substitutivity then results in the ascription of contradictory beliefs, which is sometimes thought to be objectionable in light of the fact that the agent may have made no logical mistakes. However, Kripke’s example of puzzling Pierre shows that this is not a compelling criticism of (23). As we have seen, ascriptions of contradictory statements and beliefs can be derived from (16) plus translation, without any appeal to substitutivity. Moreover, the inconsistency is genuine. Kripke’s Pierre really does say and believe both that London is pretty and that London is not pretty. But then, if the statements and beliefs of even the best reasoner can be inconsistent without his being in a position to recognize it, the mere fact that the substitutivity principle can sometimes be used to arrive at ascriptions of such inconsistency does nothing to discredit it. 22

54 • Essay One

. . . If A stopped and quizzed himself concerning what he believes, he might well sincerely utter [3] I believe that she is in danger. but not [4] I believe that you are in danger. Many people, I think, suppose that . . . [these sentences] clearly diverge in truth value, [3] being true and [4] being false. . . . But [this] view . . . is, I believe, demonstrably false. In order to simplify the statement of the argument which shows that the truth of [4] follows from the truth of [3], allow me to assume that A is the unique man watching B. Then we may argue as follows: Suppose that [3] is true, relative to A’s context. Then B can truly say that the man watching her—A, of course—believes that she is in danger. Thus, if B were to utter [5] The man watching me believes that I’m in danger. (even through the telephone) she’d speak truly. But if B’s utterance of [5] through the telephone, heard by A, would be true, then A would speak truly, were he to utter, through the phone [6] The man watching you believes that you are in danger. Thus, [6] is true, taken relative to A’s context. But, of course, [7] I am the man watching you. is true relative to A’s context. [Which is not, of course, to say that A would accept it. My addition.] But [4] is deducible from [6] and [7]. Hence, [4] is true, relative to A’s context.23 In this example, Richard is concerned with substitution of coreferential indexicals. However, the argument seems to generalize. Suppose, for example, that A believes that Ruth Barcan is F is true relative to a context. A believes that I am F should then be true relative to a corresponding context in which Ruth Barcan (i.e., Ruth Marcus) is the agent (where F is free of first-person pronouns). Suppose, in fact, that Ruth utters the sentence in a conversation with someone who knows her as “Ruth Marcus.” It would seem that this person can truly report A believes that she (pointing at Ruth) is F, or even A believes that Ruth Marcus is F. Thus, substitution of one coreferential name or indexical 23

Richard (1983, 439–41).

Direct Reference • 55

for another preserves truth-value. Since there seems to be nothing special about this example, we have a general argument for (23).24 Why, then, does substitution so often provoke resistance? The answer, I think, has to do, at least in part, with the conversational purposes served by propositional attitude ascriptions. For example, suppose that Mary’s 24 Richard himself doesn’t go this far. For one thing, his semantics for belief ascriptions is silent about sentences containing proper names. More important, however, is a weakening of (23) involving complement sentences containing two or more occurrences of indexicals and/or free variables. Let t and t' be two such terms which have the same content relative to a context C and assignment f. According to Richard, if (i) is true relative to C, f, and a circumstance E, then (ii) must be true relative to C,f, and E, but not vice versa.

(i) x believes that S (t,t) (ii) x believes that S (t, t') Both this conclusion and the semantic system that leads him to it are, in my opinion, incorrect. Nevertheless, there is an important truth underlying Richard’s observations. This truth (first suggested to me by Nathan Salmon) is brought out by (iii). (iii) a. b. c. d.

x believes that t is not identical with t'. x believes that t is not identical with t. x believes that t is not identical with itself. x believes that t is non-self-identical.

It seems evident that (a) can be true when (d) is not. The reason for this is that believing the latter involves attributing the property of non-self-identity to an object, whereas believing the former does not. In light of this, one must block either the move from (a) to (b), or the move from (b) to (c) and (d). Richard selects the first of these. According to him, (iv) x believes that S is true only if the agent believes the proposition (semantic content) expressed by S (relative to the context and assignment). Moreover, the complements of (a) and (b) express the same proposition. Nevertheless, Richard holds that (a) can be true when (b) is false. The reason for this is that on his semantics a belief ascription of the form (iv) not only reports what proposition is believed, but also places constraints on the sentence acceptance of which is responsible for the agent’s belief. In the case of (b), the agent must hold the belief in virtue of accepting a sentence containing occurrences of directly referential terms with the same character. (If the account were extended to names it would be more natural to require two occurrences of the same term.) In the case of (a), this is allowed, but not required. One problem with this account is that it is too restricted. Whatever may be the case regarding ascriptions of the form (iv), some belief ascriptions express straightforward relations to propositions. (v) a. x believes the proposition expressed at the bottom of page 437. b. The proposition expressed at the bottom of page 437 is the proposition that P. c. Therefore, x believes the proposition that P. Given the admission that t is not identical with t' and t is not identical with t express the same proposition, one can use examples of the form (v) to reinstate the very problems that the nonrelational semantics of (iv) was designed to avoid.

56 • Essay One

neighbor, Samuel Clemens, is in the habit of soliciting her opinion of his manuscripts before sending them off to the publisher. Mary thinks they are wonderful, and regards Mr. Clemens (whom she knows only under that name) as a great writer. The question is, does she think that Mark Twain is a great writer? First consider a conversation the purpose of which is to determine Mary’s opinion of various authors. The conversational participants, who use the name ‘Mark Twain’ to refer to the author, want to know Mary’s opinion of him. I, knowing Mary’s situation, report “Mary thinks that Mark Twain is a great writer.” My remark seems perfectly acceptable. However, now consider a different conversation. Mary, who is a student, has just taken a written examination; and her teacher is explaining why she failed to get a perfect score. The teacher says, “Mary did a good job, but she didn’t know that Mark Twain is a writer.” In the context of this conversation, the teacher’s remark also seems acceptable. But how can it be? Surely it is not the case that Mary thinks that Mark Twain is a great writer, while not knowing that Mark Twain is a writer at all.25 To straighten this out, we need to distinguish between the proposition semantically expressed by a sentence relative to a context, and the information conveyed by an utterance of the sentence in a conversation. In the second conversation, the proposition semantically expressed by the propositional attitude ascription is false, even though the primary information conveyed by the utterance is true—namely, that Mary didn’t know that ‘Mark Twain is a writer’ is true; and hence was not able to answer exam questions of the sort, “What is Mark Twain’s profession?.” The teacher’s utterance seems acceptable because the main information it conveys is correct. This example brings out an important point about the relationship between propositional and sentential attitudes. Attitudes like asserting and It seems to me that a better approach is to take all belief ascriptions (with the possible exception of belief de se) to express relations to propositions (semantic contents), but to distinguish the proposition expressed by the complements of (iiia) and (iiib) from the proposition expressed by the complements of (iiic) and (iiid). In this way, one can block the move from (b) to (c), while preserving (23). An account of this kind is presented in section 6 below. I am grateful to Mark Richard and Nathan Salmon for discussions of the issues in this note. 25 It might be thought that a theory that took names to be disguised descriptions (associated with them by the speaker) could render the remarks in the two conversations true by appealing to a difference in scope. But this won’t work. If the description associated by the teacher with the name ‘Mark Twain’ is something like ‘the author of The Adventures of Tom Sawyer and Huckleberry Finn’, then the teacher’s remark will be false no matter what the scope of the description.

Direct Reference • 57

believing are relations between individuals and propositions. However, often these attitudes arise in connection with attitudes toward sentences— e.g., uttering and accepting. Although propositional attitude ascriptions report relations to particular propositions, they often suggest corresponding relations to certain sentences. For example, a competent speaker of English typically (though not always) knows that ‘Mark Twain is a writer’ is true iff he knows that Mark Twain is a writer. Thus, it is natural that the teacher’s remark should carry the metalinguistic suggestion. It is also natural that in many cases these suggestions should be important to the conversation. As John Perry has emphasized, sentential attitudes are often more significant for explaining action than propositional attitudes are.26 Think again of Richard’s telephone example. Suppose that a third party asks the question “Why doesn’t A tell B that she is in danger?.” (We assume that A knows his conversational partner under the name ‘B’ and accepts ‘You are B’ in the context.) It is tempting to try to explain A’s behavior by saying, “A doesn’t know that B is in danger.” But this, as we have seen, is false. A better explanation is that A doesn’t accept the sentence ‘B is in danger’. The reason we are tempted by the propositional attitude ascription is that normally we would expect A to accept the sentence iff he thought that B was in danger. However, in this case the usual correlation between sentential and propositional attitudes breaks down. As a result, the explanation suggested by the propositional attitude ascription is correct, even though the ascription itself is false. The general thesis, then, is that the substitutivity principle (23) is correct; and that resistance to it is based on a failure to properly distinguish the semantic information expressed by a sentence relative to a context from the information conveyed by an utterance of it in a given conversation.27 If this is correct, then the main objection to assumptions A2, A4, and compositionality is eliminated, and the case against A1a is strengthened.

6. What becomes of the difficulties in section 3 once this assumption is given up? Taking the argument in (8) as a representative example, we see that the move from (8b) to (8c), and ultimately to (8d), is no longer warranted. 26

Perry (1977, 1979). This thesis has recently been championed by several philosophers, most notably Nathan Salmon (1986). Although I developed the arguments given above independently, I have profited from Salmon’s work on the topic. 27

58 • Essay One

In order to defend this as the proper response to the difficulty, I must explain how one might believe (or assert) an instance of an existential generalization, without believing (or asserting) the generalization itself. Let us focus in particular on the notion of belief. Then, what must be explained is how an individual might satisfy an open sentence x believes that R (t,t), for directly referential t, without satisfying x believes that R(t,t) and for some y, R(y,y), or x believes that for some y, R(y,y). It should be noted that the answer is not that the agent may never have gotten around to drawing the relevant conclusion. For the problematic derivation in (8) would proceed from true premises to a false conclusion even if the agents were perfect logicians. Thus, there must be some deeper explanation of how a person might fail to believe the existential generalization of something he already believes. There are two different aspects of such an explanation. The first is a metaphysical characterization of the nature of belief, specifying the facts in virtue of which belief ascriptions are true. The second is a specification of the objects of belief needed in a semantic theory. I will say a word about each. Regarding the former, we may think of beliefs as arising from certain kinds of mental states, together with their causal relations to objects in the environment.28 On this picture, a belief report, x believes that S, characterizes the agent as being in a mental state whose information content is identical with the semantic content of S in the context of the report. For example, an agent who is in a mental state appropriate for believing that a particular object is F will be correctly reported to believe that Phosphorus is F just in case the relevant part of his belief state is causally anchored to Phosphorus. Since Phosphorus is Hesperus, the agent will thereby believe that Hesperus is F. Suppose the agent believes that Hesperus bears R to Phosphorus. On this picture, he thereby believes of a certain object o that o bears R to o. However, it does not follow that he believes the proposition that something bears R to itself. Since none of the agent’s mental states has this as its information content, he does not believe it. If we restrict our attention to cases in which the agent is a competent speaker of a language, we can make this account less abstract by letting dispositions to assent to sentences play the role of mental states. We then assume something like (24). (24) If i is a sincere, reflective, and competent speaker, then i satisfies x believes that S relative to a context C (and assignment f) iff i is disposed to assent to some sentence S′ whose semantic 28

See Barwise and Perry (1983) for an articulation of this view.

Direct Reference • 59

content in the context of assent = the semantic content of S relative to C (and f).29 Let us suppose that the agent accepts R (Hesperus, Phosphorus) while rejecting R(Hesperus, Hesperus) and For some x R(x,x). An impeccable logician, the agent would accept the latter if he accepted any of its instances, R(a,a). However, he rejects all of these. Since the semantic content of one of the sentences he accepts is identical with the semantic content of R(Hesperus, Hesperus), he believes that Hesperus bears R to Hesperus even though he would not express his belief this way. Since the semantic content of For some x R(x,x) is not identical with the content of any sentence he is disposed to accept, he does not 29 Although this principle is a useful heuristic, it should not be regarded as an analysis of belief. Its most obvious limitation is that it does not apply to believers who are not language users. Even when applied to language users it must be restricted to cases in which the agent, i, the sentence S', and its semantic content (in the context of assent) stand in a certain (as yet not fully analyzed) recognition relation. I have in mind examples like ‘Newminister 1’ in which a proper name is introduced by a reference-fixing description ‘the first Tory Prime Minister of Britain elected in the 21st century’. (The example parallels the ‘Newman 1’ example discussed in Donnellan (1979).) In such a case, the sentence ‘Newminister 1 will be conservative’ may express a singular proposition involving a certain individual. However, assent to the sentence by a competent speaker is not sufficient for belief in that proposition. Intuitively, the manner in which the sentence presents the proposition to the agent is too indirect for assent to indicate belief. It should be noted that the cases discussed in the text (‘Hesperus’/‘Phosphorus’, ‘London’/‘Londres’, etc.) are not like this. In these cases, the agents are acquainted with the referents, they associate names with them, and they grasp the propositions expressed by sentences containing the names. What they do not do is recognize that the same referents are associated with different names, and that the same propositions are expressed by different sentences. But that is not required in order for assent to the sentences to indicate belief in the propositions they express. Similar points can be made about assertion (except that here the principles involving assent come closer to providing an actual analysis).

(i) An individual i satisfies x says (asserts) that S relative to a context C (assignment f) and circumstance E, if there is a sentence S' and context C' corresponding to E with i as agent, such that i assertively utters S' in C', and the content of S' in C'= the content of S in C (relative to f). (ii) An individual i satisfies x says (asserts) that S relative to a context C (assignment f) and circumstance E iff there are sentences S' and S", and a context C' corresponding to E with i as agent, such that i assertively utters S' in C', S" is readily inferable from S', and the content of S" in C'= the content of S in C (relative to f). If one takes (ii) to be a reasonable approximation of the notion of assertion, one can use it in place of (24) to construct arguments and explanations involving assertions corresponding to those in the text involving beliefs. (Note that (ii) allows contents not expressed by the sentence uttered to be asserted when, but only when, the conversational participants “can’t miss them.”)

60 • Essay One

believe that something bears R to itself. Thus, there is a principled way of blocking the move from (8b) to (8d). What we need now is a conception of semantic content capable of incorporating this point. Given that the move from (8c) to (8d) is unproblematic, we need a conception that blocks the move from (8b) to (8c) by assigning different semantic contents to the complement sentences in these examples. This requires the introduction of structure into contents. First consider simple sentences. (25) a. R(Hesperus, Phosphorus) b. R(Hesperus, Hesperus) c. R(Hesperus, itself) Regimenting a bit, we can think of the semantic contents of these examples as being identical with that of certain canonical representations. (26) a. R(h,p) b. R(h,h) c. [λx R(x,x)] h Where o is the referent of ‘Hesperus’ and ‘Phosphorus’ the content of (a) and (b) is, in effect, ; the content of (c) is .30 Accepting (a) leads to a belief whose object is the first of these semantic contents; accepting (c) leads to a belief whose object is the second such content. Accepting (b) typically leads to a belief in both. The reason for this has to do with the transparent linguistic relationship between (b) and (c). A competent speaker who accepts one will normally be disposed to accept the other, thereby acquiring both beliefs.31 Thus, a speaker who satisfies (27a) will standardly satisfy both (27b) and (27c). (27) a. x accepts ‘R(Hesperus, Hesperus)’ b. x believes that R(Hesperus, Hesperus). c. x believes that R(Hesperus, itself). However, not everyone who satisfies (27b) satisfies (27c). Whether or not the latter is satisfied will depend on the manner in which the agent believes that Hesperus bears R to Hesperus. If he believes it in virtue of accepting a sentence of the form ‘R(a,a)’ then he can be expected to believe 30

The idea for this lambda-treatment of (25c) was suggested to me by Nathan Salmon. Except in situations like Kripke’s Paderewski example, in which the agent misconstrues two tokens of the same name (referring to the same individual) for tokens of different (but phonologically identical) names of different individuals (Kripke 1979, 265–66). In such cases an agent might accept (b) without accepting (c), or believing what it expresses. 31

Direct Reference • 61

that Hesperus bears R to itself. However, if he believes it in virtue of accepting a sentence of the form ‘R(a,b)’, then (27c) may not be satisfied. The same point holds for (27d). (27d) x believes that for some y, R(y,y). A sincere, reflective, competent speaker who accepts R(a,a) will typically be disposed to accept For some y R(y,y), and thereby believe that which it expresses. However, someone who accepts R(Hesperus, Phosphorus) may satisfy (27b) without satisfying (27d). In order to reflect this in a semantic theory we must extend our account of structured semantic contents from atomic sentences to compound sentences of arbitrary complexity. This raises the question of how much structure is needed. Where S is an atomic sentence consisting of an n-place predicate plus n occurrences of directly referential terms, its structured semantic content consists of the content of the predicate plus the content of each term occurrence. There are two ways of thinking of this—as a complex made up of the semantic contents of all occurrences of its semantically significant parts, or as a complex made up of the contents of all occurrences of its directly referential terms, plus the content of whatever else is left over. In the case of atomic sentences, these characterizations come to the same thing. However, they generalize in different ways. The first leads to a conception of the semantic contents of sentences as structured Russellian propositions, the second to a conception of contents as partially structured intensions. For simplicity let us consider the semantic contents of sentences in a first order language with lambda abstraction, a belief predicate, and a stock of semantically simple singular terms, all of which are directly referential. On the Russellian account, the semantic content of a (free) variable v relative to an assignment f of individuals to variables is f(v), and the semantic content of a closed (directly referential) term, relative to a context, is its referent relative to the context. The semantic contents of n-place predicates are n-place properties and relations. The contents of ‘&’ and ‘-’ are functions, CONJ and NEG, from truth-values to truthvalues.32 Variable-binding operations, like lambda abstraction and existential quantification, can be treated in a number of ways. One of the simplest,

32 On this treatment, ‘&’ and ‘-’ are like directly referential terms in that their semantic contents = their extensions. This is not crucial to the general Russellian conception, which could just as well take the contents of these expressions to be entities—call them operations—which bear a relation to functions analogous to that borne by properties to the objects they apply to. It is even possible to take the contents of truth-functional operators to be properties of truthvalues. The differences between these alternatives do not affect the present discussion.

62 • Essay One

semantically, involves the use of propositional functions in place of complex properties as propositional constituents corresponding to certain compound expressions.33 On this approach, the semantic content of  [λx Rx,x] is the function g from individuals o to propositions that attribute the property expressed by R to the pair . ∃x Rx,x can then be thought of as “saying” that g assigns a true proposition to at least one object. (28) uses these ideas to assign Russellian propositions to sentences. (28) a. The proposition expressed by an atomic formula Pt1, . . . ,tn relative to a context C and assignment f is , where P* is the property expressed by P, and oi is the content of ti relative to C and f. b. The proposition expressed by a formula [λvS]t relative to C and f is , where o is the content of t relative to C and f, and g is the function from individuals o′ to propositions expressed by S relative to C and an assignment f′ that differs from f at most in assigning o′ as the value of v. c. The propositions expressed by −S and S&R relative to C and f are and respectively, where Prop S and Prop R are the propositions expressed by S and R relative to C and f, and Neg and Conj are the truth functions for negation and conjunction. d. The proposition expressed by ∃v S relative to C and f is , where SOME is the property of being a nonempty set, and g is as in (b). e. The proposition expressed by t believes that S relative to C and f is , where B is the belief relation, o is the content of t relative to C and f, and Prop S is the proposition expressed by S relative to C and f. f. The proposition expressed by a sentence (with no free variables) relative to a context C is the proposition it expresses relative to C and every assignment f. In stating clause (d), I have departed slightly from Russellian ideas in favor of a suggestion by Nathan Salmon. A purely Russellian approach would treat SOME as the property of being a propositional function that is “sometimes true.” However, since the existential quantifier is an extensional operator, it seems more natural that it should express a property of the extension of its operand (rather than a property of the propositional constituent expressed by the operand, as in the case of ‘believe’). On this 33

I am indebted to David Kaplan for this Russellian suggestion.

Direct Reference • 63

formulation, is true relative to a circumstance E iff the set of objects in E that g maps onto propositions true in E is nonempty. This discussion of truth conditions brings up an important point. Propositional contents do not replace truth-supporting circumstances in a semantic theory; rather, they supplement them with a new kind of semantic value. On this view, the meaning of an expression is a function from contexts to propositional constituents. The meaning of a sentence is a compositional function from contexts to structured propositions. Intensions (and extensions) of expressions relative to contexts (and circumstances) derive from intensions (and extensions) of propositions and propositional constituents. These, in turn, can be gotten from a recursive characterization of truth with respect to a circumstance, for propositions. For this purpose, we let the intension of an n-place property be a function from circumstances to sets of n-tuples of individuals (that instantiate the property in the circumstance); we let the intension of an individual be a constant function from circumstances to that individual; and we let the intension of a one-place propositional function g be a function from circumstances E to sets of individuals in E that g assigns propositions true in E. Extension is related to intension in the normal way, with the extension of a proposition relative to a circumstance being its truth-value in the circumstance, and its intension being the set of circumstances in which it is true (or, equivalently, the characteristic function of that set). Truth relative to a circumstance is defined as follows: (29) a. A proposition is true relative to a circumstance E iff the extension of P* in E contains . b. A proposition is true relative to E (where g is a one-place propositional function) iff o is a member of the extension of g in E (i.e., iff g(o) is true in E). c. A proposition is true relative to E iff the value of Neg at the extension of Prop S in E is truth (i.e., iff Prop S is not true in E). A proposition is true relative to E iff the value of Conj at the pair consisting of the extension of Prop S in E and the extension of Prop R in E is truth (i.e., iff Prop S and Prop R are true in E). d. A proposition is true relative to E (where g is as in (b)) iff the extension of g in E is nonempty (i.e., iff g(o) is true relative to E for some o in E). e. A proposition is true relative to E iff is a member of the extension of B in E (i.e., iff o believes Prop S in E).

64 • Essay One

According to this theory, the propositions expressed by the complements of (8b) and (8c) are (8b*) and (8c*). (8) b*. c*. (Where g is the function which assigns to any object o the proposition about o corresponding to the proposition 8b* about Hesperus.) Although the circumstances supporting the truth of these propositions are the same, the propositions themselves are different. Thus, we no longer have the result that anyone who believes the proposition expressed by the complement of (8b) thereby believes the proposition expressed by the complement of (8c). The argument in (8) is, therefore, blocked and the problematic conclusion avoided. Similar results hold for the other arguments in section 3. However, this is not the only way these results can be achieved. One of the striking features of Russellian propositions is that they encode a good deal of the syntactic structure of the sentences that express them. Sentences that are negations, conjunctions, or quantifications express propositions which are themselves negative, conjunctive, or quantificational in structure.34 Although this systematic assignment of structure to semantic contents is appealing, it goes beyond what is required by the interaction of propositional attitudes and directly referential singular terms exhibited in section 3. In each of the problematic arguments, the agent accepts, or assertively utters, a sentence of the form (30a), but fails to accept, or assertively utter, a corresponding sentence of the form (30b) (which is true in the same circumstances as (30a)). (30) a. S(t, t′) b. S(t, t′) & R In each case, the agent would accept, or assertively utter, (30b) if he knew that the directly referential terms t and t′ had the same content (and he continued to accept (30a)). However, he doesn’t know that they have the same content. In order to focus on the special difficulties created by this sort of ignorance let us suppose, for the sake of argument, that the agent is otherwise semantically omniscient. Thus, he knows, for any two expressions not containing directly referential terms, whether or not they have the same intension. 34 It is, of course, possible for sentences of one form to express propositions of another form, as happens in some cases of stipulative definition.

Direct Reference • 65

In particular, he knows this about (31a) and (31b). (31) a. λv,v′ [S(v,v′)] b. λv,v′ [S(v,v′) & R] If he thought that these expressions had the same intension, then his attitude toward (30a) and (30b) would be the same—he would either accept them both or reject them both. Since, in fact, he accepts one and rejects the other, it follows that (31a) and (31b) have different intensions. This means that whenever an argument of the sort presented in section 3 can be constructed, its problematic conclusion can be blocked by taking the semantic content of a sentence to be a complex consisting of intensions of all occurrences of its directly referential singular terms, plus an intension determined by the remainder of the sentence. The idea can be carried out using a standard style definition of truth with respect to a context and circumstance. Such a definition allows one to associate both a standard intension and a partially structured intension with every object language sentence. Standard intensions of sentences can be taken to be sets of truth-supporting circumstances. Partially structured intensions are complexes made up in part of the intensions of directly referential terms. If a sentence contains no such terms, then its partially structured intension is identified with its standard intension. We can make this more precise as follows: Let us call an occurrence of a singular term in a sentence S a structurally sensitive occurrence iff it is a free occurrence of a variable in S or it is an occurrence of a (constant) directly referential term.35 Let λv1, . . . ,vn S′ arise from S by prefixing  λv1, . . . ,vn and replacing each structure-sensitive occurrence of a singular term in S with a variable new to S, distinct variables for distinct occurrences, vi replacing the ith such occurrence. The extension of λv1, . . . ,vn S′ relative to an assignment f, context C, and circumstance E, is taken to be the function from n-tuples to truth-values of S′ relative to f′, C, and E, where f′ is just like f except (at most) for assigning oi, as the value of vi, for each i. Standard intension is determined from extensions in the normal way. For any (open or closed) sentence S, the partially structured intension of S relative to an assignment f and context C, is , where [ti] is the intension of the ith structure-sensitive occurrence of a singular term in S, relative to f and C, and [λv1, . . . ,vnS′] is the intension of λv1, . . . ,vnS′, relative to f and C. (Closed sentences have the same partially structured intensions—with respect to a context—relative to all assignments.) An individual i satisfies 35 I retain here the simplifying assumption that all directly referential terms in the object language are semantically simple.

66 • Essay One

an open sentence x believes that S, relative to f, C, and E, iff in E, i bears the belief relation to the partially structured intension expressed by S relative to f and C.36 Conceptually, this approach lies somewhere between the Russellian theory and the familiar truth-supporting circumstance conception. Like the Russellian theory, it takes propositions to be structured complexes which are both the semantic contents of sentences and the objects of propositional attitudes.37 However, unlike the Russellian theory, the constituents of these “propositions” are intensions extractable from a conventional truth definition. Moreover, the resulting “propositions” are only partially structured. For example, the partially structured contents of the complements of (8b) and (8c) are: (8) b#. (Where R′ is the intension corresponding to the four-place relation of x’s referring to y and z’s referring to v.)38 c#. (Where R′ is the intension corresponding to the four-place relation of x’s referring to y and z’s referring to v and there being a common referent of x and z.) Since R′ is not identical with R″, these contents are different. The move from (8b) to (8c) is, therefore, blocked and the problematic conclusion, (8d), is avoided. Corresponding results hold for other arguments of this type, including those in section 3. This approach represents a theoretically minimum response to the difficulties in section 3. As such it allows us to establish a minimum positive result about the relationship between direct reference and propositional attitudes, corresponding to the impossibility result, (13). (32)

If direct reference is legitimate and (some) propositional attitude verbs have a relational semantics (A4 plus A2), then (assuming compositionality and distribution over conjunction) the semantic content of a sentence, relative to a context and

36 This account has two precursors. The first is the introduction of structured meanings (characters) in Richard (1983). The second is a somewhat different use of structured intension suggested by David Kaplan (personal correspondence) in response to Richard. The account in the text is designed as an improvement on those treatments intended to capture certain insights that motivated them. 37 One could have versions of these theories in which semantic contents were not objects of the attitudes, but only by foregoing the strong motivation the attitudes provide for these theories. 38 Strictly speaking, the intensions of directly referential terms in (8b) and (8c) should be constant functions from circumstances to objects, rather than objects themselves. However this does not affect the issues at hand.

Direct Reference • 67

assignment of values to variables, must encode at least as much structure as is determined by occurrences of its directly referential singular terms (including free variables).39 Both structured Russellian propositions and partially structured intensions satisfy this requirement.

7. How then might we decide between these two conceptions of semantic content? Considerations involving the interaction of propositional attitudes and directly referential singular terms will, I believe, take us no further. However, other considerations will. The first of these involves related expressions which allow the construction of arguments corresponding to those in section 3. For example, if K and K′ are natural kind terms with the same semantic content, the potentially false (12b′) can be derived from the potentially true (12a′) by an argument paralleling the original (12). (12) a′. x believes (asserts) that the G is a K and if the G is a K′, then S. (where S is any sentence and the G is any description) b′. x believes (asserts) that S. Both this argument and the original (12) are blocked by requiring the semantic content of a sentence to encode at least as much structure as is determined by occurrences of its directly referential singular terms, plus its natural kind terms. This conclusion can be extended to include every kind of expression that is relevantly similar to directly referential singular terms and natural kind terms. The relevant feature, I suggest, is one that involves linguistic competence—in the sense that linguistic competence is important for determining what is said or believed by a speaker from what is assertively uttered or accepted by the speaker. If it is possible for a competent speaker to fail to recognize cases in which expressions of type T have the same semantic content, then it will be possible to use these expressions to construct arguments of the kind given in section 3. Blocking these arguments requires ensuring that the structure encoded in semantic contents includes that determined by occurrences of expressions of type T. 39 The significance of this result is enhanced by the defense, in section 5, of the consequence (23) of A2, A4, and compositionality. However, it should be noted that analogous results involving the encoding of structure in objects of the attitudes can be established using the weakenings in section 4.

68 • Essay One

This line of reasoning leads to the encoding of more and more structure into semantic contents, However, it might be thought that at least some expressions—including logical constructions plus certain predicates— remain immune from such considerations.40 If S is a sentence containing only such expressions, then its semantic content, on the partially structured intension approach, will just be a standard intension. If S contains only such expressions plus directly referential singular terms, then its semantic content, on this approach, will be a partially structured intension in the original sense. But this is still problematic. The difficulties posed by propositional attitude ascriptions for truthconditional approaches to semantics are not limited to cases arising from directly referential singular terms and their ilk. For example, if truthsupporting circumstances are metaphysically possible worlds, then the partially structured intension approach will assign the same semantic contents to the (a) and (b) sentences in the following examples: (33) a. b. (34) a. b.

First order logic is complete. First order logic is undecidable. First order logic is decidable. First order logic is decidable and S. (For unrelated S)

However, in both cases, many have believed or asserted that which is expressed by (a) without believing or asserting that which is expressed by (b). Switching to a conception in which truth-supporting circumstances are logically possible worlds only shifts attention to a more restricted, but similarly problematic, class of cases. Like Frege of the Grundgesetze, many of us have had the misfortune of satisfying x asserts (believes) that I, for a some logically impossible I, without thereby satisfying x asserts (believes) that I&S, or x asserts (believes) that S, for unrelated S. The problem is, I believe, inherent, in the truth-conditional approach, and, hence, cannot be solved by weakening constraints on truth-supporting circumstances still further. For example, consider a system like that of Barwise and Perry (1983), in which truth-supporting circumstances may be metaphysically impossible, incomplete, and inconsistent in the sense defined in section 2. In such a system, logically equivalent sentences are often assigned different semantic contents, which may be the objects of different propositional attitudes. As with all such approaches, however, the system incorporates principles like (7a–e), which can be gotten from standard, recursive treatments of logical constructions. Inevitably, sentences involving multiple constructions of this kind require psychologically nontrivial computations to determine their “semantic contents.” Thus, one can always 40

I leave it open whether there are such expressions.

Direct Reference • 69

find psychologically nonequivalent sentences which are true in the same circumstances, and, hence, are assigned the same content. One simple example of this kind, is given in (35). (35) a. C[the x: Ax] & D[the x: Cx] & C[the x: Bx] b. A[the x: Bx v Cx] & B[the x: Dx & Cx] & D[the x: Ax v Cx] c. B[the x: Ax] & C[the x: Bx] & D[the x: Cx] Although these sentences are assigned the same semantic content by corollaries (7a, b, e) of Ala, it takes a modest amount of calculation to determine this. Not all agents of propositional attitudes are adept at such calculations. Thus, it is possible to find agents who are willing to accept, or assertively utter, one of these sentences at a certain time, but not the others. Such agents believe, or assert, that which is expressed by the sentence they accept, or assertively utter. However, it is counterintuitive to suppose that they must thereby believe, or assert, what the other sentences express. The Russellian conception of propositions allows one to respect this intuition; the truth-supporting circumstance approach does not. A related point involves the relationship between propositional attitudes and conjunction. Surely, anyone who believes that (35a), or believes that (35c), believes that (36).41 (36) C[the x: Bx] However, this does not seem to be so with (35b). The reason for this difference is that in one case the move is from a belief in a conjunction to a belief in a conjunct, whereas in the other case it is not. Although many logical operations do not preserve belief, it would seem that simplification of conjunction does. In fact, I think this observation about conjunction and belief is correct. However, it has far-reaching theoretical significance that belies its widespread acceptance. Let us suppose that what are believed are semantic contents of sentences. On the truth-supporting circumstance approach, these contents never have conjunctive structure. At best, they are partially structured intensions, which reflect the structure determined by occurrences of directly referential terms (and related expressions), but obliterate other logical structure. Thus, on this approach, there is no more reason to think that anyone who believes that (35a) (or (35c)) believes that (36) than there is to think that anyone who believes that (35b) does. Proponents of the truth-supporting circumstance approach can, of course, countenance the move from belief in that which is expressed by a conjunction to belief in that which is expressed by the conjuncts. Indeed, 41 Here, I am using ‘(35a)’, ‘(35c)’, and ‘(36)’ not as names, but as abbreviations for the sentences they normally name.

70 • Essay One

they standardly do. However, the price to be paid is that of countenancing the move from x believes that S to x believes that S′ whenever the set of circumstances supporting the truth of S is included in the set of circumstances supporting the truth of S′. But this just substitutes the generation of unwanted inferences for the failure to capture one that is desired. In short, the truth-supporting circumstance approach doesn’t provide the right options.42 The Russellian approach offers a welcome contrast. Given the intuition that whenever an individual satisfies x believes that A & B he also satisfies x believes that A and x believes that B, the Russellian approach supplies a plausible explanation. Since objects of belief reflect the logical structure of sentences used to report those beliefs, whenever a belief is correctly reported using a conjunction the agent will believe a conjunctive proposition which includes the propositions expressed by the conjuncts as constituents. Since these constituent propositions are, so to speak, before his mind, no computation is required in order for him to arrive at beliefs in the conjuncts. We can think of this somewhat less metaphorically as follows: To believe a conjunctive proposition, , is to be in a belief state whose constituents correspond to its three main components. In the case of CONJ, this correspondence is, presumably, functional. A belief state constituent C represents CONJ only if an individual who is in a “conjunctive belief state” S, in which C relates constituent belief states S1 and S2, is also in—or disposed to be in—S1, and S2. Thus, anyone who believes a conjunction believes both conjuncts. The point to notice is that with propositions as semantic contents this result does not generalize in unwanted ways. Even though structured propositions determine truth-supporting circumstances, there is no reason to suppose that just because an agent bears the belief relation B to a proposition P, he must also bear B to Q whenever the class of truth-supporting circumstances for P is identical with, or a subclass of, the class of truthsupporting circumstances for Q. There are, then, good reasons not only for rejecting a strict truthsupporting circumstance conception of semantics, but also for adopting a Russellian approach. The reasons I have stressed rest on commonplace intuitions and assumptions about propositional attitudes. There are, of course, those who regard the attitudes as ill-behaved and problematic, and would, therefore, not accept such intuitions and assumptions. In my opinion, such pessimism is unwarranted. If I am right, a major reason why propositional attitudes have often seemed intractable is that the basic features of strict truth-theoretic 42

An analogous argument can be constructed regarding assertion.

Direct Reference • 71

semantics have been incompatible with elementary facts about them. The introduction of structured Russellian propositions, which determine, but are not determined by, sets of truth-supporting circumstances, has the potential to change that.

References Almog, Joseph, John Perry, and Howard Wettstein, with the assistance of Ingrid Deiwiks and Edward N. Zalta, eds. 1989. Themes from Kaplan. Oxford: Oxford University Press Barwise, Jon, and John Perry. 1983. Situations and Attitudes. Cambridge MIT Press. Donnellan, Keith. 1979. “The Contingent Apriori and Rigid Designators.” In Contemporary Perspectives in the Philosophy of Language, ed. Peter A. French, Theodore E. Uehling, and Howard K. Wettstein, 12–27. Minneapolis: University of Minnesota Press. Kripke, Saul A. 1979. “A Puzzle about Belief.” In Meaning and Use: Papers Presented at the Second Jerusalem Philosophical Encounter, April 1976, ed. Avishai Margalit, 239–83. Dordrecht: Reidel. Perry, John. 1977. “Frege on Demonstratives.” Philosophical Review 96:474–97. ———. 1979. “The Problem of the Essential Indexical.” Noûs 13:3–21. Richard, Mark. 1983. “Direct Reference and Ascriptions of Belief.” Journal of Philosophical Logic 12:425–52. ———. 1986. “Quotation, Grammar, and Opacity.” Linguistics and Philosophy 9:383–403. Salmon, Nathan. 1986. Frege’s Puzzle. Cambridge: MIT Press. Soames, Scott. 1985. “Lost Innocence.” Linguistics and Philosophy 8:59–71.

ESSAY TWO

Why Propositions Can’t Be Sets of Truth-Supporting Circumstances

In my article “Direct Reference, Propositional Attitudes, and Semantic Content,”1 I argued that any semantic theory satisfying certain natural and well-motivated assumptions cannot identify the semantic contents of sentences (the propositions they express) with sets of circumstances in which the sentences are true—no matter how fine-grained the circumstances are taken to be. The argument takes the form of a reductio of the following set of assumptions: A1. The semantic content of a sentence or formula (relative to a context and assignment of values to variables) is the collection of circumstances supporting its truth (relative to the context and assignment). A2. Propositional attitude ascriptions report relations to the semantic contents of their complements—i.e., x v’s that S is true with respect to a context C, assignment A (of values to variables) and a circumstance E of evaluation iff in E, the referent of ‘x’ with respect to A bears R to the semantic content of S relative to C and A. (When v is the verb ‘believes’, R is the relation of believing, when v is the verb ‘says’ or ‘asserts’, R is the relation of saying, or asserting, and so on for other attitude verbs.) A3. Many attitude verbs, including ‘say’, ‘assert’, ‘believe’, ‘know’, and ‘prove’ distribute over conjunction. For these verbs, x v’s that P & Q is true with respect to C, A, and E only if x v’s that P and x v’s that Q are too. A4. Names, indexicals, and variables are directly referential—their semantic contents, relative to contexts and assignments, are their referents with respect to those contexts and assignments. Com. If S1 and S2 are nonintensional sentences/formulas with the same grammatical structure, which differ only in the substitution of constituents with the same semantic contents (relative to their respective contexts and assignments), then the 1 Soames (1987a) and essay 1 in this volume. The argument is also given in Soames (1989).

Propositions • 73

semantic contents of S1 and S2 will be the same (relative to those contexts and assignments). The assumptions required by the argument are quite weak. A1 is true of all truth-conditional approaches to semantics that (unlike the Davidsonian approach) identify certain entities—the truth conditions of sentences/formulas—as their semantic contents (relative to contexts and assignments). The entities—which are sets of circumstances in which the sentences/formulas are true—can be conceptualized in any number of ways, along a continuum running from very fine-grained (e.g., the abstract situations of Barwise and Perry) to very coarse-grained (e.g., complete metaphysically possible world-states).2 The only relevant presupposition of A1 is that its truth requires the truth of the corollaries like A1a and A1b. A1a. A conjunction P & Q is true with respect to a context C, assignment A, and circumstance E iff P and Q are both true with respect to C, A, and E. Thus, the semantic content of a conjunction, relative to C and A, is the intersection of the semantic contents of the conjuncts, relative to C and A. A1b. An existential generalization For some x: Fx is true with respect to a context C, assignment A, and circumstance E iff there is some object o in E such that ‘Fx’ is true with respect to an assignment A′ that differs from A at most in assigning o as value of ‘x’. The semantic content of For some x: Fx relative to C and A is the set of circumstances E such that for some object o in E, o satisfies ‘Fx’ with respect to C, A, and E. The compositionality principle, Com, employed in the argument is also weak. All that is needed is a principle ensuring that substitution of expressions with the same semantic content in extensional sentences (that may occur as the complements of attitude ascriptions) preserves the semantic contents of those sentences. Thus, Com can afford to be silent about whether the semantic contents of sentences containing modal, propositional, or other intensional operators is similarly compositional.3 Finally, although A4 asserts the direct reference of names, indexicals (relative to contexts), and variables (relative to assignments), either variables alone, or variables plus indexicals would be sufficient. However, since examples involving names are simple, and easy to understand, I use a principle of direct reference that is stronger than that which is strictly required. 2

Barwise and Perry (1983). In the argument, Com is understood as presupposing that . . . α . . . β . . . and . . . α . . . α . . . have the same grammatical structure. This assumption is defended in Soames (1987b). 3

74 • Essay Two

The main illustrative example used in the reductio is R. R1. The ancients believed (asserted) that ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Phosphorus. R2. Since Hesperus is Phosphorus, this means (given A2, A4, and Com) that the ancients believed (asserted) that ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Hesperus. R3. Thus, the ancients believed (asserted) that: ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Hesperus and, for some x, “Hesperus’ referred to x and ‘Phosphorus’ referred to x. (From R2, A1a, A1b, and A2) R4. So, the ancients believed (asserted) that: for some x, ‘Hesperus’ referred to x and ‘Phosphorus’ referred to x—i.e., they believed that the names were coreferential. (From R3 and A3) The argument based on this example takes two pretheoretic facts for granted—that (R1) is true, and that Hesperus is, indeed, Phosphorus. What the argument shows is that any semantic theory T incorporating A1–A4, plus Com, is incompatible with these facts—in the sense that their existence is sufficient to show that T is incorrect. It is concluded on independent grounds that A1 is the offending assumption, and hence that the semantic content of a sentence is not the set of circumstances supporting its truth. Instead, it is argued, the semantic content of a sentence S is a structured proposition the constituents of which are the semantic contents of the constituents of S.

An Objection In his paper “Propositions, Circumstances, and Objects,” Walter Edelberg maintains that the argument fails because the reductio argument (R1–R4) is fallacious.4 His own formulation of the critical points he proposes to establish is given in the following two passages. I won’t be arguing that Soames has rejected the wrong assumptions of the reductio, though one might worry about that. Instead I will be arguing that no absurdity results from the general theoretical assumptions Soames cites . . . (1994, 2) Intriguing as Soames’s argument is, I think it rests on a mistake. For the argument is intended to defend the following claim. 4 Edelberg (1994), reprinted as “among the ten best articles to appear in print in 1994,” Philosophers’ Annual 17 (1996).

Propositions • 75

The Reductio Claim. Sentences (1) and (2) below will entail sentence (3) on any semantical theory countenancing [A1–A4, plus Com]. (1) Hesperus is Phosphorus (2) The ancients believed that ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Phosphorus (3) The ancients believed that for some x, ‘Hesperus’ referred to x and ‘Phosphorus’ referred to x. This claim is false. (1994, 6–7) Edelberg’s account of the allegedly mistaken defense of the “Reductio Claim” is as follows: Let’s suppose that the seven assumptions [A1–A4, plus Com and corollaries A1a and A1b] are true, and see how Soames tries to derive (3). From the Direct Reference principle [A4] and the truth of (1), it follows that ‘Hesperus’ and ‘Phosphorus’ have the same semantic content. So by Substitution [Com] it follows that (4) and (5) express the same proposition. 4. ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Phosphorus. 5. ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Hesperus. By the Circumstantialist Conception [A1], Truth.∃x, and Truth.& [A1b and A1a], it follows that (5) and (6) also express the same proposition. 6. ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Hesperus and for some x, ‘Hesperus’ referred to x and ‘Phosphorus’ referred to x. Since (4) and (5) express the same proposition, and so do (5) and (6), it follows that (4) and (6) express the same proposition. Given that (2) is true and that (4) and (6) express the same proposition, it follows by Truth.PA [A2] and Substitutivity [Com] that (7) is true. 7. The ancients believed that (‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Hesperus and for some x, ‘Hesperus’ referred to x and ‘Phosphorus’ referred to x). From Distribution [A3] and the truth of (7), it then follows that (3) is true. (1994, 4)

76 • Essay Two

Edelberg believes that the above reasoning is faulty, and that the thesis he dubs the “Reductio Claim” is false. Unfortunately, in attempting to demonstrate this, he does not define what ‘entails’, as used in that claim, is supposed to mean. We can, however, reconstruct from his argument a sense of entailment that fits his conclusion. Think of a semantic theory for a language incorporating assumptions A1–A4, plus Com, as being divided into three parts: (i) A recursive characterization of truth in an arbitrary model M (conforming to the theory), relative to a context C, assignment A, and circumstance E. (ii) A definition of the semantic content of a formula F—in M, relative to C and A—as the set of circumstances E supporting the truth of F—in M relative to C and A. (iii) A specification of an intended model MI that provides a domain of objects and the interpretations of the nonlogical vocabulary. A standard notion of model-theoretic entailment can then be defined for such a theory as follows: Model-Theoretic Entailment A set S of sentences model theoretically entails a sentence S* according to a theory T—incorporating A1–A4, plus Com—iff for every model M conforming to T, and every context C and circumstance E of M, if all the sentences in S are true in M with respect to C and E, then so is S*. This definition fits what Edelberg observes—namely that in a model M that assigns a pair of directly referential names a and b different referents, there may be circumstances E such that (1′) and (2′) are true in M with respect to E, even though (3′) is false in M with respect to E.5 (1′) a = b (2′) c believes that (Fa and Gb) (3′) c believes that (∃x) (Fx and Gx) The important point to notice is that even though M assigns a and b different referents, o and o′, the truth-supporting circumstances in M need not be metaphysically possible, and hence may include —which predicates the identity relation of different objects. Any such truth-supporting circumstance E is such that (1′) is true in M with respect to E. Suppose further that (2′) is true in M with respect to E. If 5 Since the semantic contents of names don’t vary with contexts, relativization to context is here suppressed.

Propositions • 77

the semantic content of (4′) in M were the same as that of (5′) and (6′), then, since the belief predicate distributes over conjunction, it would follow that (3′) was true in M with respect to E. (4′) Fa & Gb (5′) Fa & Ga (6′) Fa & Ga & (∃x) (Fx and Gx) However, since a and b have different referents in M, (4′) may (and standardly will) have a semantic content in M different from that of (5′) and (6′). Because of this, (3′) may be false in M with respect to E, even though (1′) and (2′) are true in M with respect to E. Hence, (3′) is not model theoretically entailed by (1′) and (2). In fact, according to the theory, (1′) and (2′) do not model theoretically entail (8).6 (8)

c believes that Fa and Ga.

Edelberg’s Error This argument, using the above definition of model-theoretic entailment, succeeds in establishing the falsity of what Edelberg dubs the “Reductio Claim”—the claim that sentences (1) and (2) model-theoretically entail (3) (for any T incorporating A1–A4, plus Com). However, the argument does not establish the incorrectness of the original reductio—since the reductio did not attempt to establish that claim. What the original reductio demonstrated was that no semantic theory T incorporating A1–A4, plus Com, can be correct because: (i) being correct requires assigning ‘Hesperus’ and ‘Phosphorus’ the same referent (Venus), and (ii) incorporating A1–A4, plus Comp, forces T to wrongly characterize the false R4—The ancients believed (asserted) that: for some x, ‘Hesperus’ referred to x and ‘Phosphorus’ referred to x—as a consequence of the true R1—The ancients believed (asserted) that ‘Hesperus’ referred to Hesperus and ‘Phosphorus’ referred to Phosphorus. In order to appreciate the distinction, one must remember that a semantic theory of the truth-supporting-circumstance variety is not just a characterization of truth with respect to a context and circumstance of an arbitrary model. Nor is it that plus definitions of (a) model-theoretic entailment and (b) the semantic content of a sentence (relative to a context) in a model. In addition, an intended model MI must be specified to interpret the nonlogical vocabulary. 6

This is a summary of the argument Edelberg gives (1994, 8 and 9).

78 • Essay Two

Given an intended model MI, we can define a notion of truthconditional consequence (over and above model-theoretical entailment) according to the theory as follows: Truth-Conditional Consequence Let T be a theory incorporating A1–A4, plus Com, with intended model MI. The content of a sentence (or formula) S*, relative to a context C of MI and assignment A of values to variables, is a truth-conditional consequence of the content(s) of a set S of sentences (or formulas), relative to C and A, iff for every circumstance E of MI, if all members of S are true in MI with respect to C, A and E, then so is S*. Equivalently put, the content of S* (relative to C and A) is a truthconditional consequence (in MI) of the content of S (relative to C and A) iff the set of circumstances common to each truth-conditional content expressed by a member of S in MI (relative to C and A) is a subset of the truth-conditional content (set of circumstances) expressed by S* in MI (relative to C and A). In effect, truth-conditional consequences of the content of S are what necessary consequences of the content expressed by S become when truth-supporting circumstances are not required to be metaphysically possible world-states. The point to emphasize here is that truth-conditional consequence and model-theoretic consequence (the converse of model-theoretic entailment) are very different notions (despite their similar-sounding names). Whereas the former is a relation between the semantic contents of sentences relative to contexts (and assignments, if the sentences contain free occurrences of variables), the latter is a relation between sentences themselves. Since the semantic content of a sentence, relative to a context, is what the sentence “says” or “expresses,” relative to the context, truth-conditional consequence is a notion from semantics, in the sense of a fully-fledged theory of meaning that assigns interpretations to all meaningful expressions of the language. Since model-theoretic consequence is a relation between sentences in which the interpretations of the nonlogical vocabulary are allowed to vary from model structure to model structure, it is a semantic notion only in the sense in which it is a semantic characterization of a logical concept (as opposed, say, to a proof-theoretic characterization). These two senses of ‘semantics’—theory of meaning vs. truth-based theory of logical consequence—are very different. To take just one point of contrast, consider (9a) and (9b), where a and be are names, indexicals, or variables that are coreferential (relative to a context and assignment). (9) a. a = b b. a = a

Propositions • 79

Whereas a theory of meaning incorporating A4 will characterize the semantic contents of these sentences as truth-conditional consequences of one another, a theory of logical consequence will deny that (9a) is a model-theoretic consequence of (9b). There is, of course, no conflict here—since the characterizations are noncompeting. However, it is crucial that one not conflate them, which, in essence, Edelberg does. Although his remarks are largely on-target when taken as comments on (nonstandard) theories of logical consequence in which the truth of a sentence in a model is relativized to truth-supporting circumstances that are allowed to be partial and/or metaphysically impossible, they miss the mark when taken as comments on theories of meaning in which semantic contents of sentences are constructed out of such circumstances. This error, though of fundamental importance, is not uncommon. Correcting it not only reinstates my reductio of a certain class of theories of meaning, but also helps to clear up widespread confusion about the relationship between semantic theories of meaning and semantic theories of logical consequence.

Conclusion What the original reductio established was the reductio-claim RC1. RC1. Let T be a semantic theory incorporating A1–A4, plus Com, with intended model MI. According to T, for all singular terms (names, indexicals, variables) a and b, and any context C of MI and assignment A, if a and b refer to the same thing with respect to C and A, then the semantic content of (3′)— c believes that (∃x) (Fx and Gx) —relative to C and A is a truth-conditional consequence of the semantic content of (2′)— c believes that (Fa and Gb) —relative to C and A. If one further assumes—as one must—that a semantic theory incorporating A1–A4, plus Com, assigns semantic contents to sentences relative to all actual contexts—i.e., contexts of utterance that incorporate circumstances of evaluation that actually obtain (and hence are metaphysically possible)—then the reductio can be seen as establishing RC2 as well. RC2. Let T be a semantic theory incorporating A1–A4, plus Com, with intended model MI. According to T, for all singular terms (names, indexicals, variables) a and b, if (1′)— a = b —is true in MI with respect to any actual context C, assignment A, and circumstance EC of C, then the semantic content of (3′)—  c believes that (∃x) (Fx and Gx) —relative to C and A is a

80 • Essay Two

truth-conditional consequence of the semantic content of (2′)—c believes that (Fa and Gb) —relative to C and A. These results, which are true, must not be confused with RC3, which is false (if T allows circumstances of evaluation which are metaphysically impossible). RC3. Let T be a semantic theory incorporating A1–A4, plus Com, with intended model MI. For all names and indexicals a and b, {(1′), (2′)} model-theoretically entails (3′). RC1 is enough to establish the incorrectness of semantic theories incorporating A1–A4, plus Com. For example, when a and b are names, we don’t need to consider contexts and assignments. If the names are in fact coreferential, then any semantic theory that makes them (rigid, directly referential) terms that refer to different objects is incorrect. But if a theory incorporating A1–A4, plus Com, assigns them the same referent, then it must falsely characterize the semantic content of (3′) as a truthconditional consequence of the semantic content of (2′). Either way the theory fails. Hence the reductio stands.

References Barwise, Jon, and John Perry. 1983. Situations and Attitudes. Cambridge: MIT Press. Edelberg, Walter 1994. “Propositions, Circumstances, and Objects.” Journal of Philosophical Logic 23:1–34. Soames, Scott. 1987a. “Direct Reference, Propositional Attitudes, and Semantic Content.” Philosophical Topics 15:47–87. ———. 1987b. “Substitutivity.” In On Being and Saying: Essays for Richard Cartwright, ed. J. J. Thomson, 99–132. Cambridge: MIT Press. ———. 1989. “Direct Reference and Propositional Attitudes.” In Themes from Kaplan, ed. Joseph Almog, John Perry, and Howard Wettstein with the assistance of Ingrid Deiwiks and Edward N. Zalta, 393–419. Oxford: Oxford University Press.

ESSAY THREE

Belief and Mental Representation

In “Propositional Attitudes” Jerry Fodor argues that beliefs and other propositional attitudes are “relations between organisms and internal representations.”1 However, this view is far from transparent, and Fodor fails to point out that it has several different interpretations. Most relevant are a strong interpretation, expressible as (1a) or (1b), and a weak interpretation, given as (2a) or (2b).2 (1) a. For all declarative sentences S of English, there is a mental representation M such that for all individuals i (and times t), i satisfies (at t) x believes that S, as used in a context of utterance C, iff i bears a certain relation R to M (at t), and the content of S in C is identical with the content of M, when taken as one of i’s mental representations (at t). b. For all propositions p, there is a mental representation M such that for all individuals i (and times t), i believes p (at t) iff i bears a certain relation R to M (at t), and p is the content of M, when taken as one of i’s mental representations (at t). A version of this essay was read at the August 19, 1988, Philosophy Department colloquium at the University of Washington. The material in it was first presented in a seminar there that I taught jointly with Charles Marks during the summer of 1987. I am indebted to Professor Marks for illuminating discussions of Fodor’s position and to other participants in the seminar and colloquium and to Ali Akhtar Kazmi for helpful comments. 1 Fodor (1981, 177). 2 The quantifier phrase “a certain relation R” is to be understood throughout as having wide scope over everything else; thus R is invariant across individuals, times, etc. The difference between the interpretations in (1) and those in (2) involves the scope of existential quantification over mental representations. Other interpretations of Fodor’s thesis involving intermediate scope for this quantification will be considered below. In what follows, I will not focus on the differences between the (a) and (b) versions of theses (1) and (2). Since Fodor’s primary intention is to give an analysis not just of propositional attitude ascriptions in English, but of propositional attitudes generally, I believe the (b) versions to be closer to his intent. However, he seems to prefer to talk in terms of “content” rather than “propositions,” and so might prefer the (a) versions after all. (He ignores the need to relativize content to context, and does not formulate his proposals in the explicit manner of either (a) or (b)). Fortunately, any ambivalence on this score will make no difference to us, since my main points are independent of the differences between the (a) and (b) versions of the theses.

82 • Essay Three

(2) a. For all declarative sentences S of English, individuals i (and times t), i satisfies (at t) x believes that S, as used in a context C, iff there is (at t) a mental representation M such that i bears a certain relation R to M (at t), and the content of S in C is identical with the content of M, when taken as one of i’s mental representations (at t). b. For all propositions p, individuals i (and times t), i believes p (at t) iff there is (at t) a mental representation M such that i bears a certain relation R to M (at t), and p is the content of M, when taken as one of i’s mental representations (at t). In distinguishing these proposals it is important to bear in mind two limiting cases in which the differences between them are minimized. One case involves the assumption of a many-one correspondence between (believable) propositions and (a subset of) mental representations. Suppose that for every proposition that could be believed, there were exactly one mental representation that was capable of expressing it (in at least one context). Then, if x believed p in the sense of (2), it would follow that x bore R to a representation M that everyone who believed p had to bear R to. Hence the necessary conditions for believing p laid down by (1) would be satisfied. They would also be satisfied if R were so defined that it was impossible for an individual to bear R to a representation M without simultaneously bearing it to all other mental representations (including those of others) with the same content as M. In the first of these limiting cases the assumption of a many-one correspondence has the effect of strengthening (2) so that it encompasses (1). In the second limiting case a permissive definition of R weakens (1) to the point that it is subsumed by (2). Outside of these limiting cases the proposals are substantially different. The second limiting case is ruled out by Fodor’s insistence that the relation R be “syntactically,” rather than semantically, defined. This is part and parcel of his view that any adequate philosophical analysis of belief must mesh with empirical accounts of mental processes—in particular, those of cognitive psychology. (This is his “condition V” on an adequate analysis of belief.) Thus, he says: Condition V, it will be remembered, permits us to choose among theories of PAs [propositional attitudes] in virtue of the lexico-syntactic form of the entities they assign as objects of the attitudes . . . it’s not just cost-accounting [i.e., ranking beliefs by their psychological complexity] that is supposed to be determined by formal aspects of the objects of PAs; it’s all the mental processes and properties that cognitive psychology explains. That’s what it means to speak of a computational

Belief and Mental Representation • 83

psychology. Computational principles are ones that apply in virtue of the form of entities in their domain. (1981, 201) For Fodor, mental representations are syntactic objects on which individuals perform computational operations. For the belief relation R to hold between an individual i and a representation M is for computations involving M to play a certain kind of functional role in the mental life of i. It is not enough for M to have the same content as some other formula on which the computations are actually performed. Thus, in explaining why English sentences cannot play the role of objects of R he cites believers who do not know English, and hence do not use English sentences. Of course, relations are cheap; there must be some relation which a dog bears to ‘it’s raining’ iff the dog believes that it’s raining; albeit, perhaps, some not very interesting relation. So, why not choose it as the relation in virtue of which the belief-ascription holds of the dog? . . . To put it generally if crudely, satisfying condition V depends on assuming that whatever the theory takes to be the object of a PA plays an appropriate role in the mental processes of the organism to which the attitude is ascribed. But English sentences play no role in the mental life of dogs. (1981, 192) In light of this it is clear that R is to be understood in a way that does not result in the collapse of the theses in (1) into those in (2). The other limiting case—in which the theses in (2) are strengthened to encompass those in (1)—is one in which a many-one correspondence between (believable) propositions and mental representations is assumed. Although Fodor does end up making this assumption, it does not provide a rationale for failing to distinguish (1) from (2). On the contrary, he makes the assumption because he runs (1) and (2) together, without exploring their differences. Throughout most of the article, Fodor writes as if he were proposing (1). However, at various places in the discussion it seems as if (2) were at issue. For example, in defending his view against the objection that belief is a relation to a proposition, he indicates that as long as the relation is understood to be one that is mediated by mental representations, it need not conflict with his view. I am taking seriously the idea that the system of internal representations constitutes a (computational) language . . . nothing stops us from specifying a semantics for the IRS [internal representational system] by saying (inter alia) that some of its formulae express propositions. If we do say this, then we can make sense of the notion that

84 • Essay Three

propositional attitudes are relations to propositions—viz., they are mediated relations to propositions, with internal representations doing the mediating. This is, quite generally, the way that representational theories of the mind work. So, in classical versions, thinking of John (construed opaquely) is a relation to an “idea”—viz., to an internal representation of John. But this is quite compatible with its also being (transparently) construable as a relation to John. In particular, when Smith is thinking of John, he (normally) stands in relation to John, and does so in virtue of his standing in relation to an idea of John. Similarly, mutatis mutandis, if thinking that it will rain is standing in relation to a proposition, then, on the present account, you stand in that relation in virtue of your (functional/causal) relation to an internal formula, which expresses the proposition. (1981, 200–201) This passage strongly suggests the theses in (2). To think of John one must have an idea that bears the right relation to him. However, no one of the potentially many ideas of John is privileged over all others; there is no one idea such that to think of John one must entertain it. If, as Fodor claims, believing a proposition p is like this, then there should be no single mental representation such that to believe p one must stand in the appropriate mental relation to it. Rather, there should be potentially many mental representations that express p and are capable of mediating belief in it. Nevertheless, it is evident that this is not the picture that Fodor has in mind. For example, in claiming that propositional attitudes are relations to internal representations, he is claiming that “in particular, the verb in a sentence like ‘John believes it’s raining’ expresses a relation between John and something else, and a token of that sentence is true if [and only if]3 John stands in the belief relation to that thing” (178). Fodor could not maintain that beliefs are relations to mental representations in this sense, if he thought that (2) were true and that it were possible for different mental representations to have the same content (while being independently possible objects of the relation R).4 Fodor’s commitment to (1) is most clearly illustrated by the way he motivates his view. After arguing that relations to linguistic representations 3 It is clear from the context that Fodor intends to give necessary as well as sufficient conditions for the truth of belief ascriptions. From time to time he informally expresses such truth conditions using conditionals, when strictly speaking biconditionals are in order. Another clear example of this occurs in Fodor (1981, 180, numbered paragraph 5). 4 Under these assumptions representations are not objects of the belief relation in the sense indicated in the quoted passage. For if M1 and M2 are different representations with the same propositional content, then (2) will allow John believes that S to be true either in virtue of

Belief and Mental Representation • 85

play an important role in the analysis of propositional attitudes, he considers a “Neo-Carnapian” analysis, which serves as a prototype of his own theory. Fodor initially characterizes this analysis as one in which propositional attitudes are construed “as relations between people and sentences they are disposed to utter, e.g., between people and sentences of English” (187). He does not regard the behavioral characterization of the belief relation in terms of verbal dispositions as crucial to the account, and suggests that it might be functionally characterized instead. However, he does indicate that the analysis is to be a strong one, in which to believe that Bill bit Mary is to bear the belief relation to “Bill bit Mary,” and to believe that it is raining is to bear the belief relation to “it is raining” (rather than to any sentences that happen to have the same contents as these sentences) (188). In short, Fodor’s prototheory is to be understood along the lines of (3a) or (3b).5 (3) a. For any sentence S of English, there is a sentence S′ of English such that for any individual i (and time t), i satisfies (at t) x believes that S, as used in a context C, iff i bears a certain relation R to S′ (at t), and the content of S in C is identical with the content of S′ in an appropriately related context with agent i (and time t). b. For any proposition p, there is an English sentence S′ such that for any individual i (and time t), i believes p (at t) iff i bears a certain relation R to S′ (at t), and p is the content of S′ in an appropriately related context with agent i (and time t). Fodor uses this prototheory to motivate his own account. According to him the prototheory accommodates most of the observations that indicate that beliefs are relations to sentences, but fails to account for certain obvious facts. For example: (i)

the fact that it is possible to believe that it is raining without understanding “it is raining,” or any other English sentence;

John’s bearing R to M1 (but not M2) or in virtue of John’s bearing R to M2 (but not M1). But then neither M1 nor M2 is such that the belief ascription is true iff John bears R to it. The relation Fodor is talking about in the quoted passage is clearly the semantic extension of the belief predicate. According to (2) the extension of “believe” is a relation between individuals and propositions, not mental representations. Nevertheless, it is a relation which holds between an individual and a proposition in virtue of another, psychological, relation holding between the individual and a mental representation. If (2) is correct, and if different representations can express the same proposition, then it is only in this latter psychological sense that beliefs are relations to representations. 5 Fodor’s own discussion ignores context sensitivity and assimilates S' to S. The formulation in the text is meant to correct this, and to bring out the parallels with (1).

86 • Essay Three

(ii) the fact that in certain cases different English sentences express the same belief (e.g., “John bit Mary” and “Mary was bitten by John”); (iii) the fact that some beliefs are inexpressible in English; and (iv) the fact that some individuals hold beliefs without having learned any natural language.6 In summarizing these objections Fodor notes that (i) “would be without force if only everybody (viz. every subject of true propositional attitude ascriptions) talked English,” that (ii) and (iii) “depend on the empirical likelihood that English sentences fail to correspond one to one to objects of propositional attitudes,” and that (iv) “would be met if only English were innate” (197). His strategy is simply to posit an internal language of mental representations with precisely the properties needed to avoid these objections, and to substitute its formulas for sentences of English as objects of the belief relation R in the prototheory. Indeed, I suppose an ultra hard-line Neo-Carnapian might consider saving the bacon by claiming that—appearances to the contrary notwithstanding—English is innate, universal, just rich enough, etc. My point is that this is the right kind of move to make; all we have against it is its palpable untruth. Whereas, it’s part of the charm of the internal language story that, since practically nothing is known about the details of cognitive processes, we can make the corresponding assumptions about the internal representational system risking no more than gross implausibility at the very worst. So, let’s assume—what we don’t, at any event, know to be false— that the internal language is innate, that its formulae correspond one to one with the contents of propositional attitudes (e.g., that “John bit Mary” and “Mary was bitten by John” correspond to the same “internal sentences”), and that it is as universal as human psychology; viz., that to the extent that an organism shares our mental processes, it also shares our system of internal representations. On these assumptions, everything works. (197) 6 It should be noted that (iii) is an objection only to (3b), and that in general the objections presuppose a strict, impermissive characterization of the relation R. For example, (ii) is an objection only if there are English sentences with the same content such that an individual can bear R to one without bearing R to the others. I have used (iv) to cover two Fodorian objections—the fact that humans must have prior beliefs in order to learn an initial natural language, and the fact that some organisms have beliefs even though they never learn languages.

Belief and Mental Representation • 87

The end result is Fodor’s theory, (1), of propositional attitudes. This result was not inevitable. Fodor’s objections to the prototheory are handled just as well by (2) as they are by (1). In fact, had he started with Carnap’s actual theory,7 (4)

An individual i satisfies x believes that S in English iff there is a sentence S′ and language L′ such that i is disposed to assertively utter (bears R to) S′ as a sentence of L′, and S′ in L′ is intentionally isomorphic to (has the same content as) S in English

several objections to the prototheory would not have arisen in the first place.8 At most, objection (iv) might prompt a move from (4) to (2). However, this leaves the choice between (1) and (2) wide open. To make this choice we need to look closely at the ways in which these theses differ.

7 Carnap (1956, 62). In appendix C, Carnap modified his view slightly to make it less behavioristic. In the modified view, having a disposition to utter S' is no longer treated as a necessary or sufficient condition for bearing the relation R in the analysis to S', but rather is seen as providing strong inductive evidence that one bears R to S'. It should also be noted that Carnap’s model of English was highly simplified, and did not include context sensitivity. 8 In discussing objection (ii) above Fodor says the following:

The natural way to read the Carnap theory is to take type identity of the correspondents of belief-ascribing sentences [i.e., S in x believes that S] as necessary and sufficient for type identity of the ascribed beliefs: and it is at least arguable that this cuts the PAs [propositional attitudes] too thin. . . . A way to cope would be to allow that the objects of beliefs are, in effect, translation sets of sentences; something like this seems to be the impetus for Carnap’s doctrine of intentional isomorphism. (1981, 191) However, Fodor finds this appeal to sameness of content problematic. It may well be, for example, that the right way to characterize a translation relation for sentences is by referring to the communicative intentions of speaker/hearers of whatever language the sentences belong to. (S1 translates S2 if the two sentences are both standardly used with the same communicative intentions.) But, of course, we can’t both identify translations by reference to intentions and individuate propositional attitudes (including, n.b., intentions) by reference to translations. (191) Applied to (4), which doesn’t mention translation, this objection appears to be based on a commitment to reduce semantic contents of natural language sentences to propositional attitudes of speakers—which in turn are to be explained in terms of semantic contents of sentences of some type. One who does not share this reductionist program need not share Fodor’s objection to Carnap’s actual proposal. Moreover, even if the reductionist program is accepted, the objection to appealing to different natural language sentences with the same content in an analysis of belief does not rule out appealing to different mental representations with the same content. Thus it provides no grounds for preferring (1) to (2).

88 • Essay Three

Theoretically Significant Differences between (1) and (2) The first point to notice is that the proposals in (1) carry the strong presumption that each proposition is expressed by at most a single mental representation. For if a proposition p were expressed by more than one such representation—M, M′, M*—(in the same or different contexts), then there would be no evident reason why bearing R to any of them shouldn’t count as believing p. Since this is allowed by the proposals in (2), they do not carry any presumption that every proposition is expressed by only one mental representation. The proposals in (1) also entail that the propositions expressed by mental representations can vary from one context to another only if those propositions are cognitively inaccessible outside of certain privileged contexts. For suppose that M expresses p in context C and fails to express p in other contexts (either because it expresses other propositions or because it expresses no propositions in those contexts). Suppose further that the agent i of C believes P (at the time and place of C) by virtue of bearing R to M. It will then follow from (1) that for all C′ in which M does not express p, the agent i' of C′ does not believe p (at the time and place of C′)—no matter what other mental representations i' may bear R to (at that time and place). The implausibility of this result can be illustrated by applying it to the prototheory (3b). For example, sentence (5a) Classes start today expresses different propositions on different days. It follows from (3b) that if a person believes p by virtue of accepting (bearing R to) (5a) on day d, then p cannot be believed on any other day—even if one accepts (5b) on day d − 1, (5c) on day d + 1, or (5e) on an arbitrary day (where “day d” is a proper name of a day). (5) b. Classes start tomorrow. c. Classes started yesterday. d. Classes start on day d. In short, the proposition expressed by (5a) on day d is cognitively inaccessible on other days. Analogous results hold for other context-sensitive sentences. For instance, if I believe that I live in New Jersey by virtue of accepting (6a) I live in New Jersey, and Ruth believes that she does not live in New Jersey by virtue of accepting

Belief and Mental Representation • 89

(6b) I do not live in New Jersey, then no one but me can believe that I live in New Jersey, and no one but Ruth can believe that she doesn’t live in New Jersey, and no one at all can believe the conjunction of our two beliefs—namely that I live in New Jersey and she doesn’t. Such results may well seem absurd. However if mental representations are allowed to express different propositions in different contexts, then they will apply to Fodor’s theory (1) as well. In light of this it is extremely tempting for a proponent of (1) to deny that mental representations are context sensitive—a proposal implicit in Fodor’s assumption that the formulas of our mental language “correspond one to one with the contents of propositional attitudes” (197). However, tempting or not, this strategy is unacceptable. One immediate consequence of it is to exclude indexicals from the system of mental representations. How then would one handle cases in which an individual expressed a belief p about an object o using a natural language indexical to refer to o? Presumably, the use of the indexical would have to be mapped onto a nonindexical representation that differed from the representation for every other object, while being identical with the one used by all other agents to believe the proposition p, about o. Thus, every moment of time about which one could express a belief using the indexical “now” would have to have its own unique name—the same for all believers—in the system of mental representations; similarly for every person and every grain of sand about which beliefs could be expressed using ‘I’ or ‘that.’9 The implausibility of this is an indictment of the strategy of trying to do away with indexicality in mental representations. 9 This result requires each object to have at least one mentalese name (unique to it). It also requires that if an agent believes a proposition p about an object o by virtue of bearing R to a representation containing a mentalese name n (connected to a use of an indexical that refers to o), then every agent who believes p must do so in virtue of bearing R to M, and hence must use n to refer to o. Since the point of the strategy is to ensure that propositions expressed by indexical sentences of natural language are cognitively accessible to believers generally, there is no avoiding the conclusion that different agents must use the same mentalese names for the same objects. It is worth noting that these requirements do not depend on any particular semantic analysis of names or indexicals; in particular they take no stand on the question of whether or not the semantic contents of such terms, relative to contexts, are their referents, relative to those contexts. The strategy requires that each content (propositional constituent) expressed by a mentalese name be expressed by that name alone. So long as this is guaranteed, it is noncommittal about the nature of the contents themselves. (If contents are referents, then an object can be named by only one mentalese name; if they are more fine grained than that, then an object might have more than one such name.) I am indebted to Ali Akhtar Kazmi for a discussion of this point.

90 • Essay Three

Moreover, the problem goes beyond implausibility. Like most proponents of mental representations, Fodor believes that they are “in the head.” Thus, intrinsic properties of an individual’s brain (at a given time) should determine which representations he entertains (at that time). One consequence of this is that physically indistinguishable individuals—e.g., molecule-for-molecule duplicates—must entertain the same representations.10 But then the nonindexical strategy just sketched will fail to handle indexical variants of familiar “Twin-Earth” cases.11 One such variant is David Kaplan’s example of identical twins, Castor and Pollux, who are molecular and behavioral duplicates.12 Each expresses a belief by sincerely uttering “I am older than my brother.” Since one of the twins thereby believes a true proposition (about Pollux), whereas the other believes a false proposition (about Castor), the propositions believed are different. However, since the twins are duplicates, the beliefs must arise from the same mental representation. Thus, the strategy of mapping noncoreferential uses of natural language indexicals onto different nonindexical mental representations cannot succeed. This leaves the proponent of (1) with no successful way of avoiding the seemingly absurd result that propositions expressed by indexical sentences, like those in (5) and (6), cannot be believed (or even apprehended) except in contexts in which they are expressed by those very sentences. Standard Twin-Earth cases involving names and natural kind terms show that the same sort of absurdity arises even when indexicals are not involved. For example, suppose that Oscar and Oscar′ are molecular duplicates who use the same natural kind term N to name different but observationally indistinguishable natural kinds. Both express beliefs by uttering  N is F. However the propositions believed are different; Oscar believes the proposition that kind K is F; Oscar′ believes that K′ is F. Since they are duplicates, their beliefs arise from the same representation M, which must have one propositional content when entertained by Oscar and a 10 The reason for insisting that mental representations be “in the head” in this sense involves the theoretical role of such representations in cognitive theories, as understood by Fodor and others. Mental representations are meant to capture the contributions of an agent’s internal cognitive states to explanations of his behavior. Since molecular duplicates can be assumed to be internally the same, they must be assigned the same mental representations. This is reflected in Fodor’s insistence that mental representations are syntactically, rather than semantically, individuated formulas in an internal language, and that psychological processes are formal computational operations on these syntactic objects. 11 The original Twin-Earth case was presented by Hilary Putnam (1975). The characteristic feature of such cases is that of physically and psychologically indistinguishable individuals with different beliefs. Although Putnam’s original case involved beliefs expressed using the natural kind term “water,” the phenomenon generalizes to all terms whose contents are (partially) determined by factors “outside the head.” 12 Kaplan (1989).

Belief and Mental Representation • 91

different content when entertained by Oscar′. It follows from (1) that each of these propositions p is such that an individual can believe p only if he bears R to M in a context in which p is the content of M. This means that as long as Oscar retains his belief that K is F, it will be impossible for him to acquire the belief that K′ is F. No matter that he may do what any other believer that K′ is F might do—namely encounter K′, use a natural kind term N′ to name it, and sincerely assent to N′ is F.13 As long as he continues to believe that K is F, the content of M, when entertained by him, must be that proposition—rather than the proposition that K′ is F. Thus, (1) will tell us that he cannot believe that K′ is F. In fact, as long as M cannot simultaneously have different contents when entertained in the same context by the same individual, (1) will tell us that no one can believe both propositions. Surely this is incorrect. In sum, the proposals in (1) cannot accommodate the following elementary facts: a. The same formula may express different propositions when interpreted from the point of view of different contexts, and different representational systems. b. The same proposition may be expressed by different formulas of the same system, or of different systems. Faced with obvious instances of (a) in natural language, the proponent of (1) has two choices. He may attempt to explain away the relativity of linguistic content as a superficial feature of natural language which disappears at the more fundamental explanatory level of mental representation, or he may allow such relativity in mental representations at the cost of making their contents cognitively inaccessible in a great many cases. Both of these alternatives are dead ends. The first requires the incredible assumption that each object about which one can have a belief p has a nonindexical mentalese representation, unique to it, the same for all believers of p—an assumption refuted by the various Twin-Earth cases. The second alternative founders from not being able to accommodate (b). Two individuals—Ruth and I, Castor and Pollux, Oscar and Oscar′—may share beliefs by accepting different sentences that express the same proposition. Since the proposals in (1) cannot account for these facts, they should be rejected.

Philosophical Advantages of (2) In addition to giving a more accurate account of the fundamental facts about propositional attitudes, the proposals in (2) avoid a number of 13

While rejecting, or suspending judgment on N = N'.

92 • Essay Three

undesirable philosophical implications of those in (1). One such implication involves the relationship between mental representations and natural languages. Whereas (2) comfortably allows the possibility that an individual’s system of mental representations may contain elements of the natural language(s) he speaks, (1) does not. For example, (1) requires everyone who believes that many people suffer from Acquired Immune Deficiency Syndrome to do so by virtue of being related to the same mental representation. Presumably it is possible for non-English speakers to believe this proposition even though their systems of mental representations do not include English terms like “Acquired Immune Deficiency Syndrome.” But then it follows from (1) that even the mental representations in virtue of which monolingual speakers of English believe that many people suffer from Acquired Immune Deficiency Syndrome do not include any English terms. In short, (1) implicitly denies that the natural language expressions we learn are directly involved in the propositional attitudes we hold. It is an advantage of (2) that it does not prejudge the issue in this way. A further point, closely related to this, involves Fodor’s startling claim that mentalese is entirely innate, and that natural language expressions are learned by connecting them with already understood mentalese counterparts.14 If, as (2) allows, the system of mental representations of a given person includes elements of his natural language, then, since natural language expressions are not innate, it will follow that a person’s system of mental representations is not entirely innate either—a welcome result. It will also follow that it is possible for the explanation of some of an individual’s beliefs to depend on the content of certain of his mental representations, which in turn may depend on the content of expressions in a public language. This picture fits well with accounts that see the contents of, for example, names and natural kind terms as arising from social, historical, and causal connections relating a speaker’s use of these terms to other speakers, and ultimately to objects in the world. On such an account there is a social process that is crucially involved in determining the content of some expressions in a public language, which in turn are responsible for the contents of some beliefs. It should be noted that this picture is at variance with Fodor’s reductionist conception in which the intentionality of natural language is reduced to the intentionality of propositional attitudes, which is then supposed to be reduced to the intentionality of mentalese. Although the proposals in (1) do not by themselves dictate this reductionist picture, they reinforce it by conferring an undeserved primacy on mentalese at the expense of natural language. 14 Fodor’s argument for this is given in Fodor (1975, chap. 2). A thorough critique of the argument can be found in Sterelny (1989).

Belief and Mental Representation • 93

Finally, there is the question of the epistemological and metaphysical status of (1). Fodor’s intention in offering his theory of propositional attitudes as relations to representations is to present an empirical, a posteriori thesis that nevertheless gives the essence of propositional attitudes, and supports counterfactual generalizations. This emphasis on specifying the essence of a phenomenon, though natural in a philosophical theory, suggests an assertion of necessity that is difficult to incorporate into the theses in (1). For example, consider the following attempts to strengthen (1b) to cover counterfactual circumstances. (7) a. Necessarily for all propositions p, there is a mental representation M such that for all individuals i (and times t), i believes p (at t) iff i bears R to M (at t), and p is the content of M, when taken as one of i’s mental representations (at t). b. For all propositions p, there is a mental representation M such that necessarily for all individuals i (and times t), i believes p (at t) iff i bears R to M (at t), and p is the content of M, when taken as one of i’s mental representations (at t). Neither of these is attractive. The former implies that in each world there is one representation to which individuals must bear R in order to believe a given proposition p. However, it does not require representations carrying belief in p to remain invariant across worlds. Thus, it allows an individual to believe p by bearing R to M1 in W1, M2 in W2, and so on (M1 ≠ M2 . . . ). But if one allows variation across worlds, it is hard to find grounds for denying it within worlds. In particular, if different representations can carry the same belief under different conditions, then either extensive empirical investigation or further conceptual analysis is needed to establish that these conditions are not jointly satisfied in the actual world. Since neither has been (or is likely to be) forthcoming, the proponent of (1) is pushed inexorably to (7b). But this just makes things worse by compounding the problems of the previous section to include cross-world cases in addition to the intra-world cases already discussed.15 15 On page 202, Fodor shows some reluctance to accept either (7a) or (7b). First he grants that it is conceivable that propositional attitudes are not relations to internal representations, by which he means that his thesis that propositional attitudes are relations to internal representations is not a priori. Next he considers the objection that it is empirically possible that propositional attitudes are not relations to internal representations.

it may be empirically possible that there should be creatures that have the same propositional attitudes we do (e.g., the same beliefs) but not the same system of internal representations; creatures that, as it were, share our epistemic states but not our psychology. Suppose, for example, it turns out that Martians, or porpoises, believe what we do but have a very different sort of cost accounting. We might then want to say that there are

94 • Essay Three

Other Interpretations of Fodor’s Thesis There is, then, ample reason to regard the proposals in (2) as more acceptable versions of the thesis that beliefs are relations to mental representations than those in (1). However, before settling on (2), we would do well to note that there are interpretations of Fodor’s thesis that are intermediate between (1) and (2). (1.5) a. For all propositions p, individuals i (and times t), there is a mental representation M such that i believes p (at t) iff i bears a certain relation R to M (at t), and p is the content of M, when taken as one of i’s mental representations (at t). b. For all propositions p, individuals i (and times t), there is a mental representation M with content p, when taken as one of i’s mental representations (at t), such that i believes p (at t) iff i bears R to M (at t). These proposals do not have all the obvious difficulties of (1). Thus, the question arises as to whether or not they are acceptable alternatives to (2). They are not. All (1.5a) tells us is that for any proposition p, individual i, (and time t) there will be at least one representation M that makes the following biconditional true: i believes p (at t) iff i bears a certain relation R to M (at t), and p is the content of M, when taken as one of i’s mental representations (at t). One way for a biconditional to be true is for both sides to be false; and one way for the right-hand side of this biconditional to be false is for p not to be the content of M. Thus, (1.5a) leaves open the bizarre possibility that i might fail to believe p (e.g., that i lives in New Jersey) simply by virtue of the fact that some mental representation (e.g., one underlying ‘2 + 2 + 4’) does not have p as content (when taken as one of i’s mental representations at a given time). This possibility could be discounted if it could be guaranteed that bearing R

translation relations among systems of internal representation (viz., that formally distinct representations can express the same proposition). . . . Whether we can actually make sense of this sort of view remains to be seen. (1981, 202) Fodor’s mention of translation seems to be a way of indicating a potential retreat from (1) to (2), should the objection prove correct. But what would it be for it to be correct? If the possibility mentioned in the objection is the normal non-epistemic kind, the correctness of the objection depends not just on what turns out to be the case in the actual world, but also on what could have been the case (in counterfactual circumstances). Once it is admitted that there could have been creatures of the sort Fodor hypothesizes, (7b) must be rejected, as must (7a), if it is further acknowledged that these creatures could have coexisted with us in a world in which we retained our actual psychology.

Belief and Mental Representation • 95

to a representation expressing p were sufficient for believing p. However, (1.5a) doesn’t guarantee this. It fails to predict that if (at t) i bears R to some, or even all, mental representations of i that have p as content, then i believes p (at t). Thus, (1.5a) is too weak. This doesn’t mean that it is untrue; since it is a consequence of (2b) it had better be true. It does mean that (1.5a) cannot be an adequate theory of belief. Thesis (1.5b) suffers from different, but related problems. According to it, bearing R to all one’s representations that have p as content is sufficient for believing p, and bearing R to at least one such representation is necessary for believing p. However, (1.5b) is silent about cases in which one bears R to some but not all of one’s representations that have p as content.16 Moreover, it requires something that is obviously too strong—namely, that each believer have (at any given time) mental representations sufficient to express all propositions. Thus, it too is inadequate.17 16 In my view, it is possible for an individual i to believe p by virtue of bearing R to a representation M1 that expresses p, even if he fails to bear R to another of his representations, M2, that (unknown to him) also expresses p. This point is connected to the discussion of examples (9)–(13) in the next section, and to the literature cited in note 21. 17 Fodor (1987, 17) makes the following claim.

Claim 1 (the nature of propositional attitudes): For any organism O, and any attitude A toward the proposition P, there is a (“computational”/“functional”) relation R and a mental representation MP such that MP means that P, and O has A iff O bears R to MP. Aside from the obviously undesirable feature of allowing the relation R to vary not just from attitude to attitude (belief, hope, desire, etc.) but also from organism to organism and proposition to proposition, this formulation is a straightforward generalization of (1.5b), and is therefore inadequate for the same reasons as (1.5b). However, it is not clear that Fodor’s formulation really captures his intention. For example, in explaining what he means by the claim, he says: A cruder but more intelligible way of putting claim 1 would be this: To believe that such and such is to have a mental symbol that means that such and such tokened in your head in a certain way; it’s to have such a token ‘in your belief box,’ as I’ll sometimes say. (1987, 17) But this is an informal statement of (2), rather than (1.5b). Moreover, on page 20 Fodor says that claim 1 has the following consequences: For each tokening of a propositional attitude, there is a tokening of a corresponding relation between an organism and a mental representation; and for each tokening of that relation, there is a corresponding tokening of a propositional attitude. Although both of these are consequences of (2), only the first is a consequence of claim 1 (or of 1.5b). So perhaps by the time of Psychosemantics Fodor really means to adopt (2). If so, this is a significant and unacknowledged change from his position in “Propositional Attitudes.” I am indebted to Ali Akhtar Kazmi for directing my attention to Fodor’s comments on page 17 of Psychosemantics.

96 • Essay Three

Propositions, Mental Representations, and the Explanation of Behavior In light of this, it appears that (2) is the most promising version of the thesis that beliefs are relations to mental representations. However, if one does adopt (2), one must be careful to note that the objects of belief— the things believed—are contents or propositions expressed by mental representations, rather than mental representations themselves. Moreover, one must build into these propositions much of the syntactic structure of the representations that express them; otherwise one will lose benefits crucial to Fodor’s appeal to representations in the first place. Such benefits are illustrated by the following remarks. It’s plausible to claim that there is a fairly general parallelism between the complexity of beliefs and the complexity of sentences that express them. So, for example, I take it that “the Second Punic War was fought under conditions which neither of the combatants could have desired or forseen” is a more complex sentence then, e.g., “it’s raining”; and, correspondingly, I take it that the thought that the Second Punic War was fought under conditions which neither of the combatants could have desired or forseen is a more complicated thought than the thought that it’s raining. (1981, 188–89) The view that beliefs are relations to mental representations is meant to explain this parallelism. A theory of propositional attitudes specifies a construal of the objects of the attitudes. It tells for such a theory if it can be shown to mesh with an independently plausible story about the “cost accounting” for mental processes. A cost accounting function is just a (partial) ordering of mental states by their relative complexity. Such an ordering is, in turn, responsive to a variety of types of empirical data, both intuitive and experimental. Roughly, one has a “mesh” between an empirically warranted cost accounting and a theory of the objects of PAs when one can predict the relative complexity of a mental state (or process) from the relative complexity of whatever the theory assigns as its object (or domain). . . . Again, roughly: to require that the complexity of the putative objects of PAs predict the cost accounting for the attitudes is to impose empirical constraints on the notation of (canonical) belief-ascribing sentences. So, for example, we would clearly get different predictions about the relative complexity of beliefs if we take the object of a PA to be the correspondent [complement] of the belief ascribing sentence than if we take it to be, e.g., the correspondent [complement] transformed into disjunctive form. (189–90)

Belief and Mental Representation • 97

In short, Fodor suggests that (8a) and (8b) differ in psychological complexity, and that this difference mirrors the difference in syntactic complexity between (8c) and (8d). (8) a. b. c. d.

the belief that P ⊃ Q the belief that (P & Q) v (− P & Q) v (− P & − Q) P⊃Q (P & Q) v (− P & Q) v (− P & − Q).

This suggestion is plausible. Moreover, if it is correct, then the propositions expressed by (8c) and (8d) cannot be identical. They cannot be identical, since the state of believing one of them places fewer demands on an individual’s psychological system, in a sense that Fodor takes to be measurable, than does the state of believing the other. This means that propositions with the same truth conditions may nevertheless be distinct. In short, propositions cannot be what I have elsewhere called sets of truthsupporting circumstances.18 Let us suppose, with Fodor, that syntactic complexity is a good measure of psychological complexity, which in turn is reflected in the propositions believed. It is then appropriate to view propositions as themselves syntactically structured objects constructed out of the contents of the constituents that make up the representations that express them.19 This conception of content guarantees that although many different representations may express the same proposition, such representations will in general share the same syntactic structure, and hence have the same psychological complexity. This is what is required by the combination of (2) with Fodor’s account of the psychological complexity of beliefs.20 In what sort of cases, then, can we expect different mental representations to express the same proposition? Presumably in cases of the sort illustrated by the following pairs. (9) a. b. (10) a. b. 18

Today is F. (Said at 11:55 p.m.) Yesterday was F. (Said 10 minutes later.) This (pointing at o) is F. That (pointing at o from a slightly different perspective) is F.

Soames (1985, 1988, 1989). See the articles mentioned in the previous note, plus Salmon (1986). 20 The combination of (2) with Fodor’s account of psychological complexity requires contents of sentences to be structured propositions; but this does not mean that if (2) were replaced by (1), the more familiar conception of contents as truth conditions, or sets of truth-supporting circumstances, could be maintained. On that conception there is no way of preventing the construction of semantically equivalent sentences that differ greatly in syntactic complexity—something that would make Fodor’s account of the psychological complexity of beliefs impossible. 19

98 • Essay Three

(11) a. b. (12) a. b. (13) a.

London is pretty. Londres est jolie. Catsup is a tomato-based condiment. Ketchup is a tomato-based condiment. Everyone who has heard of catsup believes that catsup is catsup. b. Everyone who has heard of catsup believes that catsup is ketchup.

I take it that in the case of each of these pairs the two sentences express the same proposition, in the relevant contexts. The question then arises as to whether the two sentences are mapped onto the same or different mental representations (in the contexts). A good reason for assuming the latter is that competent speakers who understand them may nevertheless be unaware that they express the same proposition (in the relevant contexts).21 If such speakers are unaware that the sentences express the same proposition, then they may be expected to react differently to them. Given the role of mental representations in explaining the linguistic and other behavior of speakers, one will thus be led to posit different mental representations underlying the sentences in the relevant contexts.22 This conclusion has important consequences for cognitive explanations of behavior. According to Fodor, the paradigm for such explanations is one where propositional attitudes interact causally and do so in virtue of their content. And the paradigm of this paradigm is the practical syllogism. . . . John believes that it will rain if he washes his car. John wants it to rain. So John acts in a manner intended to be a car-washing. (1981, 183) According to Fodor, this account might be counterfactual-supporting in at least the following sense: John wouldn’t have car-washed had the content of his beliefs, utilities, and intentions been other than they were. . . . To say that John’s mental states interact causally in virtue of their content is, in part, to say that such counterfactuals hold. (183)

21 For discussions supporting this characterization of these examples see: (i) in the case of (9) and (10), Kaplan (1989) and Perry (1979); (ii) in the case of (11), Kripke (1979, appendix A); (iii) in the case of (12), Salmon (1989, 1990); (iv) in the case of (13), Church (1954), Soames (1987, sec. 9), and the introduction to Salmon and Soames (1988). 22 Recall that (2) allows cases in which the mental representation of a sentence contains the words in the sentence, and may even be identical with the sentence itself.

Belief and Mental Representation • 99

Fodor continues: If there are true, contingent counterfactuals which relate mental state tokens in virtue of their contents, that is presumably because there are true, contingent generalizations which relate mental state types in virtue of their contents. So . . . we can schematize etiologies like the one above to get the underlying generalizaton: if x believes that A is an action x can perform; and if x believes that a performance of A is sufficient to bring it about that Q; and if x wants it to be the case that Q; then x acts in a fashion intended to be a performance of A. (183) According to Fodor, explanation in cognitive science requires generalizations of this basic form. In his view, we can’t state the theory-relevant generalization that is instantiated by the relations among John’s mental state unless we allow reference to beliefs of the form if X then Y; desires of the form that Y; intentions of the form that X should come about; and so forth. Viewed one way (material mode), the recurrent schematic letters require identities of content among propositional attitudes. Viewed the other way (linguistically), they require formal identities among the complements [or better, the mental representations of the complements—my addition] of the PA-ascribing sentences which instantiate the generalizations of the theory that explains John’s behavior. (184) But if (2) is correct, and if different mental representations may express the same proposition without the agent realizing that they do, then Fodor’s conception of psychological explanation cannot be accepted as stated. More precisely, claims of the form (14) will not require formal identity of mental representations corresponding to different occurrences of the same schematic letter in specifications of the agent’s beliefs and desires. (14) If x believes that A is an action that x can perform, and if x believes that performing A will bring it about that Q; and if x wants it to be the case that Q; then x will act in a fashion intended to be a performance of A. Because of this, claims of the form (14) cannot be expected to be true, exceptionless, universal generalizations. The reason for this is easy to see. One can believe that doing A will bring it about that Q by bearing the belief relation R to a mental representation (15a). (15a) IF I DO A, THEN M1.

100 • Essay Three

One can desire to bring it about that Q by bearing the desire relation to a mental representation (15b). (15b) M2 In this example M1 and M2 share the same content, that Q. However, if the agent fails to realize this, then he may well lack any inclination to do A. In such a case the agent has the requisite beliefs and desires, described in terms of content; however, he lacks the expected behavioral disposition because he holds the beliefs and desires in an unusual way—namely by virtue of being related to distinct mental representations that appear to him to be unrelated. Since cases like this can be expected to occur on at least some occasions, the pre-established harmony that Fodor imagines holding between explanations of behavior that appeal to contents of propositional attitudes and explanations that appeal to internal computational operations breaks down. In my view, this partial breakdown should not be cause for alarm. Generalizations like (14) require ceteris paribus clauses in any case, independent of any difficulties noted here. It is hard to see why we should be overly concerned about including in such clauses some indication that, for example, the agent apprehends the content that Q in the same way (i.e., via the same representations) in the relevant beliefs and desires. Moreover, whatever questions there may be regarding the details of such generalizations, such questions provide no grounds for a generalized doubt about the correctness of particular pretheoretic explanations of actions in terms of beliefs and desires. John may do A because he wants it to be the case that Q and believes that doing A will bring this about, even if not everyone with those beliefs and desires would do A. This is no more mysterious than the observation that John may fall and break his leg because he steps on a banana peel, even though not everyone who steps on a banana peel suffers a similar fate. Thus, the partial breakdown of Fodor’s pre-established harmony need not threaten most ordinary explanations of actions in terms of beliefs and desires. Nor does it threaten (2). Indeed, I can see no reason to doubt that the analysis of propositional attitudes given in (2) is superior to that given in (1). This is not to say that (2) has been established. Important questions remain about the exact nature and status of mental representations. Although I will not try to answer them here, I do hope to have shown that if such representations play a role in the analysis of propositional attitudes, then they do so along the lines of (2) rather than (1).

Belief and Mental Representation • 101

Appendix: Scientific Psychology and the Notion of Narrow Content In chapter 1 of his recent book Psychosemantics, Fodor sums up his main point about the relationship between psychology and common sense as follows: “An explicit psychology that vindicates commonsense belief/desire explanations must permit the assignment of content to causally efficacious mental states and must recognize behavioral explanations in which covering generalizations refer to (or quantify over) the contents of the mental states they subsume” (1987, 14–15). The contents he has in mind are propositions that specify that which is believed or desired, and are the referents of “that”-clauses (that P, that Q, etc.) in standard propositional attitude ascriptions (1987, 11). The content-involving causal generalizations of the common-sense-vindicating psychology are supposed to parallel familiar commonsense generalizations of the sort illustrated by (14) above. For Fodor, the key feature of this account is the correspondence between the causal powers of mental states and their (propositional) contents (1987, 12). Mental representations are seen as providing the link between the two (1987, 16–19). In particular, beliefs and desires with the same (different) causal powers are thought to be relations to the same (different) representations. If, in addition, beliefs and desires with the same (different) propositional contents involve the same (different) representations, then the contents of beliefs and desires will match up one-to-one with their causal powers. On such a picture it is no wonder that causal explanations of behavior in terms of propositions believed and desired should be effective. The difficulty, of course, is that this picture is inaccurate. I have argued that mental states involving different mental representations, with different causal powers, sometimes share the same propositional content, thereby creating exceptions to explanatory generalizations like (14). Twin-Earth type cases (both indexical and nonindexical) illustrate a related shortcoming. In such cases mental states involving the same representations, with the same causal powers, are assigned different propositional contents, due to differences in the surrounding environment. Although these cases do not threaten the truth of Fodor’s commonsense explanatory generalizations, they do raise the specter of different psychological explanations—in terms of different propositions believed and desired—of identical behavior of molecule-for-molecule twins. Fodor doesn’t like this (1987, 37). In chapter 2 of Psychosemantics, he argues that the notion of content needed for scientific psychology must be individuated in terms of the causal powers of mental states, which in turn supervene on brain-states. Twin-Earth cases show that commonsense (propositional) contents of

102 • Essay Three

beliefs and desires are not so individuated and do not so supervene. Thus, Fodor concludes that scientific psychology requires a notion of “narrow content” different from propositional content (44–45). His proposal is that narrow contents should be identified with functions from “contexts” to propositional contents.23 What, if anything does that mean? Well it’s presumably common ground that there’s something about the relation between Twin-Earth and TwinMe in virtue of which his “water”-thoughts are about XYZ even though my water-thoughts are not. Call this condition that’s satisfied by {TwinMe, Twin-Earth} condition C (because it determines the Context of his ‘water’-thoughts). Similarly, there must be something about the relation between me and Earth in virtue of which my water-thoughts are about H2O even though my Twin’s “water”-thoughts are not. . . . But now we have an extensional identity criterion for mental contents: Two thought contents are identical only if they effect the same mapping of thoughts and contexts onto truth conditions. Specifically, your thought is content-identical to mine only if in every context in which your thought has truth condition T, mine has truth condition T, and vice versa. It’s worth reemphasizing that, by this criterion, my Twin’s ‘water’thoughts are intentionally identical to my water-thoughts; they have the same contents even though, since their contexts are de facto different, they differ, de facto, in their truth conditions. In effect, what we have here is an extensional criterion for “narrow” content. The “broad content” of a thought, by contrast, is what you can semantically evaluate; it’s what you get when you specify a narrow content and fix a context. (48)

23 Fodor (1987) introduces this notion on page 47, where he describes narrow contents as “functions from contexts and thoughts onto truth conditions.” There, Fodor assumes for the sake of argument that propositional contents can be identified with truth conditions. He indicates that he would be content with a more general formulation in terms of functions from contexts and thoughts to propositional contents, if the identification were questioned. His inclusion of thoughts among the arguments of narrow contents (of thoughts) is more puzzling. It is clear that “thought” in this discussion stands for the mental state of believing a proposition, rather than the proposition believed. Since mental state tokens tend not to survive even tiny variations in contexts, a thought whose narrow content allows it to have different propositional contents in a variety of different contexts should presumably be a mental state type. But then, the narrow content of a mental state type will be a function from arbitrary contexts to propositions (it expresses in those contexts), not a function from arbitrary contexts and arbitrary mental state types to propositions (expressed by those types in those contexts). This is how I shall understand Fodor’s proposal. (An alternative would be to take a narrow content as belonging to a mental state token, and as being defined over pairs of contexts and tokens of that type. The issues raised below are neutral between these alternatives, and could be expressed under either formulation.)

Belief and Mental Representation • 103

Fodor’s distinction between narrow and broad content resembles David Kaplan’s distinction between character and content. However, there are important differences. For example, Kaplan observed that whereas speakers of a language use indexicals to express different propositions in different contexts of utterance, nonindexicals are not used in this way. In particular, the propositional contents expressed by English speakers using the word “water” do not vary from one context to another.24 This is reflected in the fact that the characters of nonindexical expressions in a language— names, natural kind terms, other general terms, etc.—are always constant functions (that take the same propositional content as value in each context of utterance). Thus, for all contexts of utterance C and C′, and all nonindexical expressions N and N′ of a language, if the (propositional) content of N in C is identical with the (propositional) content of N′ in C′, then the characters of N and N′ will be identical. This is not true for Fodorian narrow contents. The narrow contents of all representations (expressions) for which Twin-Earth type cases can be constructed—which includes virtually all representations (expressions)25—will be nonconstant functions (that assign different propositional contents as values to different Fodorian contexts).26 As a result, representations E and E′ will share the same narrow content only if they have the same propositional content in all Fodorian contexts, including 24 This will hold even if, descending the steps of my spacecraft after landing on TwinEarth, I say “That [gesturing at Twin-water] is water.” Even though the utterance occurs on Twin-Earth, the content of my utterance of ‘water’ will be the same as the content of ‘water’ utterances on Earth. Of course if I stay on Twin-Earth long enough I may, through interaction with the natives, lapse into Twin-English, in which case my ‘water’ utterances will come to share the content of those of my twins. However, as long as I continue to speak English—as it is spoken back home—my ‘water’ utterances will retain their normal English contents. Note the contrast with a genuine indexical like ‘that.’ My (English) utterance of ‘that’ on Twin-Earth stands for Twin-water, even though no ‘that’ utterances on Earth have ever stood for Twin-water. In Kaplan’s terminology, the content of the English word ‘water’ does not vary from context to context, and similarly for the content of the Twin-English word ‘water.’ Since the contents, and characters, of the two are different, the English word ‘water’ is not identical with the Twin-English word ‘water’; instead, they are phonological and orthographic twins, or homonyms. A similar analysis is given to ordinary cases in which different individuals are said to “have the same name.” In Kaplan’s system, such cases typically involve homonymy— phonologically similar names with different contents that remain fixed across contexts. 25 Fodor (1987, 29) notes the generality of Twin-Earth type cases. 26 Since even nonindexical expressions have nonconstant functions as narrow contents, Fodorian contexts must somehow provide such expressions with propositional contents, in addition to assigning contents to indexicals. Fodor’s goal is to use this framework to give a naturalistic account of how expressions acquire representational content. The general idea seems to be that the propositional content of a representation arises from the causal properties of its tokens, including both relations to mental events entirely internal to the agent (encoded in narrow contents) and relations to other agents and objects in the world (encoded in contexts).

104 • Essay Three

all variations of Twin-Earth type cases. But this condition will almost never be met. For nearly every pair of distinct representations, we can imagine a single context (Twin-Earth type environment) in which they stand for different things.27 Thus, we can typically expect E and E′ to have the same Fodorian narrow contents if and only if E is the same representation as E′.

27 With indexicals, the case is transparent. Imagine a representational system with demonstratives ‘that1’, ‘that2’, etc., which can be used to refer to arbitrary objects (or arbitrary objects within a specified range or ranges). Even if each demonstrative behaves in accord with the same rule (e.g., one that determines its referent or content in a context by tracing its causal connections to objects in the environment), the existence of at least one context in which ‘that1’ and ‘that2’ have different referents/contents will be sufficient to guarantee that they (always) have different narrow contents. Since there will typically be such contexts, different demonstratives can be expected to have different narrow contents. Given the parallel between the narrow contents of indexicals and nonindexicals, as well as their similar roles in Fodor’s account of Twin-Earth cases, one would expect the point to carry over to representations generally. This can be illustrated with the help of Hilary Putnam’s (1975) well-known “elm”/“beech” example. Putnam says that he cannot tell the difference between a beech tree and an elm tree, and even that his “concept” of a beech is the same as his “concept” of an elm. Nevertheless the two words, as used by him, refer to different things—beeches and elms—and therefore have different propositional contents. In each case, Putnam uses the word to express the same propositions about the same things that other speakers do. Thus, each word has the same reference and propositional content for him as it does for other speakers in the linguistic community. In Fodor’s scheme, Putnam would associate different mental representations with the two words and these representations would have different narrow contents, despite the fact that Putnam has no mental resources for distinguishing beeches from elms. The point generalizes across the board. Imagine, for example, a related character, Putnam*, who is in an analogous position regarding ‘ketchup’ and ‘catsup’. Putnam* lives in a community just like ours except for the fact that the red tomato-based condiment that comes in transparent bottles marked ‘ketchup’ is slightly different from the red tomatobased condiment that comes in transparent bottles marked ‘catsup’. Though the difference between the two can be discerned by experts, most speakers, including Putnam*, can’t distinguish them. He may even be unsure whether the two words refer to the same or different things. Still the words ‘ketchup’ and ‘catsup’, as used by Putnam* and other speakers, would, in Fodor’s system, be associated with different mental representations with different narrow contents. Finally, if our own community of English speakers contains one of Putnam*’s molecule-for-molecule twins, then his distinct ‘ketchup’ and ‘catsup’ representations will have different narrow contents, despite the fact that ‘ketchup’ and ‘catsup’ are fully synonymous in English. (This example is adapted from Salmon (1990)). The general point to be noted is that different representations will virtually always have different Fodorian narrow contents, no matter how similarly they function in the internal cognitive economies of their agents. The reason for this is that Fodorian narrow contents are, by definition, sensitive to everything capable of influencing the propositional content of a representation in a context. Since environmental factors are typically capable of this, and since they may interact with distinct representations in different

Belief and Mental Representation • 105

This means that, unlike the case with propositional content, one cannot drive a wedge between representations and their narrow contents. In criticizing the analysis of belief given in (1a) and (1b), I noted that they could not accommodate the following elementary facts: a. The same formula may express different propositions when interpreted from the point of view of different contexts, and different representational systems. b. The same proposition may be expressed by different formulas of the same system, or of different systems. If reference to propositions in (a) and (b) is replaced by reference to Fodorian narrow contents, then this criticism will not apply. However, the same replacement will not save the proposals in (1). The resulting principle in the case of (1a) is just like the original except for the addition of “narrow” before each of the two occurrences of ‘content.’ The falsity of this principle is easily shown. Imagine that I believe that I live in New Jersey by virtue of bearing R to representation M. Now imagine a different individual J in an epistemic circumstance similar to mine. Like me, J believes that he lives in New Jersey. J also says to himself “I live in New Jersey,” and J also bears R to M. According to the modified principle (la-narrow), J satisfies “x believes that I live in New Jersey” when said by me in a context C with me as agent. But this is wrong. J doesn’t believe that I live in New Jersey; J believes that he lives in New Jersey. Thus, the appeal to narrow content won’t save (1a). Nor will it save (1b). If reference to narrow contents replaces reference to propositions, then (1b-narrow) will identify narrow contents as the objects of belief. Since what we believe are propositions, this is incorrect. Fodor is well aware of this. . . . the content that an English sentence expresses is ipso facto anchored content, hence ipso facto not narrow. So, in particular, qua expression of English “water is wet” is anchored to the wetness of water (i.e., of H2O) just as, qua expression of Tw-English, “water2 is wet” is anchored to the wetness of water2 (i.e., of XYZ). And of course, since it is anchored to water, “water is wet” doesn’t— can’t—express the narrow content that my water-thoughts share with ways, distinct representations will in general have different Fodorian narrow contents. Exceptions are possible only when factors entirely internal to the agent determine that a pair of different internal representations must have the same content, independent of external considerations. For example, it might be possible (even if pointless) to have two formally distinct internal representations that functioned as first person singular pronouns with the same narrow content. However, outside unusual cases like this the generalization should hold.

106 • Essay Three

my Twin’s. Indeed, if you mean by content what can be semantically evaluated, then what my water-thoughts share with Twin “water”thoughts isn’t content. Narrow content is radically inexpressible, because it’s only content potentially; it’s what gets to be content when—and only when—it gets to be anchored. We can’t—to put it in a nutshell—say what Twin thoughts have in common. This is because what can be said [my emphasis] is ipso facto semantically evaluable; and what Twinthoughts have in common is ipso facto not. . . . Looked at the other way around, when we use the content of a sentence to specify the content of a mental state (viz., by embedding the sentence to a verb of propositional attitude), the best we can do— in principle, all we can do—is avail ourselves of the content of the sentence qua anchored; for it’s only anchored sentences that have content. (1987, 50) Since (1a) and (1b) cannot be saved by appealing to narrow content, they are unsalvageable—as are (1.5a) and (1.5b). The proper interpretation of the thesis that propositional attitudes are relations to representations remains that given in (2). Although narrow content is not explicitly mentioned in (2), there is no problem accommodating it. If we assume that propositions result from applying narrow contents to Fodorian contexts, and that the propositional content of a mental representation M of an individual i in a context C with i as agent is the result of applying the narrow content of M to C, then (2b) will be true iff (2c) is true.28 (2c) For all propositions p, individuals i (and times t), i believes p (at t) iff there is (at t) a mental representation M, with narrow content N, and context C, with i as agent (and t as time), such that i bears a certain relation R to M (at t), and N(C) = p. The end result is not an alternative to the analysis given in (2b), but an elaboration of it. I emphasize this because it has been suggested to me29 that appeal to narrow content provides “an obvious way” of avoiding both (2) and my critique of Fodor. In particular, (1c) has been suggested as “the obvious alternative” to my (2b). (1c) For all narrow contents N, there is a mental representation M such that for all individuals i, and contexts C, i believes the 28 These principles could, of course, also be relativized to possible worlds. In the case of (2c) the result would be: For all propositions p, individuals i (times t and worlds w), i believes p (at t, w) iff there is (at t, w) a mental representation M, with narrow content N, and a context C, with i as agent (t as time and w as world), such that i bears a certain relation R to M (at t, w), and N(C) = p. 29 By an anonymous referee for this volume.

Belief and Mental Representation • 107

proposition p that results from applying N to C iff i bears R to M, and N is the narrow content of M, and i is agent of C. This proposal is hopeless. Since it fails to give necessary and sufficient conditions for believing p it won’t do as an analysis of belief—for reasons analogous to those given in connection with (1.5a). Even worse, (1c) is plainly false. For example, let C be a context with Bill as agent, let N be such that N(C) = the proposition that Bill lives in New Jersey, and let Mary be distinct from Bill. Since Mary is not the agent of C, the righthand side of the biconditional “Mary believes the proposition that Bill lives in New Jersey iff Mary bears R to M, and N is the narrow content of M, and Mary is agent of C,” obtainable from (1c), is false, thereby requiring the falsity of the left-hand side to preserve the truth of the whole. Thus, (1c) has the absurd consequence that since Mary is not identical with Bill, she doesn’t believe that Bill lives in New Jersey. Perhaps the idea behind (1c) can be sympathetically recast as (1d). (1d) For all narrow contents N, there is a mental representation M with N as narrow content such that for every individual i and context C with i as agent, i believes the proposition p that results from applying N to C iff i bears R to M. Although this avoids the absurdity of (1c), it is still plainly inadequate. I have already indicated that different mental representations can be expected to have different Fodorian narrow contents—which means that, in general, an individual narrow content will be associated with just one mental representation. Using this assumption plus a fact about propositional attitudes, we can derive a contradiction from (1d). I argued in the text (in the discussion of examples (9)–(13)) that sometimes an agent may have different mental representations that express the same proposition. Such an agent may bear R to one of these representations while not bearing R to others. For example, i might bear R to M1, with narrow content N1, but not bear R to M2, with narrow content N2 (distinct from N1), in a context C with i as agent, where N1(C) = N2(C) = p. It now follows from (1d) that if M1 is the only mental representation with narrow content N1, and M2 is the only representation with narrow content N2, then i both believes p and does not believe p. Since this is unacceptable, someone who adopts Fodor’s notion of narrow content ought to reject (1d) as false.30 As a result, the introduction of narrow content does not affect the analysis of belief given in (2). 30 Of course, someone with views different from Fodor’s might try to develop a different notion of narrow content, in which distinct mental representations are allowed to share the same narrow content. However, even then (1d) would not be an acceptable analysis of belief—as indicated by the way in which it would fail to provide necessary and sufficient

108 • Essay Three

However, it does raise a question about Fodor’s conception of the relationship between scientific psychology and commonsense belief/desire explanations of behavior. In chapter 1 of Psychosemantics, he claims that the proper way to do scientific psychology is one that will “vindicate” commonsense generalizations like (14), which achieve their explanatory effect by generalizing over propositional contents. However, in chapter 2, he takes Twin-Earth cases to show that the notion of content required by scientific psychology is narrow, rather than propositional. What, then, becomes of the “vindication” of common sense? Clearly, generalizations framed in terms of propositional contents will not appear in Fodor’s imagined science. Nor, since narrow contents are neither believed nor desired, will generalizations that simply substitute narrow for propositional contents in principles like (14). At a minimum, if we find new objects to replace propositions, we will also need new relations to replace (commonsense) believing and desiring, which are relations to propositions. Perhaps we should call these new relations ‘narrowly-believing’ and ‘narrowly-desiring.’ But now one must ask whether any such new theoretical inventions are needed. Having framed narrow content so as to match up one to one with mental representations, Fodor has no need to burden a properly austere scientific psychology with both. As far as the imagined explanation of behavior is concerned, one might as well omit appeals to narrow content altogether, and rely entirely on computational relations to representations. Nothing would be lost, since representations and narrow contents have been designed to mirror one another anyway. Suppose that some such view of scientific psychology were shown to be correct. We would then have a serious science of the mental which gave true explanations of behavior in terms of the mental syntax of agents. It is worth noting that the truth of such explanations needn’t conflict with the truth of typical content-involving commonsense explanations of behavior. Thus, such a science needn’t be thought of as refuting particular commonsense belief/desire explanations of behavior. However, vindication is more than the absence of refutation, and one might conditions for believing p. Suppose, in the case of the above example, that N1 and N2 were identical, even though M1 and M2 were distinct. Although no contradiction would then be forthcoming, (1d) would make no prediction about whether i believed p or did not believe p. In general, where N is any narrow content and C is a context with i as agent and N(C) = p, the most that could be said would be that a necessary condition for i to believe p was for i to bear R to at least one mental representation with narrow content N, and a sufficient condition for i to believe p was for i to bear R to all mental representations (of others as well as of i himself) with narrow content N. Since this is not enough for an analysis of belief, (1d) would still pose no threat to the analysis of belief given in (2).

Belief and Mental Representation • 109

wonder whether such a science had provided any vindication of our commonsense explanations at all. This issue, I think, depends on the relative priority of mental representations and narrow contents. If one thinks of mental representations as being somehow given in advance of any considerations of their narrow contents, then the invocation of such contents standing in a one-to-one relation to representations will naturally seem to be a theoretically superfluous excrescence. However, if mental representations are themselves individuated (in large part) by appeal to narrow contents—i.e., by appeal to prior commonsense judgments about the propositions believed by agents in different contexts31—then a representation-based psychology of the sort imagined by Fodor might well depend on, and even vindicate, our commonsense conception of propositional attitudes. Although this would be a welcome result, it is not my intention to endorse it, or any other detailed conception, as a prescription for how scientific psychology ought to proceed. While I expect our commonsense conception of the attitudes to play a significant role in the development of scientific psychology, I cannot predict just what that role will be. This uncertainty does not infect my attachment to the attitudes themselves. In my opinion, our commonsense conception of propositional attitudes has an authority that is largely independent of its ultimate role in future science; and I am confident that we will find room in the world for beliefs and desires, no matter what surprises science may hold in store. My concern has been with how to explicate our commonsense conception. I agree with Fodor that mental representations are important for this. However, I have insisted that such an explication should be understood along the lines of (2) rather than (1). It has been the burden of this appendix to show that this result stands, whether or not Fodorian narrow content is incorporated into scientific psychology.

References Carnap, Rudolf. 1956. Meaning and Necessity. 2nd ed. Chicago: University of Chicago Press. Church, Alonzo. 1954. “Intensional Isomorphism and the Identity of Belief.” Philosophical Studies 5:65–73. Reprinted in Propositions and Attitudes, ed. Nathan Salmon and Scott Soames, 159–68 (New York: Oxford University Press, 1988). Fodor, Jerry A. 1975. The Language of Thought. New York: Crowell. 31

As illustrated in the discussion of the examples in note 27.

110 • Essay Three ———. 1981. “Propositional Attitudes.” In RePresentations: Philosophical Essays on the Foundations of Cognitive Science, 177–203. Cambridge: MIT Press. First published in The Monist 61 (1978). ———. 1987. Psychosemantics: The Problem of Meaning in the Philosophy of Mind. Cambridge: MIT Press. Kaplan, David. 1989. “Demonstratives: An Essay on the Semantics, Logic, Metaphysics, and Epistemology of Demonstratives and Other Indexicals.” In Themes from Kaplan, ed. Joseph Almog, John Perry, and Howard Wettstein with the assistance of Ingrid Deiwiks and Edward N. Zalta, 481–563. New York: Oxford University Press. Kripke, Saul A. 1979. “A Puzzle about Belief.” In Meaning and Use: Papers Presented at the Second Jerusalem Philosophical Encounter, April 1976, ed. Avishai Margalit, 239–83. Dordrecht: Reidel. Reprinted in Propositions and Attitudes, ed. Nathan Salmon and Scott Soames, 102–48 (New York: Oxford University Press, 1988). Perry, John. 1979. “The Problem of the Essential Indexical.” Noûs 13:3–21. Reprinted in Propositions and Attitudes, ed. Nathan Salmon and Scott Soames, 83–101 (New York: Oxford University Press, 1988). Putnam, Hilary. 1975. “The Meaning of ‘Meaning’.” In Language, Mind, and Knowledge, ed. Keith Gunderson, 358–98. Minneapolis: University of Minnesota Press. Salmon, Nathan. 1986. Frege’s Puzzle. MIT Press. ———. 1989. “How to Become a Millian Heir.” Noûs 23:211–20. ———. 1990. “A Millian Heir Rejects the Wages of Sinn.” In Propositional Attitudes: The Role of Content in Logic, Language, and Mind, ed. C. Anthony Anderson and Joseph Owens, 215–47. Stanford, Calif.: Center for the Study of Language and Information. Salmon, Nathan, and Scott Soames, eds. 1988. Propositions and Attitudes. New York: Oxford University Press. Soames, Scott. 1985. “Lost Innocence.” Linguistics and Philosophy 8:59–71. ———. “Substitutivity.” 1987. In On Being and Saying: Essays for Richard Cartwright, ed. J. J. Thomson, 99–132. Cambridge: MIT Press. ———. 1988. “Direct Reference, Propositional Attitudes, and Semantic Content.” In Propositions and Attitudes, ed. Nathan Salmon and Scott Soames, 197–239. New York: Oxford University Press. First published in Philosophical Topics 15:47–87. ———. 1989. “Direct Reference and Propositional Attitudes.” In Themes from Kaplan, ed. Joseph Almog, John Perry, and Howard Wettstein with the assistance of Ingrid Deiwiks and Edward N. Zalta, 393–419. Oxford: Oxford University Press. Sterelny, Kim. 1989. “Fodor’s Nativism.” Philosophical Studies 55:119–41.

ESSAY FOUR

Attitudes and Anaphora

In this essay I will investigate how pronouns anaphoric on singular term antecedents are understood, and what a semantic theory should say about them. I propose to examine these questions by appealing to propositional attitude ascriptions containing such pronoun/antecedent pairs. In so doing I hope not only to advance our understanding of a particular linguistic construction, but also to illustrate a productive yet underappreciated methodology. If one is interested in characterizing what a certain sentence means, one can scarcely do better than attend to the assertions that utterances of the sentence are standardly used to make, and the beliefs they can reliably be taken to express. The reason for this is that when we ascribe such assertions and beliefs to a speaker on the basis of what he says, the truth of our ascription generally requires us to be more faithful, in characterizing what he says or believes, to the meaning of the sentence he utters than we would have to be if we were merely interested in providing the truth conditions of his remark. Consequently, judgments about the truth conditions of attitude ascriptions, N says that S or N believes that S, can often be used to discriminate between different semantic proposals regarding the interpretation of S, even when the proposals are extensionally and intensionally equivalent.1 They can also be used to raise difficult issues and problems for familiar semantic analyses that might otherwise seem unproblematic. My aim will be to illustrate both of these points regarding semantic analyses of pronouns anaphoric on singular term antecedents?2 The problems I am interested in arise from a theoretical background that includes the following assumptions. 1 Two semantic interpretations of a sentence (expression) are intensionally equivalent iff they assign the sentence (expression) the same function from truth-supporting circumstances—e.g., possible worlds—to truth-values (extensions). 2 The material in the first section of this essay recapitulates, in a shorter space, the main conclusions drawn in Soames (1989–90). In the following sections I discuss problems that arise when one considers questions about the proper scope of the mechanism for interpreting anaphoric pronouns presented in that article, as well as questions about whether that mechanism needs to be supplemented with, or even supplanted by, other mechanisms for producing analogous interpretations.

112 • Essay Four

Assumption 1: Sentences express propositions (contents) relative to contexts of utterance. A semantic theory specifies meanings of sentences as functions from contexts to propositions. Assumption 2 (Compositionality): Let S and S′ be extensional sentences that are grammatically constructed in the same way from corresponding constituents el, e2, . . . ,en and e′l, e′2, . . . ,e′n respectively. If the semantic content of ei in context C is identical with the semantic content of e′i in context C′ (for each i), then the semantic content of S in C is identical with the semantic content of S′ in C′— i.e., they express the same proposition. Assumption 3: A sincere, reflective, competent speaker who assertively utters S in a context C typically asserts, often among other things, the proposition expressed by S relative to C. Assumption 4: An individual i satisfies x says (asserts) that S relative to a context C iff i bears the relation of asserting to the proposition expressed by S relative to C. Assumption 5: An individual i satisfies x v’s that S (where v = ‘believes’, ‘knows’, ‘proves’, ‘expects’, etc.) relative to a context C iff i stands in a certain relation R to the proposition expressed by S relative to C. Assumption 6 says that anaphoric pronouns with quantified antecedents function as bound variables, while Assumption 7 spells out what this function amounts to. Assumption 6: Anaphoric occurrences of pronouns with c-commanding quantified antecedents function as bound variables.3 3 The notion of c-command is commonly used by linguists to characterize the scope of quantifiers in natural language. Although details of the definition vary slightly from theorist to theorist, the basic idea is that a constituent A of a sentence S c-commands a nonoverlapping constituent B of S iff the first branching node that dominates A in a constituent structure representation of S also dominates B. The scope of a quantifier in an English sentence, and hence the domain within which it can bind a pronoun/variable, is standardly said to consist in the elements it c-commands. For example, the quantifier ‘every man’ c-commands, and hence binds, the associated pronouns in ‘Every man loves his mother’ and ‘Mary told every man that he was unreliable’; but it does not c-command, and so cannot bind, the pronouns in ‘If every man dates Mary, then he likes Mary’ or ‘The woman who dated every man likes him’. It should be understood that the precise syntactic characterization of the domain of quantification in English is not at issue here. The points that will be at issue are: (i) the interpretation of occurrences of pronouns that clearly and uncontroversially function as variables bound by their quantified antecedents (Assumptions 6 and 7), (ii) the treatment of parallel cases in which the c-commanding antecedent is a singular term, and (iii) the analysis of cases analogous to (ii) except for the fact that the antecedent does not c-command the anaphoric pronoun, and so occupies a position from which variable binding by ordinary quantifiers is not normally possible.

Attitudes and Anaphora • 113

Assumption 7: The truth conditions of a sentence Quantifier v ( . . . v . . . ) are defined in terms of the truth conditions of the formula ( . . . v . . . ) relative to assignments of objects to variables. The semantic content of a variable v relative to an assignment f is the referent of v relative to f, which is the object that f assigns to v. Thus, the contribution made by v to the truth conditions of a formula relative to f is simply the object that f assigns as value of v. These assumptions have important consequences for the analysis of examples like those in (1). (1) a. At least one planet is such that Ralph says, and believes, that it is visible in the morning. b. Each woman detested the man who said that she was silly. According to our assumptions, (la) is equivalent to (2), which is true iff Ralph asserts and believes the proposition expressed by (3) relative to an assignment of a planet p to ‘x’.4 (2) [∃x: Planet x](Ralph says, and believes, that x is visible in the morning). (3) x is visible in the morning. It is a job of the semantics to tell us what proposition this is. According to Assumption 7 the semantic content of ‘x’ relative to this assignment is its referent p. Since the proposition expressed by a sentence is determined by its grammatical structure plus the semantic contents of its parts, the proposition expressed by (3) relative to this assignment is a singular proposition about p to the effect that it is visible in the morning. I use the set theoretic construction (4) to represent it. (4) . Thus (la) is true iff there is at least one planet p such that Ralph asserted and believed the singular proposition (4) about p. Next we must ask what it is to assert and believe a singular proposition. To answer this question we need to note what is required in order 4 Although the assumptions speak explicitly about sentences, I intend them to be understood as also applying to “open sentences—i.e., formulas containing free variables, or the natural language equivalents of such. For example, Assumption 1 calls for the assignment of propositions to formulas relative to assignments of values to pronoun/variables, Assumption 2 guarantees that substitution which does not change the contents of subsentential constituents relative to appropriate contexts and assignments does not change the content of the entire (open) sentence with respect to those contexts and assignments, and Assumption 4 implies that an ascription x says that S is true relative to an assignment f, and context c iff the referent of x with respect to f bears the relation of asserting to the proposition expressed by the (open) sentence S relative to f and c.

114 • Essay Four

for examples like (la–b) to be true. In the case of (la) it is sufficient for Ralph to sincerely and reflectively utter one of the sentences in (5), fully comprehending and accepting what he said. (5) a. That [pointing at p] is visible in the morning. b. Venus (Phosphorus) is visible in the morning. Similarly, (lb) will be true if for each woman w, there was exactly one man m who assertively and comprehendingly uttered a sentence of the sort indicated in (6), and w detested m. (6) a. She [pointing at w] is silly. b. You [addressing w] are silly. c. N (where N names w) is silly. This account is generalized in Assumption 8. Assumption 8: In order to assert and believe a singular proposition about an object o it is sufficient to sincerely, comprehendingly, and assertively utter a corresponding sentence containing a name or demonstrative that refers to o. More generally, in order to assert and believe a singular proposition about an n-tuple of objects o1, . . . ,on it is sufficient to sincerely, comprehendingly and assertively utter a corresponding sentence containing names or demonstratives referring to the oi’s.5 With these assumptions in place I want to review some familiar interpretations of anaphoric pronouns of the sort that occur in examples (7) and (8). (7) a. b. (8) a. b.

John loves his mother. Mary says, and believes, that John loves his mother. Mary told John that he wasn’t John. John fooled Mary into thinking that he wasn’t John.

In each of these cases an anaphoric pronoun occurs with a singular term antecedent. Typical characterizations of anaphora tell us that an anaphoric element is one that inherits its interpretation from its antecedent. Thus, theorists have typically asked what it is that these pronouns inherit. The answer has standardly been that they inherit the reference of their antecedents. Thus, it has often been assumed that the 5 Note, Assumption 8 allows but does not require names and indexicals to be directly referential—i.e., for their semantic contents to be their referents. By the same token, it allows for the possibility that sentences containing names and indexicals might express singular propositions analogous to (4), but it does not require that they do. See Soames (1989–90, 195–97) for a discussion of these points.

Attitudes and Anaphora • 115

interpretation of anaphoric pronouns like these is given by a simple rule which states that the anaphoric pronoun refers to the same thing as its antecedent. Rule R: An anaphoric occurrence of a pronoun with a singular term antecedent T refers to the same thing as T. It is important to see that, from the point of view of our assumptions, Rule R is incomplete. In order to interpret a sentence containing an anaphoric pronoun we need to specify the proposition it expresses, and in order to do this we need to know the semantic content of the pronoun, namely that which it contributes to the proposition. There are two main possibilities to consider. The first possibility is that reference is all there is to the interpretation of these anaphoric pronouns. The Anaphoric Direct Reference Hypothesis: The semantic content of an anaphoric occurrence of a pronoun with a singular term antecedent is its referent. This hypothesis together with Rule R constitutes one possible alternative for interpreting the pronouns we are interested in. The second possibility is given by the pair consisting of Rule R, and the anti-direct-reference analysis of names. Rule Rs: An anaphoric occurrence of a pronoun with a singular term antecedent T has the same semantic content as T. The Anti-Direct-Reference Analysis of Names: Proper names express senses in addition to referring to objects. These senses allow coreferential names to differ in semantic content, and thus make it possible for substitution of such names to affect what proposition is expressed. It is important to see that neither of these alternatives accounts for the full range of cases that needs to be covered. First consider the treatment of (7b) by the direct reference account. When combined with our background assumptions, this account tells us that (7b) is true iff Mary asserts and believes the singular proposition about John expressed by the formula (7c) John loves x’s mother relative to an assignment of the man John to the variable ‘x’. Assumption 8 guarantees that it is sufficient for Mary to assert and believe this proposition that she sincerely and comprehendingly utter (7d). (7d) John loves your mother (addressing John). There is no requirement in these predicted truth conditions that Mary realize that the person she is addressing is the person whose love she is

116 • Essay Four

reporting; in fact, she may be convinced that the person she is addressing is not John. Thus there is no requirement that Mary say or believe that John is an own-mother-lover. This, I take it, conflicts with our ordinary understanding of at least one anaphoric interpretation of (7b). That understanding is one in which Mary is characterized as asserting and believing a proposition that attributes to John the property of loving one’s own mother. If this understanding is correct, then (7a) has an interpretation that is not captured by the conjunction of Rule R and the Anaphoric Direct Reference Hypothesis. To capture this interpretation one might opt for Rule Rs plus the Anti-Direct-Reference Analysis of Names. One would then get the result that (7b) is true iff Mary asserts and believes the proposition expressed by (7e). (7e) John loves John’s mother. On this account it would not be sufficient for Mary to assert and believe the proposition expressed by (7e) that she sincerely and comprehendingly utter (7d) in a context in which ‘you’ demonstratively refers to John. Rather the proponent of this account may maintain that in order to assert and believe (7e) one must recognize the purported lover and the person whose mother is said to be loved as one and the same. However, even if this account were accepted for (7b), it fails for (8a) and (8b). We know, for example, that (8b) will be true if John’s actions lead Mary sincerely and comprehendingly to accept sentences (8c) or (8d) in a context in which the demonstratives refer to John, despite Mary’s unawareness that they are coreferential with the name ‘John’. (8) c. He [pointing at John] isn’t John. d. You [addressing John] aren’t John. The combination of Rule Rs plus the Anti-Direct-Reference Analysis of Names cannot account for this. On that view, the truth of (8b) requires Mary to have accepted proposition (8e). (8e) John isn’t John. Moreover, it is crucial to the view that just as sincerely and comprehendingly uttering (7d) does not count as asserting or believing proposition (7e), so sincerely and comprehendingly uttering (8c, d) does not count as asserting or believing proposition (8e). In short, the very features posited by the analysis to account for the sentences in (7) make it incapable of handling those in (8). The difference between the sentences in (7) and those in (8) is that in (8) there is an intervening propositional attitude verb that takes the pronoun, but not the antecedent, in its scope. It may be tempting to think that we

Attitudes and Anaphora • 117

could rely on a directly referential analysis of the pronouns in cases like (8), where there is an intervening propositional attitude verb, while relying on the anti-direct-reference analysis in cases like (7), where there is no such intervention. However, as (9) shows, it is obvious that this won’t work. (9) The children thought that Mary told John that he wasn’t John. Just as (7b) has a natural anaphoric interpretation in which Mary is characterized as believing a proposition that attributes to John the property of being one who loves one’s own mother, so (9) has a natural interpretation in which the children are characterized as believing a proposition that attributes to John the property of being one who was told by Mary that one wasn’t John. Neither the analysis which treats the pronoun as a directly referential term that inherits the reference of its antecedent, ‘John’, nor the analysis which treats it as a non–directly referential term that inherits the content of the name ‘John’ is capable of capturing this interpretation. What we need is a unified account that handles each of the examples in (7), (8), and (9). One way of providing such an account would be to follow Richard Montague in treating demonstratives and grammatically proper names as quantifiers rather than singular terms. Anaphoric pronouns c-commanded by such antecedents would not inherit the sense or reference of their antecedents, but rather would function as bound variables. On this account, (7a) may be represented as (10a) and understood on a par with (10b). (10) a. John loves his mother. [John x](x loves x’s mother). b. Every man loves his mother. [Every x: Man x](x loves x’s mother). Just as (10b) expresses a proposition that attributes the property of loving one’s own mother to every man, so (10a) expresses a proposition that attributes this property to John. Both of these propositions are complexes consisting of the property of loving one’s own mother plus a higher-order property expressed by the quantifier. In the case of (10b), the higher-order property is one that is true of precisely those properties that are true of every man; in the case of (10a) it is one that is true of precisely those properties that are true of John. Thus, the proposition expressed by (10a) is (11). (11) . Since one cannot believe this proposition without realizing that the person whose love is in question is the person whose mother is supposed

118 • Essay Four

to be loved, the analysis of proper names as quantifiers accounts for (7b).6 It also handles (8a), which is represented as (12a), and understood on a par with (12b). (12) a. Mary told John that he wasn’t John. [John x](Mary told x that x wasn’t John). b. Mary told some man that he wasn’t John. [Some x: Man x](Mary told x that x wasn’t John). On this analysis, (12a) is true iff the formula (13) Mary told x that x wasn’t John is true relative to an assignment of the man John to x. According to our background assumptions, this will hold if Mary sincerely and comprehendingly addressed John by assertively uttering (8d) [‘You aren’t John’]. In short, the combination of these assumptions plus the analysis of names and demonstratives as quantifiers handles all the cases we have so far considered. There is another way to get these desirable results that does not involve giving up the intuitively compelling view that grammatically proper names and demonstratives are genuine singular terms. According to it, anaphoric pronouns with c-commanding singular term antecedents are variables bound by an abstraction operator introduced by the anaphoric relation itself. On this analysis, (7a) can be represented as (14a), which we may take as expressing the singular proposition (14b), and (8a) can be represented (15).7 (14) a. λx (x loves x’s mother) John. b. . (15) λx (Mary told x that x wasn’t John) John. 6 To get this result we must, of course, depart from Montague’s characterization of propositions as sets of possible worlds. On that characterization there is no difference between the proposition expressed by ‘[John x](x loves x’s mother)’ and the proposition expressed by ‘x loves x’s mother’ relative to an assignment of the man John to the variable ‘x’—since the two are intensionally equivalent. In place of that characterization I will assume that propositions are structured complexes built up out of the semantic contents of the syntactic constituents of the sentences that express them. 7 In Soames (1989–90, 204) I described this analysis as one in which anaphoric pronouns with singular term antecedents are abstraction operators rather than singular terms. This characterization is effectively criticized by Nathan Salmon (1992). I agree with Salmon that a better way of conceptualizing the proposal is to view the pronouns as variables bound by an abstraction operator which is not associated with any constituent of the sentence that is overtly present in surface structure. It was just this bound-variable conceptualization that I was referring to in the last paragraph of Soames (1989–90, 204–5 n. 11).

Attitudes and Anaphora • 119

In this way, the abstraction analysis also accounts for all the problematic cases so far considered.8 Another case that should be mentioned is (16). (16) At least one criminal is such that Mary says, and believes, that he loves his mother. Just as (7b) has a natural anaphoric interpretation in which Mary is characterized as saying and believing that John is an own-mother-lover, so (16) has a natural anaphoric interpretation in which Mary is characterized as saying and believing of some criminal that he is an own-mother-lover. One cannot get this interpretation by treating both anaphoric pronouns in (16) as variables bound by the same quantified antecedent, and analyzing the sentence as equivalent to (17a). (17a) [∃x: Criminal x](Mary says, and believes, that x loves x’s mother). According to Assumptions 2, 3, 4, 5, and 7, (17a) is equivalent to (17b), (17b) [∃x: Criminal x] [∃y: y = x] (Mary says, and believes, that x loves y’s mother) which is true iff Mary asserts and believes the singular proposition expressed by ‘x loves y’s mother’ relative to an assignment of the same criminal to x and y. According to Assumption 8, it is sufficient for Mary to assert and believe this proposition that she sincerely and comprehendingly accept, and assertively utter, (17c) in a context in which T1 and T2 are names or indexicals that denote the same individual. (17c) T1 loves T2’s mother. It is not required that T1 and T2 be the same term, or that Mary realize that they are codesignative. Thus, (17a) does not require Mary to characterize any individual as loving his own mother, and so fails to express the desired interpretation of (16). To get the desired reflexive interpretation, we need to take the antecedent of the pronoun ‘his’ to be the c-commanding pronoun ‘he’, rather than the quantifier phrase. Predicate abstraction can be used to give the interpretation of the anaphoric relation between the pronouns in the complement clause. Thus, (16) is equivalent to (18a). (18a) At least one criminal is such that Mary says, and believes, that λx (x loves x’s mother) he. 8 See Soames (1989–90, 205–8) for a discussion of how the following example might be used to construct an argument for preferring the predicate abstraction analysis to the quantificational treatment.

120 • Essay Four

The anaphoric occurrence of he may now be taken to be a variable bound by its quantified antecedent. This results in (18b), which has the desired truth conditions. (18b) [∃y: Criminal y] (Mary says, and believes, that λx (x loves x’s mother)y).

A Problem All of this seems promising. However, there is a significant unsolved problem that needs to be addressed—namely, that of determining the scope of the bound variable analysis of anaphoric occurrences of pronouns with (grammatically) singular term antecedents (names, demonstratives, and pronouns). It is not unreasonable to suppose that if such occurrences of pronouns really are bound variables, then the scope of the operators that bind them should be subject to the usual syntactic constraints; and so should be limited to constituents c-commanded by the operators. In light of this, it might be thought that an occurrence of a pronoun with a grammatically singular term antecedent can be interpreted as a variable bound by a quantificational construal of the antecedent, or by a predicate abstraction operator associated with the anaphoric relation itself, only if the antecedent c-commands the pronoun, and so occupies a position from which an ordinary quantifier could bind it. However, this appears not to be so, as is illustrated by (19a). (19a) Ralph says/believes that any man who dates Susan likes her. Here the antecedent, Susan, does not c-command the pronoun her. As a result, if the name is replaced by an ordinary quantifier, the quantifier cannot bind the pronoun. (19b) *Ralph says/believes that any man who dates every woman likes her. In light of this it might seem that we should not be able to bind the pronoun in (19a) either by treating Susan as a quantifier, or by introducing an abstraction operator to interpret the anaphoric relation. But this is puzzling, since it seems that (19a) has the same sort of reflexive interpretation as our earlier examples. In particular, it seems as if we could construct an argument, parallel to those given earlier, to show that (19a) has an anaphoric interpretation whose truth would not be guaranteed by Ralph’s sincere, comprehending assent to (19c), in a situation in which he did not realize that his use of the demonstrative her was coreferential with the name Susan.

Attitudes and Anaphora • 121

(19c) Any man who dates Susan likes her [pointing to Susan]. Moreover, if (19a) has a reflexive interpretation, then surely (19d) does too. (19d) At least one woman is such that Ralph says/believes that any man who dates her likes her. There are three possible ways of dealing with examples like these that should be considered. The first possibility is to maintain that, appearances to the contrary, they do not have reflexive interpretations. In the case of (19a), this would mean arguing that the attitude ascription does not have an interpretation whose truth requires Ralph to have ascribed to Susan the property of being one who is liked by any man who dates one, or even to have recognized that the person said to be liked is the same as the person said to be dated. The problem with this is that it flies in the face of our ordinary, pretheoretic intuitions that sentences like (19a) do have such interpretations. It is, of course, possible that such intuitions might be confused and mistaken. However, if they are, what is the status of corresponding intuitions about reflexive interpretations of examples like ‘John loves his mother’ and ‘John injured himself ’, in which the antecedent c-commands the pronoun? Are these intuitions also incorrect, or is there some principled distinction between the two classes of cases? My own view is that we should be hesitant about rejecting any of these intuitions. Although they are certainly not infallible, they do have evidential weight and should be respected, unless we determine that there is no theoretically plausible way of doing so. The second possible way of dealing with examples like (19a, d) is to extend the bound variable analysis to cases in which the antecedent does not c-command the pronoun. On this analysis, (19a) could be represented as (19a′), and (19d) could be represented as (19d′). (19) a′. Ralph says/believes that (λx[any man who dates x likes x] Susan). d′. [∃y: Woman y] (Ralph says/believes that (λx[any man who dates x likes x] y)). This solution is technically possible, and could very well turn out to be correct. Certainly, nothing I have to say will refute it. However, it should be recognized that adopting it would leave us with an unanswered question. Why can a pronoun anaphoric on a non-c-commanding antecedent function as a bound variable if the antecedent is a name, demonstrative, or another pronoun/variable, but not function as a bound variable if the antecedent is an ordinary quantifier? The third possible way of dealing with examples like those in (19) avoids this question by (i) limiting the bound variable interpretation of

122 • Essay Four

anaphora to cases in which the antecedent c-commands the pronoun, and (ii) positing other interpretive mechanisms to produce reflexive-type interpretations in the remaining cases. What makes this idea especially attractive is the possibility that these extra interpretive mechanisms may be needed independently for other reasons. If, in addition, they can be shown to produce the right results for the cases that are problematic for the bound variable analysis, then the resulting theory would cover the needed range of cases in a maximally desirable way. In the remainder of this essay, I will try to determine whether this strategy is workable.

Unbound Pronouns: One Strategy for Solving the Problem With this in mind, let us compare the sentences in (19) to those in (20). (20) a. Ralph says that any man who dates an intelligent woman likes her. b. Ralph says that any man who dates the woman over there likes her. I will assume here that definite and indefinite descriptions are quantifiers whose scopes are confined to constituents they c-command. On this assumption, the examples in (20) illustrate the fact that quantifiers can sometimes be antecedents of pronouns they do not c-command, and therefore do not bind. The unbound analysis of these pronouns, derived from Gareth Evans, Martin Davies, and Stephen Neale, treats them as disguised descriptions.9 On this analysis (20a, b) express the same propositions as (21a, b). (21) a. Ralph says that any man who dates an intelligent woman likes the intelligent woman (or all the intelligent women) he dates. b. Ralph says that any man who dates the woman over there likes the woman over there. The point to notice is that on this analysis we can assign (20a, b) a sort of pseudo-reflexive interpretation without having to extend the scope of the bound variable analysis. Suppose now that the account of unbound anaphora used in (20) can be applied to (19a), so as to yield: (22a) Ralph says/believes that any man who dates Susan likes Susan (or the x: x = Susan). 9

Evans (1977); Davies (1981); Neale (1990).

Attitudes and Anaphora • 123

In the presence of our other assumptions, this will give us what we need only if names are not directly referential, i.e., only if the semantic content of a name is not exhausted by its referent. For if Susan is directly referential, then (22a) will be true iff (22b) is true relative to an assignment of the woman, Susan, to the variables ‘x’ and ‘y’. (22b) Ralph says/believes that any man who dates x likes y. Moreover, it is sufficient for the truth of this that Ralph sincerely, comprehendingly, and assertively utter a sentence of the form (22c) Any man who dates t like t′, where the terms ‘t’ and ‘t′’ are names or indexicals that both refer to Susan. Since it is not required that Ralph be aware that they refer to the same person, we cannot get the pseudo-reflexive truth conditions we are after in this way. Perhaps this is simply so much the worse for the view that proper names are directly referential. Thus, it might be proposed that names contribute more than their referents to the propositions expressed by sentences containing them. Moreover, it might be maintained that the proposition expressed by a sentence of the form (23a) . . . α . . . α . . . is such that a speaker who asserts or believes it by virtue of accepting, or assertively uttering, a sentence of the form (23b) . . . γ . . . δ . . . must realize that γ and δ have the same content. Given this much, one could then postulate two different anaphoric processes relating pronouns with singular term antecedents. The first, involving variable binding (in the form, let us say, of predicate abstraction), is responsible for reflexive interpretations of examples like (7)–(9), and is limited to cases in which the antecedent c-commands the pronoun. The second, more general, process does not require the antecedent to c-command the pronoun, and is operative in examples like (19a) and (20). A pronoun is interpreted by this process as if it were a repetition of its singular term antecedent. The idea, then, would be to account for competent speakers’ intuitions of reflexive, or pseudo-reflexive, interpretations across a wide range of linguistic environments by positing two types of semantic processes that give rise to such interpretations. The generality of this account is illustrated by (24). (24) a. Mary thinks that Hesperus is a star, but it isn’t. b. Mary thinks that Hesperus is a star, but Hesperus isn’t a star. c. Mary thinks that Phosphorus is a star, but it isn’t.

124 • Essay Four

First consider (24a). By treating ‘Hesperus’ as nondirectly referential, and it as a repetition of ‘Hesperus’, the account we are considering assigns it an interpretation that is both pseudo-reflexive and distinct from the interpretation of (24c). A similar point is illustrated by the discourse, (25), in which anaphoric occurrences of pronouns are not in the same sentence as their antecedent. (25) Discourse: Bill told us the following story: Mary entered the building and started up the stairs. Sam heard her footsteps and looked up. Before she reached the top of the stairs he was up and moving toward the back door. When she came in, the office was empty. The aim of the account we are considering is to give this discourse a semantic interpretation from which it follows that Bill said that one and the same person entered the building, started up the stairs, was heard by Sam before reaching the top, and entered an empty office. Since separate sentences are involved it doesn’t seem plausible to invoke variable binding to get this result. We can, of course, view the occurrences of ‘she’ as repetitions of the antecedent, ‘Mary’. But this will produce the kind of pseudo-reflexive interpretation we want only if names are not directly referential, which is, of course, just what the alternative we are considering maintains. There are, however, two important problems with this rosy scenario. The first is that of coming up with an acceptable theory of the semantic contents of names and indexicals as something other than their referents. I am not sanguine about this. The second problem is more specific, and is illustrated by (19d), which, as I have said, surely has a reflexive, or pseudo-reflexive interpretation, if (19a) does. (19d) At least one woman is such that Ralph says/believes that any man who dates her likes her. In this example, neither occurrence of the pronoun ‘her’ c-commands the other. Thus, if the abstraction analysis is restricted to cases involving c-command, then (19d′) (19d′) [∃y: Woman y] (Ralph says/believes that (λx[any man who dates x likes x] y)) is not a possible representation of (19d), and there is little choice but to take the quantifier phrase to be the antecedent of both pronouns, as is indicated in (19e): (19e) [∃x: Woman x] (Ralph says/believes that any man who dates x likes x).

Attitudes and Anaphora • 125

But then, given our assumptions, we don’t get a reflexive interpretation— for the same reason that (17a) did not provide a reflexive interpretation of (16).

Annotated Propositions: Another Strategy The lesson to be learned is that our problem with delimiting the scope of abstraction anaphora cannot be solved by any account that contents itself with providing nondirectly referential treatments of names and indexicals. In addition, quantification must be subjected to a similar treatment, which makes the task all the more difficult. In order to appreciate this difficulty, it is necessary to avoid a certain temptation. The temptation is to suppose (i) that names have senses capturing their cognitive significance for speakers, (ii) that substitution of coreferential names with different senses may fail to preserve the cognitive significance of a sentence, (iii) that the truth of an attitude ascription x says/believes that S requires the cognitive significance of S to match the cognitive significance of the agent’s attitude, and (iv) that a quantified sentence ∃xFx is true iff there is some object o and name a of o such that Fα/x (the result of substituting a for free occurrences of x) is true. What makes this tempting in the present case is that it would lead one to think that (19e) is true just in case some instance of it of the sort illustrated by (22a) is true. (22a) Ralph says/believes that any man who dates Susan likes Susan. It might then seem that if we had a pseudo-reflexive account of examples like (22a), then the quantificational cases would present no further problem. What makes this something to be avoided is that (i)–(iv) do not account for the nature of quantification into attitude ascriptions, as is shown by examples like (26) and (27). (26) There is at least one planet which has the characteristic that Ralph said/believed, in the morning, that it was visible only in the morning and Ralph said/believed, in the evening, that it was visible only in the evening. (27) There is at least one man who is such that Mary believes that he is a saint and Margaret believes that he is a sinner. It is a fact about English that (26) is true in a situation in which, in the morning Ralph looks up and says ‘Phosphorus is visible only in the morning’ (or ‘That [pointing at Venus] is visible only in the morning’) and in the evening looks up and says ‘Hesperus is visible only in the evening’ (or

126 • Essay Four

‘That [pointing at Venus] is visible only in the evening’)—even though he doesn’t realize that he has talked about the same thing on both occasions. In particular, the truth of (26) does not require the existence of some name α of Venus such that α is visible only in the morning and α is visible only in the evening captures Ralph’s cognitive perspectives, morning and evening. Similarly, the truth of (27) is compatible with the possibility that Mary and Margaret may have completely different cognitive perspectives on (and representations of) the same man. Examples like these show that when we quantify into an attitude ascription we abstract away from the cognitive perspective of the agent in a way that falsifies the conjunction of (i)–(iv). Thus, if the goal is to determine whether our problem with delimiting the scope of abstraction anaphora can be solved by a semantic theory that allows reflexive interpretations to arise not only from multiple occurrences of the same name or indexical, but also from multiple occurrences of the same variable, we must look for some treatment of quantification into the attitudes other than the one given by (i)–(iv). Such a theory has recently been developed by Mark Richard.10 Two main aspects of the theory are relevant to us. The first is Richard’s introduction of words into ordinary, run-of-the-mill propositions. For example, consider sentences (28a, b) and the Russellian proposition (28c) corresponding to them. (28) a. Hesperus = Hesperus. b. Hesperus = Phosphorus. c. . For Richard, the proposition expressed by a sentence is an amalgam of the Russellian proposition corresponding to it plus the words of the sentence itself. Call these annotated propositions. The annotated propositions expressed by (28a) and (28b) are (29a) and (29b), respectively. (29) a. . b. . Just as formulas express Russellian propositions relative to assignments of values, so they express annotated propositions relative to such assignments. For example, the annotated propositions expressed by (30a, b) relative to an assignment of the planet Venus to ‘x’ and ‘y’ are (3la, b). (30) a. x = x. b. x = y. 10

Richard (1990).

Attitudes and Anaphora • 127

(31) a. . b. . The key point here is that substitution of coreferential singular terms, even coreferential variables, fails to preserve the proposition expressed. The next aspect of Richard’s system that is crucial to our discussion is the analysis of belief ascriptions given in (32). (32)

 x believes that S is true relative to a context of utterance C and assignment A of values to variables iff the individual denoted by ‘x’ relative to A stands in a certain relation R to the annotated proposition expressed by S relative to C and A.

It is important here to be clear about the relation R, the content of which is informally given in (33). (33) In order to believe the annotated proposition P expressed by the complement clause that S in the context C in which the belief ascription is made, the agent must accept a sentence S′ that expresses (in the agent’s context) an annotated proposition Q which is obtainable from P by application of a function f that (i) maps the words in P onto those in Q (without affecting the nonlinguistic Russellian content of P, Q), and (ii) satisfies contextual constraints in C regarding which words in the ascriber’s context may be used to represent words used by the agent. In short, in order for an ascription x believes that S to be true, the agent must accept a sentence S′ whose words are functionally related to the words of S in an appropriate way, and whose Russellian content is exactly the same as that of S. We need not be concerned here with how, according to Richard, conversational contexts are supposed to constrain acceptable mappings of the ascriber’s words onto the words used by the agent.11 Suffice it to say that in contexts in which no constraints are operative substitution of 11 According to Richard the intentions of the speaker, and to some extent the audience, determine a set of restrictions on acceptable correlation functions, functions which map the ascriber’s words onto those of the agents of attitude ascriptions. A restriction consists of a triple the first element of which is a person u (corresponding to a potential agent of an attitude ascription), the second element of which is a pair consisting of a word w and its Russellian content (e.g., a name used by the speaker together with the referent of that name), and the third element of which is a set s of words with that content (e.g., names whose use by the agent can be represented in the ascriber’s context by w). A correlation function f is said to obey the contextual restrictions relevant to an individual u iff for each contextual restriction whose first element is u, f(w) is a member of s. An individual i satisfies an as

128 • Essay Four

coreferential singular terms is essentially free, thereby capturing standard Russellian intuitions; whereas in contexts in which there are constraints on which words can represent which others substitution is restricted, thereby capturing standard Fregean intuitions. That, anyway, is the idea. What is important to us is not whether this idea is ultimately successful; rather what is important is a feature of the analysis common to all contexts—namely, the fact that the mapping from the ascriber’s words to the agent’s words is a function; defined on expression types. This means that when the complement clause in the ascriber’s sentence contains multiple occurrences of the same expression, even if that expression is a variable, the ascription can be true only if the agent accepts a corresponding sentence containing multiple occurrences of a corresponding expression. For example, in Richard’s system it is not sufficient for the truth of (19e) that Ralph sincerely, comprehendingly, and assertively utter, and accept, a sentence, Any man who dates t likes t′, in which t and t′ are singular terms that refer to the same woman. In order for (19e) to be true, t and t′ must, in addition, be the same term, and be recognized to be such by Ralph. In this way Richard assigns a kind of reflexive interpretation to cription, x believes that S, as uttered in a context C, only if there is a correlation function obeying all the restrictions in C relevant to i that maps S onto a sentence S' which is accepted by i, and which expresses the same Russellian proposition in i’s context as S does in C. The picture then is this: The content of a speaker’s ascription of a propositional attitude to an agent a is determined in part by the speaker’s intention to use certain words to represent words used by a—for example, to use the word ‘Hesperus’ to represent the Babylonian translation of ‘Hesperus’ when ascribing attitudes to a particular Babylonian. Note, however, that the claim that a speaker intends to use one word to represent another when ascribing attitudes to an individual is itself an attitude ascription—one that gives rise to the same kind of puzzles that motivate Richard’s treatment of belief ascriptions. One problem worth noting in this connection is the following: Suppose a speaker says to himself “I intend to use the name n to represent the name m when ascribing attitudes to the F”. Suppose further that a is the F, but the speaker does not know this. Is the speaker’s intention sufficient to establish a restriction of which the individual a is the first element? If not, the speaker’s ascriptions will be assigned essentially Russellian truth conditions, contrary, I believe, to Richard’s motivations. (Think of ascriptions of beliefs to ancient Babylonian astronomers by someone with no de re beliefs about them.) If, on the other hand, the speaker’s intention is sufficient to establish such a restriction, then it would seem all too easy for the speaker’s intentions to establish incompatible restrictions. Suppose, for example, the speaker says to himself not only “I intend to use the name n to represent the name m when ascribing attitudes to the F”, but also “I intend to use the name n to represent the name m* when ascribing attitudes to the G”, where m and m* are different names, and the speaker is unaware of the fact that the F is the G. It would seem that we would then have incompatible restrictions involving the individual a who, in fact, is both the F and the G. Since no correlation function can satisfy these incompatible restrictions, it would follow that no ascription of an attitude to a by the speaker (in the relevant context) could be true—and that doesn’t seem right.

Attitudes and Anaphora • 129

(19e), and thereby to the English sentence (19d), which it can be taken to represent.12 The picture then is this: The content of a speaker’s ascription of a propositional attitude to an agent a is determined in part by the speaker’s intention to use certain words to represent words used by a—for example, to use the word ‘Hesperus’ to represent the Babylonian translation of ‘Hesperus’ when ascribing attitudes to a particular Babylonian. Note, however, that the claim that a speaker intends to use one word to represent another when ascribing attitudes to an individual is itself an attitude ascription—one that gives rise to the same kind of puzzles that motivate Richard’s treatment of belief ascriptions. One problem worth noting in this connection is the following: Suppose a speaker says to himself “I intend to use the name n to represent the name m when ascribing attitudes to the F”. Suppose further that a is the F, but the speaker does not know this. Is the speaker’s intention sufficient to establish a restriction of which the individual a is the first element? If not, the speaker’s ascriptions will be assigned essentially Russellian truth conditions, contrary, I believe, to Richard’s motivations. (Think of ascriptions of beliefs to ancient Babylonian astronomers by someone with no de re beliefs about them.) If, on the other hand, the speaker’s intention is sufficient to establish such a restriction, then it would seem all too easy for the speaker’s intentions to establish incompatible restrictions. Suppose, for example, the speaker says to himself not only “I intend to use the name n to represent the name m when ascribing attitudes to the F”, but also “I intend to use the name n to represent the name m* when ascribing attitudes to the G”, where m and m* are different names, and the speaker is unaware of the fact that the F is the G. It would seem that we would then have incompatible restrictions involving the individual a who, in fact, is both the F and the G. Since no correlation function can satisfy these incompatible restrictions, it would follow that no ascription of an attitude to a by the speaker (in the relevant context) could be true—and that doesn’t seem right. Next recall (19a), in which there is an anaphoric relationship between the proper name ‘Susan’ and the pronoun ‘her’ within the embedded 12 Richard’s account accommodates (26) above by taking its truth to be guaranteed in a context C if there is a planet p plus correlation functions f and g such that f maps ‘it is visible only in the morning’ onto α is visible only in the morning and g maps ‘it is visible only in the evening’ onto β is visible only in the evening, and in the morning Ralph accepted α is visible only in the morning while in the evening Ralph accepted β is visible only in the evening, and α is visible only in the morning expresses in Ralph’s morning context the same Russellian proposition as ‘it is visible only in the morning’ expresses in C with respect to an assignment of p to the variable ‘it’, and β is visible only in the evening expresses in Ralph’s evening context the same Russellian proposition as ‘it is visible only in the evening’ expresses in C with respect to an assignment of p to the variable ‘it’.

130 • Essay Four

clause. In cases like this, Richard suggests that the anaphoric pronoun be treated as a repetition of its antecedent.13 Thus, the annotated proposition expressed by the clause will contain multiple occurrences of the name ‘Susan’, making the attitude ascription reflexive. Richard would presumably apply this analysis even to examples like (7b), Mary says, and believes, that John loves his mother, which formed part of the motivation for the bound variable analysis discussed above. Does this mean that the bound variable analysis is superfluous within Richard’s framework? Examples like those in (8) suggest that it is not. (8) a. Mary told John that he wasn’t John. b. John fooled Mary into thinking that he wasn’t John. Treating the anaphoric pronouns in these examples as repetitions of their antecedents would result, in Richard’s system, in ascriptions of absurd, (pseudo) reflexive assertions and beliefs to Mary. It is tempting to think, therefore, that in these cases the pronouns themselves should appear along with their referent John in the annotated propositions expressed by the sentences. However, this creates problems for examples like (9), which has a reflexive interpretation in which the children are characterized as believing that John has the property of being one who was told by Mary that one wasn’t John. (9) The children thought that Mary told John that he wasn’t John. This interpretation cannot be captured by taking the pronoun to be a repetition of its antecedent. However, if the pronoun is treated simply as a distinct term which also refers to the man John, then the annotated proposition (9a) < . . . . . . . . . . . . > expressed by the sentential complement of the verb ‘thought’ will not be reflexive in the right way. In particular, it will not indicate that the children recognized that the person Mary was talking to is the person she claimed not to be John. Short of appealing to the bound variable analysis, the only remaining hope of capturing the reflexivity of (9) in Richard’s system would be to expand his conception of a contextual restriction constraining acceptable mappings f from words in the complement clause to words used by the children to include something along the lines of (9b). 13

Richard (1990, 214–15).

Attitudes and Anaphora • 131

(9b) f(‘John’) ≠ f(‘he’), but the children accept the concatenation of f(‘John’), f‘=’, and f(‘he’) in their context. The idea is to ensure that if (9a) were true, as used in the context, then the children would have to be aware that the person Mary was addressing was the person she claimed not to be John. However, there are serious problems with this strategy. For one thing, we want (9c) to be a logical consequence of the conjunction of (9) with the claim that John exists. (9c) Someone is such that the children thought that Mary told him that he wasn’t John. In Richard’s system, (9c) can be represented as (9d). (9d)

∃x(the children thought that Mary told x that x wasn’t John).

On the proposal we are considering, the truth of (9) guarantees that the children accept both (9e) and (9d). (9) e. Mary told β that α wasn’t β. f. α = β. For Richard, the truth of (9d) requires that the children accept something of the form (9g). (9g) Mary told α that α wasn’t β. It will follow from the fact that the children accept (9e, f) that they also accept (9g), only if it is logically guaranteed that the set of sentences someone accepts is closed under the relevant sort of substitution of equals for equals. Richard maintains that, as a general rule of thumb, a person will accept the results of such substitutions. However, he concedes that it is probably possible for someone to accept both S and an identity statement α = β without accepting all logical consequences S′ that result from substituting one of these terms for the other.14 As a result, the proposal we are considering for capturing the reflexivity of (9) fails to characterize as logically valid an inference that should be so characterized. A second difficulty with the proposal is illustrated by (9h), which involves another level of embedding. (9h) The men said that the children believe that Mary told John that he wasn’t John. 14

Richard (1990, 212).

132 • Essay Four

Here the context C of utterance would have to restrict the function f mapping words of the speaker onto those of the men in the manner indicated by (9i). (9i) f(‘John’) ≠ f(‘he’), but the men accept the concatenation of f(‘John’), f‘=’, and f(‘he’) in their context, C′. This ensures that the truth of (9h) in the speaker’s context C requires the men to have assertively uttered something of the form (9j) , while accepting the coreferentiality, in their context, C′, of distinct terms α and β. (9j)

The children believe that Mary told α that β wasn’t α.

However, it is not enough to capture the fact that the truth of (9h) requires the men to have characterized the children as holding a reflexive belief that attributes to John the property of being one who was told by Mary that one wasn’t John. Moreover, there seems to be no reasonable way to capture this with the proposal we are examining. These problems can all be avoided by analyzing anaphoric pronouns with c-commanding antecedents as bound variables (in particular, as variables bound by a predicate abstraction operator). Introducing this analysis into Richard’s system would result in two different mechanisms for achieving reflexive interpretations—one involving variable binding of the sort discussed above in the first section of this essay, and the other involving multiple occurrences of the same expression. In particular, a pronoun with a singular term antecedent could be treated as a bound variable in cases like (7)–(9), in which the antecedent c-commands the pronoun; while a pronoun with a non-c-commanding singular term antecedent could be treated as a repetition of the antecedent. Between the two, all cases of reflexivity that we have looked at thus far could be handled. A further reason, from the perspective of Richard’s system, for making room for the bound variable analysis within it is illustrated by (34d). (34) a. b. c. d.

Mary told John that he wasn’t John. Mary told Bill that he wasn’t Bill. Mary told Dick that he wasn’t Dick. In short, Mary told each man that he wasn’t that man.

Looking at (34d) in isolation one would naturally suppose that both ‘he’ and ‘that man’ function as variables bound by the quantifier, ‘each man’, and that (34d) is equivalent to (34e). (34e) [∀x: Man x] (Mary told x that x was not x). However, in Richard’s system this cannot be right, since in that system the truth of (34e) requires Mary to have recognized in each case that she was,

Attitudes and Anaphora • 133

absurdly, saying that someone was not himself. Thus, some other means must be found for capturing the nonreflexive interpretation of (34d). Once the predicate abstraction analysis has been introduced into Richard’s system, the way to do this would seem to be as follows. First, (35a) is assigned as the analysis of (34a). (35a) λx[Mary told x that x wasn’t John] John. Next we universally generalize on the two occurrences of ‘John’ in (35a) to produce (35b), which provides the nonreflexive interpretation of (34d). (35b) [∀y: Man yl (λx[Mary told x that x wasn’t y] y). Here, predicate abstraction, paradigmatically a device for producing reflexive interpretations, is used in our imagined modification of Richard’s system to provide a nonreflexive interpretation of (34d). It does this by introducing a variable-binding operator distinct from the quantifier that allows the expressions ‘he’ and ‘that man’ to be treated as different variables.15

The Return of the Problem This completes my investigation of strategies that limit the bound variable interpretation of anaphora to cases in which the antecedent c-commands the pronoun, while positing other interpretive mechanisms to produce reflexive-type interpretations in the remaining cases. Although the version of this strategy that combines the bound variable analysis with Richard’s account of annotated propositions and attitude ascriptions is able to accommodate a substantial range of data, it does not solve all our problems. This is shown by examples like (36). (36) Ralph said that any man who sees Susan in that costume will tell you that she isn’t Susan. This example seems to have a reflexive interpretation that parallels the interpretation of (9). However, since in this case the antecedent ‘Susan’ does not c-command the pronoun, we cannot treat the pronoun as a bound variable. Moreover, in the Richard system the pronoun cannot be taken to be simply a repetition of the antecedent. Thus, even with all the resources at our disposal, we still don’t have an adequate means of dealing with (36). This suggests that we haven’t gotten things right. 15 Of course, to make this stick one would have to argue that the semantic structure (35b) was appropriately related to the syntactic structure assigned to (34d) by a reasonable grammar; I have not tried to do that.

134 • Essay Four

In fact, we are left with a version of the very problem that motivated our exploration of the Richard system in the first place. On the presumption that anaphoric occurrences of pronouns with singular term antecedents can function as bound variables only if they are c-commanded by their antecedents, how can we account for apparent reflexive interpretations in other syntactic environments? The answer seems to be that we can’t, or, at any rate, that we don’t see a way of generating the needed reflexive interpretations in the full range of cases. Given this, perhaps we should abandon the presumption, and allow variable binding in a wide range of syntactic environments, including those in which the singular term antecedent does not occupy a position from which a natural language quantifier phrase can standardly bind the pronoun. Certainly, this alternative deserves careful consideration. Until it, or some other alternative, is worked out in detail we cannot claim to have an adequate account of attitudes and anaphora.16

References Davies, Martin. 1981. Meaning, Quantification, and Necessity. London: Routledge and Kegan Paul. Evans, Gareth. 1977. “Pronouns, Quantifiers, and Relative Clauses (I).” Canadian Journal of Philosophy 7:467–536. Neale, Stephen. 1990. Descriptions. Cambridge: MIT Press. 16 One possible alternative combines the idea that grammatically proper names are quantifiers whose scope is limited to the constituents they c-command with the treatment discussed above of unbound pronouns with non-c-commanding quantified antecedents. On this alternative the pronoun in (36) could be treated as a quantifier (a repetition of its antecedent) that is assigned wide scope in the manner indicated in (i).

(i) Ralph said that [Susan x](any man (who sees Susan in that costume) will tell you that x isn’t Susan). Of course, this strategy will produce an appropriate reflexive interpretation only if coreferential proper names and/or indexicals are not intersubstitutable, so that Ralph’s assertion of (ii) [She (pointing at Susan) x](any man (who sees Susan in that costume) will tell you that x isn’t Susan) is not sufficient for the truth of (i). Moreover, it would still leave quantified examples like (19d) (and similar examples involving quantificational variants of (36)) to be accounted for. Perhaps one could invoke Richard’s framework for these cases. However, the resulting theory would be highly eclectic, and sufficiently different from any presently existing account that it is unclear precisely what its consequences would be, or whether it could be made attractive.

Attitudes and Anaphora • 135 Richard, Mark. 1990. Propositional Attitudes. New York: Cambridge University Press. Salmon, Nathan. 1992. “Reflections on Reflexivity.” Linguistics and Philosophy 15:53–63. Soames, Scott. 1989–90. “Pronouns and Propositional Attitudes.” Proceedings of the Aristotelian Society 90, part 3: 191–212.

This page intentionally left blank

PA RT T WO

Modality

This page intentionally left blank

ESSAY FIVE

The Modal Argument: Wide Scope and Rigidified Descriptions

The Modal Argument In Naming and Necessity,1 Saul Kripke gives three types of argument against semantic theories that analyze the meaning, or reference, of proper names in terms of the meaning, or denotation, of descriptions associated with those names by speakers. One type consists of semantic arguments designed to show that, typically, the referent of a proper name n, as used by a speaker s, is not linguistically determined to be the denotation of any description, or set of descriptions, associated with n by s. One such argument is that a speaker’s use of n may uniquely refer to an object o, even though the speaker has no uniquely denoting description at all associated with n; other arguments maintain that a speaker’s use of n may refer to o even though the speaker associates n with descriptions that pick out something other than o. The second type of argument given by Kripke is epistemic. These arguments are the designed to show that the epistemic status of sentences containing names is different from the epistemic status of corresponding sentences containing descriptions. For example, if D is the description associated with a name n by speakers, then the proposition expressed by the sentence If n exists, then n is D is typically not knowable a priori even though the proposition expressed by If D exists, then D is D is knowable a priori. This supports the conclusion that D does not, in fact, have the same meaning as n. The third, and final, type of argument used by Kripke against description theories of names is the modal argument. This argument is based on the observation that the modal profile of sentences containing names often differs from the modal profile of corresponding sentences containing descriptions. This fact is used to show that the meanings of names are not given by the descriptions associated with them by speakers. I will be concerned with this argument. 1

Kripke (1980).

140 • Essay Five

My reconstruction of modal argument is as follows. The Modal Argument (1) Proper names are rigid designators. (2) Therefore proper names do not have the same meanings as nonrigid descriptions. So, if N is a proper name, and D is a nonrigid description, then the sentences N is F and D is F typically do not have the same meaning, or express the same proposition. (3) Since the descriptions commonly associated with names by speakers are nonrigid, typically the meanings of names are not given by those descriptions. So, if N is a name and D is a description associated with N by speakers, then the sentences N is F and D is F typically do not have the same meaning, or express the same proposition. The most important step in the argument is the first one, which requires establishing that names are rigid designators. The way this is done can be illustrated using the name ‘Aristotle’. To say that ‘Aristotle’ is a rigid designator is to say that it denotes the same thing in (or at) all possible worlds. The reason we think it does this is that we think that the truth-values, at different worlds, of sentences containing the name always depend on the properties of one and the same individual at those worlds. For example, we take the sentence ‘Aristotle was a philosopher’ to be true at a world w iff a certain individual—the person who was actually Aristotle—was a philosopher in w. Since a sentence α is F is true at an arbitrary world w iff the denotation of α at w is in the extension of F at w, we conclude that for any arbitrary world w ‘Aristotle’ denotes in w the individual who was Aristotle in the actual world. The key point here is the claim that the truth-value of the sentence ‘Aristotle was a philosopher’ at a world w always depends on whether or not the person we call ‘Aristotle’ in the actual world is a philosopher in w. Why do we think this? Couldn’t people in w have given the name ‘Aristotle’ to some other person, and so taken the sentence to be about him? Of course they could; but that is irrelevant. When we say that the sentence ‘Aristotle was a philosopher’ is true at w, we are saying that the sentence, as we actually understand it, is true when taken as a description of how things stand, according to w. In other words, to say that a sentence is true at a world w is to say that the claim or proposition that we actually use the sentence to express would be true if w obtained. Thus, our ultimate ground for thinking that the name ‘Aristotle’ is a rigid designator is our conviction that there is a certain individual x, such that for every possible world w, the proposition that Aristotle was a philosopher is true at w iff x was a philosopher at w, and similarly for other

The Modal Argument • 141

propositions. This feature of the name differentiates it from a description like ‘the teacher of Alexander’. The proposition that the teacher of Alexander was a philosopher is true at an arbitrary world w iff one and only one person taught Alexander at w, and that person was a philosopher at w. Since different people teach Alexander at different worlds, the description ‘the teacher of Alexander’ is not rigid. Hence, by the modal argument, it does not give the meaning of the name ‘Aristotle’. As I have already indicated, the modal argument was just one of several arguments given by Kripke against description theories of proper names. As such it was never intended to constitute, all by itself, a decisive refutation of all such theories. Rather, it was intended to be used in conjunction with the other arguments to produce that result. Nevertheless, the modal argument has been the main focus of attention for proponents of descriptivism, who have developed two main strategies for challenging it. Both strategies claim that names are semantically equivalent to descriptions, but descriptions of a certain special sort. According to the first strategy, names are equivalent to descriptions that are semantically required to take wide scope over modal operators occurring in the same sentence.2 This strategy amounts to a denial that names are rigid designators, plus an alternative proposal to account for the semantic data on which the doctrine of rigidity is based. According to the second strategy for challenging the modal argument, names are semantically equivalent to rigidified descriptions, and so are rigid designators. Since speakers have at their disposal the linguistic resources to convert ordinary nonrigid descriptions into corresponding rigid descriptions, proponents of the second strategy take the meanings of proper names to be given by the rigidified descriptions that speakers associate with them.3 My aim will be 2 This strategy was suggested by Dummett (1973, 110–51). He also defends a variant of the strategy in Dummett (1981, chap. 9 and appendix 3). The initial (1973) variant of the strategy maintains that the thesis that names are rigid designators just is the thesis that they take wide scope over modal operators. (See in particular pages 128 and 134.) In response to Kripke’s criticism in the preface of Naming and Necessity, Dummett (1981) presents a second variant of the view, which acknowledges that rigidity and wide scope are alternative theoretical notions used by semantic theorists to account for pretheoretic semantic facts and intuitions. What is common to the two variants is the claim that all genuine pretheoretic semantic facts and intuitions bearing on the dispute can be accommodated by treating names as nonrigid descriptions that take wide scope over modal expressions. A more recent version of the wide-scope position is given by David Sosa (1996, chap. 3, “Russell and Rigidity”). It was this work, plus conversations with David, that provided the chief initial impetus for the present essay. 3 This strategy is discussed sympathetically by Jason Stanley (1997a, sec. 5). A view different in some details, but similar in spirit to the view that names are rigidified descriptions, is advocated by Michael Jubien (1993).

142 • Essay Five

to examine these two strategies in more detail, to demonstrate that they won’t work, and to explain why.

The Analysis of Proper Names as Wide-Scope Descriptions Recall the conclusion we reached earlier about our grounds for thinking that the name ‘Aristotle’ is a rigid designator. We think that ‘Aristotle’ is rigid because we believe that which is expressed by principle (GR). GR. There is a certain individual x, such that for every possible world w, the proposition that Aristotle was a philosopher is true at w iff x was a philosopher at w, . . . and so on for other propositions expressed using the name ‘Aristotle’. Note that (GR) contains an occurrence of the name ‘Aristotle’, embedded under a modal quantifier—one ranging over possible worlds. Suppose we replace this occurrence of the name with a nonrigid description,  the G, which denotes the man Aristotle in the actual world, and which is required to take large scope over all modal predicates, operators, and quantifiers in the same sentence. This replacement of the name ‘Aristotle’ by a wide-scope description gives us a simulated rigidity principle, (SR i), whose content is explicitly given by (SR ii).4 SR.

(i) There is a certain individual x, such that for every possible world w, the proposition that the G was a philosopher is true at w iff x was a philosopher at w, . . . and so on for other propositions expressed using the name ‘Aristotle’. (ii) [the y: Gy] (there is a certain individual x, such that for every possible world w, the proposition that y was a philosopher is true at w iff x was a philosopher at w, . . . and so on for other propositions expressed using the name ‘Aristotle’)

Since the description the G denotes Aristotle, principle (GR) is true iff (SR) is true. The proponent of the wide-scope analysis now asserts that 4 In discussing the wide-scope analysis, I will employ formal representations in which definite descriptions are restricted quantifiers, the scopes of which, like that of other quantifiers, are the formulas to which they are immediately prefixed. 5 To say that a name is synonymous with a wide-scope description is to say that arbitrary sentences containing the name express the same propositions as corresponding sentences in which the description is substituted for the name and given the appropriate wide scope. This point can also be put in terms of the distinction between “assertive content” and “ingredient sense” drawn by Michael Dummett (1973, 446–47; 1981, 572–73). Roughly speaking the assertive content of a sentence (in a context of utterance) is the proposition expressed by the sentence (in that context), while the ingredient sense is what the sentence

The Modal Argument • 143

the name ‘Aristotle’ is synonymous with the wide-scope description the G that appears in (SR).5 Thus, he maintains that our original reason for taking the name ‘Aristotle’ to be rigid—namely (GR)—really is nothing more than (SR), which simulates rigidity. On this view, the semantic intuitions underlying the original rigidity claim are compatible with a treatment of proper names as having descriptive contents. Moreover, the proponent of this analysis argues that facts about propositional attitudes show that names really do have descriptive meanings, and propositions expressed by sentences containing names are identical with those expressed by sentences containing descriptions. The argument is based on the widely held presumption that often it is possible to assert or believe the proposition expressed by a sentence A is F without asserting or believing the proposition expressed by a corresponding sentence B is F, even though the two sentences differ only in the substitution of coreferential names. For example, it is widely presumed to be a fact that one can assert and believe the proposition expressed by the sentence ‘Hesperus is seen in the evening’ without asserting or believing the proposition expressed by the sentence ‘Phosphorus is seen in the evening’, despite the fact that the names ‘Hesperus’ and ‘Phosphorus’ are coreferential. Proponents of description theories claim that the explanation of this putative fact is that speakers associate the names with different, nonequivalent descriptions, the E and the M, respectively; hence the proposition expressed by ‘Hesperus is seen in the evening’ is just the proposition expressed by the E is seen in the evening and the proposition expressed by ‘Phosphorus is seen in the evening’ is just the proposition expressed by the M is seen in the evening. Everyone agrees that one can assert and believe one of these descriptive propositions without asserting or believing the other. The driving force behind the wide-scope analysis is the desire to preserve this explanation of the behavior of names in propositional attitude constructions, while also explaining their behavior in modal constructions. In propositional attitude ascriptions, the different descriptions associated by speakers with codesignative names are invoked to explain the apparent possibility of substitution failure. In modal constructions, the

contributes to the assertive contents of larger sentences in which it may be embedded. Dummett insists that there are natural examples in which two sentences, S1 and S2, have the same assertive content (express the same proposition) in a context of utterance, but have different ingredient senses because Operator S1 and Operator S2 have different assertive contents (express different propositions) in the relevant context. Phrased in these terms, when I say that according to the wide-scope analysis names are synonymous with widedescriptions I am making a claim about the assertive contents of sentences, not about the ingredient senses of sentences or expressions.

144 • Essay Five

wide scope given these descriptions is used to explain the apparent rigidity of names, and the accompanying guarantee of substitution success.6 That, as I see it, is the basic idea behind the wide-scope analysis. I will now try to state the analysis a bit more precisely. In doing so, we let S(n) be a sentence of English containing an occurrence of a name n; we let d be a description and S(d) be the result of substituting d for each occurrence of n; similarly, we take S(x) be the result of replacing each occurrence of n with the variable ‘x’. According to the analysis, the proposition expressed by S(n) is the proposition expressed by S(d), on an interpretation in which each occurrence of d (that replaces an original occurrence of n in S(n)) is given wide scope over every modal operator, modal predicate, and modal quantifier in S(x), except those for which doing this would involve removing d from the scope of some propositional attitude verb. When S(n) contains no modal operators, modal predicates, or modal quantifiers, the proposition it expresses is the same as the proposition expressed by S(d).7 Let us now consider some example sentences. If the content of the name n is given by the description the G, then the proposition expressed by N is F is the proposition expressed by (The x: Gx) Fx. Similarly, the proposition expressed by (Necessarily) John believes that n is F is the proposition expressed by (Necessarily) John believes that [(the x: Gx) Fx]. However, the propositions expressed by Necessarily n is F and  Necessarily if n is F, then something is both F and G are the propositions expressed by (The x: Gx) Necessarily [Fx] and (The x: Gx) Necessarily [Fx ⊃ ∃ y (Fy & Gy]. Finally, a point of clarification. In stating the analysis I have made use of the notion of the proposition expressed by a sentence. The analysis 6 Where the names ‘Hesperus’ and ‘Phosphorus’ are (semantically) associated with the codesignative descriptions the E and the M, respectively, it is assumed that the ascriptions Jones believes that Hesperus is F and Jones believes that Phosphorus is F may have different truth-values, since it is possible to believe that which is expressed by the E is F without believing that which is expressed by the M is F (and vice versa). However, given that the descriptions are coreferential (in the actual world) the modal sentences Necessarily Hesperus is F and Necessarily Phosphorus is F must have the same (actual) truthvalue, since the object which, in the actual world, is denoted by both descriptions either has the property expressed by F necessarily (in which case both modal sentences are true) or does not have the property necessarily (in which case both are false). 7 In addition, I will adopt a further proviso in the discussion that follows: an occurrence of a description corresponding to a name that is not required to take wide scope over a modal operator in the same sentence will be interpreted as taking the smallest possible scope. This proviso is heuristic, and is adopted to reduce the number of ambiguities we will have to consider in the examples that follow. This reduction will not affect the force of the criticisms to be developed, since in each case the problem will be that the analysis assigns certain interpretations to sentences that they do not, in fact, have.

The Modal Argument • 145

states that sentences of different sorts containing names express the same propositions as sentences of various kinds containing descriptions. For this reason it is worth saying a word about what I am assuming about propositions. I assume that, in addition to being expressed by sentences, propositions are both bearers of truth-value and objects of attitudes, such as believing and asserting. To say this is just to say that there are some things which can be asserted and believed, and that what is asserted and believed may also be true or false (either necessarily or contingently). That there are such things seems to be one of the evident commitments of our ordinary, prephilosophical speech. For purposes of this essay, the expression ‘propositions’ is simply a name for these things, whatever they turn out to be. No theoretical assumptions about their structure or about the nature of the relations, like belief and assertion, that we bear to them will be needed. We are now ready to criticize the analysis.

Arguments against the Wide-Scope Analysis The Basic Argument According to the analysis, the proposition expressed by the sentence If n is F, then something is both F and G is the proposition expressed by the sentence If the G is F, then something is both F and G. This gives us premise 1 of our argument. P1. The proposition that if n is F, then something is both F and G = the proposition that if the G is F, then something is both F and G Next we add premise 2. P2. The proposition that if the G is F, then something is both F and G is a necessary truth [((the x: Gx) Fx) ⊃ ∃y (Fy & Gy)] Clearly, C ought to follow from P1 and P2. C. The proposition that if n is F, then something is both F and G is a necessary truth [Fn ⊃ ∃y (Fy & Gy)]. However, on the wide-scope analysis, it does not follow, since, according to the analysis, C just is claim C′. C′.

The G is such that the proposition that if it is F, then something is both F and G is a necessary truth (the x: Gx) [ Fx ⊃ ∃y (Fy & Gy)]

146 • Essay Five

The problem for the wide-scope analysis is that whereas the argument from P1 and P2 to C is clearly valid, the analysis wrongly characterizes it as invalid. According to the analysis both P1 and P2 are true, while C— i.e., C′—may be false (when F and G are unrelated and the property expressed by G is not an essential property of the thing that actually has it). The reason that the wide-scope analysis has this consequence is that it treats linguistic constructions containing modal operators like ‘necessarily’, or modal predicates like ‘is a necessary truth’, as inherently shifty. In each case, the modal element combines syntactically with an argument A, a sentence in the case of the operator, ‘necessarily’, a noun phrase in the case of the predicate, ‘is a necessary truth’. When the argument A contains no proper names, the modal element is applied to the proposition expressed, or denoted, by A (depending on whether A is a sentence or a noun phrase). However, when A does contain a proper name, the modal operator, or predicate, is not applied to the proposition expressed or denoted by A; rather it is applied to a different proposition.8 To simplify matters let us focus simply on the modal predicate ‘is a necessary truth’.9 According to the wide-scope analysis, this predicate can be seen as expressing a modal property of propositions. When the predicate is combined with a term α that denotes a proposition p, the resulting sentence α is a necessary truth attributes the property of being necessarily true to p, provided that α does not itself contain any proper names. However, when α does contain a proper name the sentence α is a necessary truth does not attribute any property to the proposition denoted by α. For example if α is the proposition that Fn, and the name n is associated with the description the G, then the sentence α is a necessary truth is interpreted in such a way that it is true if and only if there is a unique individual o which has the property expressed by G, and which

8 For an illustrative comparison, see David Kaplan’s construal of ‘Ralph believes that’ as a “shifty” operator (1986, sec. 5). 9 In using the modal predicate, ‘is a necessary truth’, to give the argument, I am assuming the following equivalences:

(i) (ii) (iii) (iv)

The proposition that S is a necessary truth iff that S is necessary truth. That S is a necessary truth iff it is a necessary truth that S. It is a necessary truth that S iff it is necessarily true that S. It is necessarily true that S iff necessarily S.

Given these equivalences, the proponent of the wide-scope analysis cannot avoid the basic argument by claiming that the description associated with a name n takes wide scope in the constructions necessarily n is F and it is necessarily true that n is F while taking small scope in the construction the proposition that n is F is a necessary truth. The fact that such a strategy would require denying at least one of (i)–(iv) is itself an argument against it.

The Modal Argument • 147

is such that the singular, Russellian proposition that predicates F-hood of o has the property of being a necessary truth.10 This is why the wide-scope analysis of names is forced to treat some arguments with the apparent (grammatical) form (I) as invalid. I.

(i) α = β (ii) α is a necessary truth (iii) β is a necessary truth

When α contains no proper names but β does, the analysis treats (ii) as predicating the property of being necessarily true of a certain proposition p; the analysis views (i) as identifying p with proposition q; yet the analysis denies that (iii) predicates the property of being necessarily true of q. Because of this, (i) and (ii) may be characterized as true, while (iii) is characterized as false. (An analogous point holds for examples using the operators ‘it is necessarily truth that’ and ‘necessarily’.) The lesson to be drawn is clear. The wide-scope analysis purports to provide a correct description of the meanings of English sentences containing proper names, definite descriptions, and modal expressions. The fact that it wrongly characterizes obviously valid arguments as invalid shows that it fails to do this. This failure is not mitigated by the fact that the semantics it provides these sentences is conceptually coherent. There could be a language that worked in the way characterized by the widescope analysis, and in such a language many arguments of the (grammatical) form (I) would be invalid. But such a language is not English as we now understand it. It is not even clear that we should be willing to describe such a language as containing proper names in the sense that English does. Because of this the wide-scope analysis fails to throw light on how names actually function in English, or other natural languages.11

10 I here make an additional assumption—that the proposition denoted by the proposition that Fx, and expressed by Fx, with respect to an assignment of o as value of ‘x’ is the singular, Russellian proposition that predicates the property expressed by F of o. Other choices are theoretically possible, but this is by far the most natural and straightforward. Although I believe the choice to be correct, and will maintain the assumption throughout, the overall argument against the wide-scope analysis does not depend on it. 11 The argument given in this section applies to several positions in the literature, including the two variations of Dummett’s views mentioned in note 2. In the preface to Naming and Necessity (1980), Kripke criticized the identification of rigidity with wide scope in Dummett (1973), acknowledged that some intuitions about the truth-values of modal sentences containing names can be accounted for either by treating names as rigid or by treating them as nonrigid, wide-scope descriptions, and argued that we nevertheless have pretheoretic semantic intuitions about the modal profile of (the propositions expressed by) sentences containing names that cannot be accounted for by the wide-scope analysis. In appendix 3 of The Interpretation of Frege’s Philosophy, Dummett responds. He admits that

148 • Essay Five

A Variation on the Argument Before leaving this argument I would like to call attention to an implicit assumption I have invoked. I have assumed that the descriptive contents attributed by the analysis to proper names are also expressed in English by ordinary descriptive phrases of the form the G. This is worth mentioning because some proponents of the wide-scope analysis

according to the wide-scope analysis names are not rigid (1981, 594–95), and he maintains that the only genuinely pretheoretic semantic intuitions bearing on his dispute with Kripke concern the conditions under which assertive utterances of sentences express truths. Claims about the modal profile of sentences (relative to contexts)—i.e., about the truth-value at alternative possible worlds of that which a sentence expresses (in a context)—are decreed not to be directly testable by appeal to pretheoretic intuition, but to be matters of theoretical choice. (See, for example, p. 582.) For Dummett this means that we have no pretheoretic intuitions that we can bring directly to bear on the question “Is (that which is expressed by) S true, at all worlds, some worlds, or a certain world w?”; rather, the best we can do is appeal to intuitions that bear on the different, but related, question, “Is it the case that Necessarily S, Possibly S, or At world w, S is true?” He asserts that when intuitions are restricted in this way the wide-scope analysis can explain all of the genuinely pretheoretic semantic intuitions that the rigidity thesis can account for. (See pp. 577–79.) The argument involving (I) given above shows that this claim is incorrect, provided that there is a description, the G, that gives the content of the name. In that case, the wide-scope analysis will characterize as invalid, inferences classified as valid by pretheoretic intuitions (about the conditions under which various sentences express truths) that Dummett presumably deems to be legitimate. The argument also refutes a different, and more restricted, thesis advocated by Gareth Evans (1985). Evans is concerned with the special case in which the referent of a name, n, is semantically fixed to be the denotation of a description, the G, that is used in a stipulative introduction of the name. Evans argues (1985, 181) that in this sort of case (i) and (ii) have the same content—in my terminology, express the same proposition. (i) If there was a unique G, then the G was G. (ii) If there was a unique G, then n was G. Evans notes that there is an obvious objection to this view, which he credits to Kripke. Since the proposition expressed by (i) is necessary and the proposition expressed by (ii) is not, the two sentences cannot express the same proposition (1985, 182). Evans responds to this objection as follows: “I agree that sentences containing names embed differently under modal operators than do sentences containing descriptions, but it is perhaps the main point of this paper that the conclusion which Kripke draws from this fact follows only upon a questionable view of the connection between the content of an utterance and its modal properties” (1985, 182). Evans then goes on to sketch a semantic theory according to which sentences are assigned both contents (which serve as objects of propositional attitudes) and conditions under which the sentences are true at arbitrary possible worlds (which serve as the arguments for modal operators). It is further maintained that sentences may have the same contents (express the same proposition), even though they are associated with different conditions for being true at arbitrary worlds. Because of this, Evans argues that (i) and (ii) can express the same proposition even though (iii) and (iv), which differ only in the substitution of (ii) for (i), have different truth-values.

The Modal Argument • 149

have denied it.12 And, of course, if, for some n, there is no synonymous description, the G, then we will not be able to formulate any true premise of the form P1, and the above argument will be blocked. In addition, some proponents of the wide-scope analysis have used the possibility that names may have descriptive contents, even if they do not have the same contents as any ordinary descriptive phrases, to support a surprising and mysterious doctrine—namely that it makes no sense to attribute modal properties to propositions expressed by sentences containing names. This mysterious doctrine can be motivated as follows: Suppose that the name n has a wide-scope descriptive content which is not the content of any ordinary descriptive phrase in English. In that case we will not be able to formulate any true claim of the sort, ‘n is F’ expresses the

(iii) It is necessarily true that if there was a unique G, then the G was G. (iv) It is necessarily true that if there was a unique G, then n was G. The argument involving (I) given in the text—which is essentially just a reworking of Kripke’s original objection to Evans—shows that this is incorrect, assuming, as I do, that It is necessarily true that S, That S is necessarily true, and The proposition that S is necessarily true are equivalent. Like Dummett, Evans gives an empirically incorrect account of the semantics of English, since his position wrongly characterizes certain intuitively valid arguments as invalid. (There are a number of other important errors and confusions in Evans’s discussion, some of which are pointed out in Soames (1989, 148–50)). It should also be noted that although Kripke’s original objection to Evans’s claim that (i) and (ii) express the same proposition is correct, his own discussion in Naming and Necessity contains the seeds of Evans’s confusion on this point. There Kripke seems to suggest that one could know the proposition expressed by (ii) a priori, on the basis of a reference-fixing definition of the name. This was naturally taken by many to indicate that knowing the proposition expressed by (i) and knowing the proposition expressed by (ii) come to pretty much the same thing. From here it seemed a short step to identify the two propositions. My own view is that (ii) is not knowable a priori. See my review of Evans (Soames 1989). See also Donnellan (1979) and Salmon (1987–88). The argument in the text involving (I) also provides strong support for the rigidity thesis RT, questioned by Stanley (1997a, 1997b). RT. The rigidity of proper names demonstrates that utterances of sentences containing proper names, and utterances of sentences differing from those sentences only in containing nonrigid descriptions in place of proper names differ in content. [In my terminology: The rigidity of proper names demonstrates that the proposition expressed by a sentence containing a name, relative to a context, differs from the proposition expressed, relative to the same context, by a corresponding sentence in which a nonrigid description is substituted for the name.] If this principle were false, then for some name and description, we would have (i) and (ii) of the intuitively valid argument (I) characterized as true, while (iii) was characterized as false. Since any semantic theory leading to this result is inadequate, no adequate semantic theory of English falsifies RT. 12 See Dummett (1973, appendix to chap. 5, esp. pp. 135–37).

150 • Essay Five

proposition that the G is F. We can say, truly, that n is F expresses a proposition, and, using the sentence n is F, we can entertain this proposition, and make a variety of judgments about it. Strangely, however, we cannot assess its modal profile, its truth or falsity in different possible worlds. For, if we try to do this we find ourselves asking some such question as Is it the case that in world w n is F?. But in asking this question we have embedded n under a modal operator, thereby causing the descriptive content of n to “hop over” the content of the operator and take large scope. This has the effect of transforming our question into a different one from the one we intended. We intended to ask about the truth-value in w of the proposition expressed by n is F. We ended up asking about the truth-value in w of the singular Russellian proposition consisting not of the descriptive content of n together with the property expressed by F, but of the individual denoted by that descriptive content in the actual world, together with the property expressed by F. Moreover, the proponent of the mysterious doctrine asserts, there is no other way in which we can ask the question we intended. Instead, we must face the fact that questions about the modal profile of propositions expressed by sentences containing names don’t make sense. The best we can do is raise questions about the truth-values of propositions expressed by larger sentences containing modal operators, under which the proper names are embedded.13 How should we respond to this strange and desperate doctrine? Let us begin by supposing, for the sake of argument, that proper names have nonrigid, wide-scope descriptive contents which are not the contents of any descriptive phrase in English. If so, then propositions expressed by sentences containing names will not be expressible by us in any other way. For example, it may be that the proposition expressed by the sentence n is F is not expressed by any other sentence in our language. Still, there is nothing to prevent us from describing that proposition. Indeed we have already done so—it is the proposition expressed by the sentence  n is F, a proposition that consists of the descriptive content of the name n together with the property expressed by F. Since we can describe the proposition in this way nothing prevents us from using our description to ask about its modal profile. For example, we may ask “What is the truth-value of the proposition expressed by the sentence n is F in world 13 Dummett holds (i) that, typically, an ordinary proper name does not have the content of any single description, (ii) that it makes no intuitive sense to ask about the modal profile of (the proposition expressed by) a sentence containing a name, and (iii) that our intuitions are restricted to assessing the truth-values of sentences in which names are embedded under modal operators. As will be seen, this position, like the slightly more extreme doctrine presented in the text, provides no effective means of avoiding criticisms of the sort illustrated by (I) above.

The Modal Argument • 151

w?” This question does, of course, contain a modal phrase. However, since the name n occurs within quotes in the question, it is not given wide scope over that phrase. The alleged descriptive content of the name is not even a constituent of the propositional content of the question; so there is nothing here to be given wide scope. Thus we have succeeded in doing what the mysterious doctrine tells us can’t be done. We have asked an intelligible question about the modal profile of a proposition expressed by the sentence n is F. Moreover, it is not difficult to see how to go about answering it. We know that the proposition expressed by this sentence consists of the allegedly nonrigid descriptive content of the name together with the property expressed by F. We also know that, in general, any proposition expressed by a sentence α is Φ is true at a world w iff the denotation in w determined by the content of α is something which in w has the property expressed by Φ. Since we have been told that the name n has a descriptive content we know that the denotation determined by this content at a world is whatever individual, if any, uniquely possesses the relevant descriptive characteristics at that world. Surely this is something that is determinable in many cases—for if it is determinable in the actual world which individual corresponds to the descriptive content of a name, there is no reason the same shouldn’t hold true for other possible worlds. But this just means that often we can determine correct answers to questions about the modal profile of the proposition expressed by n is F. With this in mind, all that remains for us to do is to reformulate the original counterargument (I) against the wide-scope analysis, so that it applies even to those versions of the analysis which maintain that names have wide-scope descriptive contents that are not the contents of any descriptive phrases in English. The first premise of the reformulated argument is P1a, which surely is undeniable. P1a. The proposition that n is F = the proposition expressed by the sentence ‘n is F’. Next consider claims (1) and (2). (1) The proposition expressed by the sentence ‘n is F’ is true at world w. (2) The proposition that n is F is true at world w. According to the wide-scope analysis, (1) will be characterized as true iff the descriptive content ascribed to n picks out, at w, an individual that has the property expressed by F at w; (2) will be characterized as true iff that same descriptive content picks out, in the actual world, an individual that has the property expressed by F at w. If the descriptive sense ascribed to n picks out different individuals at different worlds (as it

152 • Essay Five

must if the appeal to wide scope is to have a point), then for some worlds, the corresponding claims (1) and (2) will be characterized as having different truth-values. Because of this, the wide-scope analysis will fail to characterize the pair of inferences, (i) from P1a and (1) to (2), and (ii) from P1a and (2) to (1), as jointly valid. Since in fact they are both valid, the wide-scope analysis fails. A Related Confusion It is illuminating to note that essentially the same failure can be expressed in a slightly different way. As before, we begin with the undeniable premise P1a, to which we add the trivial truth P2a. P2a. For all worlds w, the proposition that n is F is true at w iff the proposition that n is F is true at w. From these two premises, Obv is an obvious consequence. Obv. For all worlds w, the proposition expressed by the sentence ‘n is F’ is true at w iff the proposition that n is F is true at w. According to the wide-scope analysis, however, P1a and P2a are true, while Obv is false. To see this, imagine that n is synonymous with the wide-scope description, the G. (If there is no such ordinary description available in the language, let the G be the x: x = n, where the name is taken to have a descriptive content that is uniquely satisfied by different individuals at different worlds.) Then, according to the wide-scope analysis, Obv is equivalent to (3). (3) (the x: Gx) (for all worlds w) [the proposition expressed by ‘(the x: Gx) [x is F]’ is true at world w iff the proposition that x is F is true at w] But when G expresses a property that different objects may have at different worlds, (3) will be false. Since, according to the wide-scope analysis, (3) is equivalent to Obv, the analysis mischaracterizes it as false (and the inference from P1a and P2a to Obv as invalid). This failure of the analysis is related to a persistent confusion about its content. A striking feature of the relationship between rigidity and wide scope is the frequency with which the two are confused.14 In particular, the wide-scope analysis has often been mischaracterized, even by proponents, as claiming that names are rigid. Although this is a mistake, it is an 14 See, for example, the early views of Dummett, discussed in notes 2 and 11. See also my criticism of Gareth Evans’s discussion of so-called E-type pronouns (Soames 1989, 145–46).

The Modal Argument • 153

understandable one. Recall our earlier discussion of the name ‘Aristotle’, in which I mentioned that the wide-scope analysis can account for the truth of the principle (GR), which constituted our original grounds for taking the name to be rigid. GR. There is a certain individual x, such that for every possible world w, the proposition that Aristotle was a philosopher is true at w iff x was a philosopher at w, . . . and so on for other propositions expressed using the name ‘Aristotle’. What I did not point out at the time was that in order to get from (GR) to the claim that ‘Aristotle’ is rigid we need an instance of Obv involving that name. ObvA. For all worlds w, the proposition expressed by the sentence ‘Aristotle was a philosopher’ is true at w iff the proposition that Aristotle was a philosopher is true at w; ditto for other examples involving the name. Together (GR) and ObvA entail R. R.

There is a certain individual x, such that for every possible world w, the proposition expressed by the sentence ‘Aristotle was a philosopher’ is true at w iff x was a philosopher at w, . . . and so on for other propositions expressed using the name ‘Aristotle’.

This is what is needed for the rigidity of ‘Aristotle’. For it is R that guarantees that the sentence ‘Aristotle was a philosopher’, as we now understand it, will be true at a world w iff a certain individual—the person who was actually Aristotle—was a philosopher in w. Given that a sentence α is F is true at an arbitrary world w iff the denotation of α at w is in the extension of F at w, we conclude that for any arbitrary world w, ‘Aristotle’ denotes in w the individual who was Aristotle in the actual world. Although the wide-scope analysis accommodates (GR), it characterizes ObvA, and R, as false. Thus it wrongly characterizes names like ‘Aristotle’ as nonrigid. The fact that proponents of the analysis have not always recognized this suggests that they too may have simply taken Obv for granted, thereby implicitly endorsing as genuine some of the pretheoretic semantic intuitions denied by the analysis. Argument 2 The second argument against the wide-scope analysis is a simple variation of the first that does not employ any premise explicitly identifying propositions. Instead, it is based on the following scenario: Bill assertively utters

154 • Essay Five

the sentence If n exists, then n is F, where F expresses some essential (but hidden and nonobvious) property of the bearer of the name n—such as the property of originating from a certain bit of genetic material. Suppose further that the bearer of n is the unique object with the property expressed by G, that speakers associate the nonrigid description the G with n, and that there is no necessary connection between the properties expressed by F and G.15 In such a case the following premises, P1 and P2, will be true, and recognized as such by the wide-scope analysis. P1. Bill asserted that if n exists, then n is F. P2. It is a necessary truth that if n exists, then n is F. However, C, which, in fact, follows from P1 and P2, may wrongly be characterized by the analysis as false. C. Bill asserted a necessary truth This is clear when the argument is symbolized (in accord with the wide-scope analysis) as follows:16 P1′. Bill asserted [that: n exists ⊃ Fn] P2′. (the x: Gx) [ (x exists ⊃ Fx)] C′. ∃p [Bill asserted p and p is a necessary truth] The key point is that, according to wide-scope analysis, P2 does not attribute necessity to that which Bill is said, in P1, to have asserted. According to the analysis, the truth of P2 requires the necessity of that which is expressed by the open formula (x exists ⊃ Fx), relative to an assignment to the variable ‘x’ of the unique object which has the property expressed by G. By contrast, the truth of P1 requires Bill to have asserted that which is expressed by the sentence (n exists ⊃ Fn). But, according to the widescope analysis, that which is expressed by (n exists ⊃ Fn) is not identical with that which is expressed by the formula (x exists ⊃ Fx), relative to 15 As before, it is not essential to the argument that there be a descriptive phrase in English expressing the descriptive sense attributed to the name. However, the argument is more simply presented if we assume that there is such a phrase. 16 I assume here that C involves quantification over objects of assertion—i.e., propositions, or, in Dummett’s terminology, the assertive contents of sentences (in contexts). The argument could, of course, be restated slightly to bring this out.

P1. Bill asserted the proposition that if n exists, then n is F. P2. The proposition that if n exists, then n is F is necessarily true. C. Bill asserted a proposition that is necessarily true. I also assume that English sentences like C cannot be represented adequately by standard substitutional quantification into sentential position. See Richard (1990, 75–78) for a brief sketch of some of the problems facing attempts to treat apparent instances of objectual quantification over propositions in English substitutionally.

The Modal Argument • 155

any assignment of an object to ‘x’. Rather, it is a descriptive proposition involving the sense of the name n. Since this proposition is not necessary, C is characterized as false in a situation in which P1 and P2 are characterized as true. As a result, the analysis wrongly characterizes an argument which is, in fact, valid as invalid. Because of this, the interpretations of sentences provided by the analysis are incorrect.17 Argument 3 The third argument against the analysis is based on examples of a slightly different type. (4) Necessarily, if Bill asserts (believes) that n is F, and n is F, then Bill asserts (believes) something true. (5) Necessarily, if Bill asserts (believes) that n is F, and everything Bill asserts (believes) is true, then n is F. Although sentences of this type express obvious truisms, many of them are wrongly characterized as false by the wide-scope analysis. The problem arises because different occurrences of the name n are assigned different scopes, and so end up being evaluated at different worlds. Note that in each sentence the modal operator takes the entire conditional in its scope. According to the wide-scope analysis both occurrences of the name are replaced by occurrences of an equivalent description— the G.18 Since one of these occurrences is in the content clause of a propositional attitude verb, its scope remains confined to that clause. Since the other occurrence is not in the scope of any propositional attitude verb, it is obligatorily assigned wide scope over the modal operator. The resulting symbolizations are the following:19 (4′) (the x: Gx) [(Bill asserts/believes [that: (the y: Gy) Fy] & Fx) ⊃ ∃p [(Bill asserts/believes p) & p is true]] 17 Like the argument based on (I) above, this argument also applies to the other positions mentioned in note 11. 18 The point made in note 15 applies here as well. 19 As before, I assume objectual quantification over propositions; however here this assumption is not needed for the argument. For example, consider the following version of 4', in which objectual quantification over propositions has been replaced by substitutional quantification into sentential position: (the x: Gx) [(Bill asserts/believes that [(the y: Gy) Fy] & Fx) ⊃∃S [(Bill asserts/believes that S) & S]]. This sentence is true iff there is a unique individual o that satisfies G, and for all worlds w, if Bill asserts/believes that [(the y: Gy) Fy] & Fx is true at w, with respect to an assignment of o to ‘x’, then ∃S [(Bill asserts/believes that S) & S] is true at w. Let w be a world in which o satisfies F, the unique object that satisfies G does not satisfy F, but Bill asserts/believes that which is expressed by  the G is F. Then w is a world at which Bill asserts/believes that [(the y: Gy) Fy] & Fx is true (with respect to an assignment of o to ‘x’). However, w may also be a world at

156 • Essay Five

(5′) (the x: Gx) [(Bill asserts/believes [that (the y: Gy) Fy] & (p) [(Bill asserts/believes p) ⊃ p is true]) ⊃ Fx] These examples pose two problems for the wide-scope analysis. The first is that each asserts the existence in the actual world of a unique individual with the property expressed by G. However, it is not obvious that any such existential claim is entailed by the original English sentences. More generally, the wide-scope analysis is incompatible with the existence of meaningful proper names that (i) do not denote any individual existing in the actual world, but (ii) sometimes occur embedded under modal operators (outside the scope of propositional attitude verbs) in true sentences of English. If proper names of this sort exist in English, then the widescope analysis is false.20 The second problem posed by these symbolizations involves cases in which the description the G is a nonrigid designator. For example, suppose that in the actual world o is the unique individual that has the property expressed by G, and that w is a possible world satisfying the following conditions: (i) o has the property expressed by F, but not G, in w; (ii) in w, Bill asserts the proposition expressed by the G is F; (iii) Bill doesn’t assert anything else (or anything else true) in w; and (iv) either there is nothing in w that uniquely satisfies G, or whatever object in w uniquely satisfies G does not satisfy F in w. The existence of such a world w falsifies (4′). (5′) is falsified by a world w′ in which (i) Bill believes the proposition expressed by the G is F; (ii) all of Bill’s other beliefs in w′ are true in w′; (iii) there is a unique object in w′ that satisfies G, and it also satisfies F in w′; but (iv) o does not satisfy F in w′. Since (4′) and (5′) are analyses that would be assigned to (4) and (5) by the wide-scope analysis, the analysis incorrectly characterizes these obvious truths of English as false. On the basis of all these arguments, I conclude that the wide-scope analysis of proper names is incorrect.

The Analysis of Names as Rigidified Descriptions I now turn to the other main descriptivist challenge to Kripke’s modal argument. This is the view that proper names are synonymous with rigidified which ∃S [(Bill asserts/believes that S) & S] is false. Suppose that ∃S (Bill asserts/believes that S) is true at w, because the substitution instance Bill asserts/believes that [(the y: Gy) Fy] is true at w. Still, the sentence (the y: Gy) Fy (or even the sentence Fn) is false at w. (Note, in the previous sentence the description, and name, are mentioned rather than used, and so cannot, on pain of quantifying into quotes, be given wide scope over ‘is false at w’.) Hence, according to the wide-scope analysis even the substitutional version of 4’ is false. 20 I am indebted to Mike Thau for drawing my attention to this point.

The Modal Argument • 157

versions of the descriptions associated with them by speakers. On this view, names are rigid designators; hence no appeal to wide-scope is needed to account for the substitutivity of codesignative names in modal constructions. However since codesignative names may be associated with different descriptive information, they are not inter-substitutable everywhere; most notably they are not inter-substitutable in propositional attitude constructions. In assessing this view it is crucial to understand how, according to it, names are to be rigidified. There are two main alternatives in the semantic literature, only one of which is promising for the rigidification analysis. The view in question involves using the actuality operator to construct definite descriptions. Syntactically, ‘Actually’ combines with a sentence or formula to form a more complex sentence or formula. Semantically, ‘Actually’ is an indexical, like ‘I’, ‘now’, and ‘here’. As such, its content—that which it contributes to propositions expressed by sentences containing it—varies from one context of utterance to another. For example, the sentence ‘I am hungry now’, used by me at time t, expresses a proposition that is true at an arbitrary world w iff at w Scott Soames is hungry at t; the same sentence used by Saul Kripke at t′ expresses a different proposition, one that is true at a world iff Kripke is hungry at t′ in that world. Similarly, the sentence ‘Actually Kripke wrote Naming and Necessity’ used by anyone in the actual world, Aw, expresses a proposition that is true at an arbitrary possible world iff in Aw Kripke wrote Naming and Necessity; the same sentence used by a speaker at a different world w* expresses a proposition that is true at an arbitrary world iff in w* Kripke wrote Naming and Necessity. It will be apparent from this explanation that whenever S is a true sentence, Actually S is a necessary truth. The corresponding fact about descriptions is the following: whenever a definite description, the x: Fx denotes an individual o in the actual world, the rigidified description the x: Actually Fx denotes o in all possible worlds in which o exists (and never denotes anything else). This follows directly from the standard semantics of ‘the’ and ‘Actually’. According to the semantics of ‘the’, the denotation of a description the x: Sx at an arbitrary world w is the unique object, if any, existing at w that satisfies the open formula Sx, at w. Where Sx is the formula Actually Fx, an object satisfies it at w iff that object satisfies F at the actual world. The position we are considering now claims that ordinary proper names are synonymous with rigidified descriptions of this sort. For example it might be claimed that the name ‘Aristotle’, as used by a particular speaker, is synonymous with the description ‘the actual teacher of Alexander’, which in turn is understood as ‘the x: actually x taught Alexander’. Two criticisms of this view can be found in the existing

158 • Essay Five

literature. I will mention them very briefly, and then put them aside in order to focus on a new criticism. One of these criticisms concerns the question of whether proper names (like variables relative to assignments) designate their referents even with respect to worlds in which those individuals don’t exist. David Kaplan and Nathan Salmon have argued, quite plausibly, that proper names should be understood in this way.21 If they are right, then proper names are not equivalent to descriptions that have been rigidified using the actuality operator—since these descriptions designate an object at a world only if the object exists at that world.22 Another criticism that I will not discuss is based on Kripke’s original epistemic arguments against description theories. These arguments were designed to show that typically the sorts of descriptions D associated by speakers with a name n are such that the proposition expressed by If n exists, then n is D is not knowable a priori, even though the proposition expressed by If D exists, then D is D is knowable a priori. These arguments illustrate the difficulty of identifying the contributions made by proper names to propositional attitude ascriptions with the contents of descriptions that speakers typically associate with the names. Since these arguments typically seem to hold even when D is an ‘Actually’-rigidified description, it is difficult to find specific descriptions that allow the analysis to get off the ground.23

21 Kaplan (1973, appendix X; 1989b, sec. 4); Salmon (1981, 32–40). The most striking examples employed in the arguments exploit the parallels between temporal and modal semantics, and involve sentences like ‘Plato is dead’ and ‘Locke anticipated Kripke’. The first of these is a sentence of the form n is F. On the usual view such a sentence is true at a time t iff n designates at t something that is in the extension of F at t. But then since ‘Plato is dead’ is true now, the name ‘Plato’ must now designate something—and what else other than the now nonexistent Plato? The point here is generalized to cover a variety of cases. 22 The argument for this conclusion depends on taking sentences containing definite descriptions to make existence claims, and hence on taking the domain of (the x: Fx) relative to a world (or circumstance of evaluation) to be a subset of the set of individuals existing at the world (or circumstance). This assumption is explicit in Salmon (1981). Either it is not made, or it is ignored, in Kaplan (1989a, 577). 23 It is sometimes suggested that the content-giving description associated with a name is something of like the x: actually x stands at the end of a causal-historical chain of such and such type connecting x to this use of the name ‘n’. (Imagine ‘such and such type’ being filled out with a correct account of the way reference is actually determined, and imagine the referent of this use of the name ‘n’ being determined by the context of utterance.) However, it seems clear that no such proposal can be correct, since descriptions of this sort do not, in general, give the contents contributed by names to propositional attitude ascriptions. If they did, then when I attributed to someone the belief that Venus is a star, I would be attributing to him a belief about a certain one of my uses of the name ‘Venus’, as well as a belief about the specific sorts of causal-historical chains that connect uses of names to their bearers. Clearly no such beliefs are being attributed to the ancient Babylonians when I say that they believed that Venus was a star.

The Modal Argument • 159

I want, however, to waive these difficulties, in order to concentrate on a further problem which, by itself, is sufficient to show that the contents of proper names are not given by rigidified descriptions of the form the x: Actually Fx. The problem involves the interaction of modal and propositional-attitude constructions; it is based on the elementary observation that not only individuals in the actual world, Aw, but also inhabitants of other worlds, have beliefs about Aristotle, for example. I, along with many others in the actual world, believe that Aristotle was a philosopher; and it is not unreasonable to suppose that we also believe of the actual world, Aw, that Aristotle was a philosopher in it. A similar point holds for a great variety of different possible worlds w. In w, I, along with others, believe that Aristotle was a philosopher; in addition, we may also believe of the world w, that Aristotle was a philosopher in it. However, in w we need not have any beliefs about the actual world Aw, which may be quite remote from w. This fact provides the basis for the following argument against the analysis of proper names as ‘Actually’-rigidified descriptions. P1. It is possible to believe that Aristotle was a philosopher without believing anything about the actual world Aw. In particular, there are worlds w* in which agents believe that Aristotle was a philosopher, without believing of Aw that anything was F in it, and hence without believing of Aw that the unique thing that was F in it was a philosopher. P2. Necessarily one believes that the actual F was a philosopher iff one believes of the actual world, Aw, that the unique thing that was F in it was a philosopher. C1. It is not the case that necessarily one believes that Aristotle was a philosopher iff one believes that the actual F was a philosopher. P3. If the content of ‘Aristotle’, as used in a context C, were identical with the content of the actual F, as used in C, then (i) the contents of (propositions expressed by) Aristotle was G and  The actual F was G in C would be the same, (ii) the propositions expressed by α believes that Aristotle was G and α believes that the actual F was G, in C, would be necessarily equivalent, and (iii) C1 above would be false. C2. The content of ‘Aristotle’, as used in a context, is not the same as the content of the actual F, as used in that context. Each premise in this argument is true. First consider P1. Surely it is a datum that agents could have believed that Aristotle was a philosopher even if things had been quite different from the way they, in fact, are. Must these agents also have had beliefs about the actual world? In asking this, I am, of course, asking about the world that I call ‘actual’ here and

160 • Essay Five

now; the world provided by the context for my present remarks. Presumably, in some merely possible world the agents there have no direct acquaintance, or epistemic contact, with this world that I am now calling ‘actual’; nor, in many cases, will they possess any uniquely identifying descriptions of it. As a result, often there will be no way for them to form beliefs about the actual world.24 They may, of course, have beliefs about worlds they call ‘actual’, but that is another matter. An agent who sincerely, and assertively, utters, in a possible world w, a sentence ‘Actually the earth is round’ expresses his belief of w that in it the earth is round. We may even decide that whenever an agent believes a proposition p, at a world w, he also believes of w that p is true in it. Such a principle would explain how all of us in the actual world have beliefs about the actual world.25 However, it does not provide a way for agents in other worlds to share those beliefs. Next consider P2 and P3. These premises are based on the standard Kaplan-style indexical semantics for ‘Actually’,26 plus an account of propositional attitude ascriptions as reporting relations to the propositions expressed by their content clauses. The relevant semantic ideas are given in (6).27 (6) For any possible context of utterance C, the sentence The actual F was G, expresses in C a proposition that says of the world, Cw of C, that the unique thing that “was F” in it “was G”. The proposition expressed by Jones believes that the actual F was G, in context C, is true when evaluated at an arbitrary world w, iff Jones believes, in w, the proposition expressed by The actual F was G in C. Hence, the proposition expressed by Jones believes that the actual F was G, in C, is true when evaluated at an arbitrary world w, iff in w, Jones believes of Cw that the unique thing that “was F” in it “was G”. It follows that when the actual world, Aw, is the world of the context, the proposition expressed by Jones believes that the actual F was G is true at an arbitrary world w iff in w Jones believes, of Aw, that the thing that “was F” in it also “was G”. 24 Even in the unlikely case in which there is an agent in some merely possible world who both believes that Aristotle was a philosopher and also has some beliefs about the actual world Aw, there is no reason to suppose that included among his beliefs is the belief that whoever was the F in Aw was a philosopher. (Remember that the F is the description that we, in the actual world, associate with the name.) 25 Whether or not we all employ an indexical actuality operator. 26 Kaplan (1987b). 27 In stating (6) I ignore complications that would result from adding a temporal dimension, and taking truth-values to be determined at time/world pairs rather than simply at worlds.

The Modal Argument • 161

In addition to being highly plausible in itself, (6) is something that the proponent of the view that names are ‘Actually’-rigidified descriptions cannot afford to deny. His view requires that ‘Actually’ be an operator which rigidifies a description while allowing it to retain its descriptive content. This requirement dictates that the content of the x: actually Fx, as used in a context C, be a descriptive condition involving the property expressed by F which, when applied to an arbitrary world w, is satisfied by the unique individual in w that has that property somewhere—not in w obviously, but in the world given by C. This just is the Kaplan semantics for ‘Actually’, given in (6). Of course, (6) and P3 implicitly include an account of the semantics of propositional attitude ascriptions which has the consequence that expressions with the same content in a context C can be substituted for one another in attitude ascriptions without change in the truth conditions of those ascriptions in C. As before, this view is both plausible in itself and difficult for the descriptivist to deny. His view is that proper names are synonymous with, and hence have the same contents in the same contexts as, ‘Actually’-rigidified descriptions. The main point of appealing to descriptions in the first place was to provide names with the content needed to explain their contributions to attitude ascriptions. This would be lost if the descriptivist were now to deny that expressions with the same content are inter-substitutable in attitude constructions.28 As a result, P2 and P3 must be accepted. But then since the argument is valid, and each of the premises is true, it follows that proper names are not equivalent to descriptions rigidified using the actuality operator.29 Hence, this version of descriptivism is false. 28 If this principle were denied, there would be nothing to prevent dispensing with descriptions altogether, and taking the contents (meanings) of names to be their bearers. 29 A related argument, similar in spirit to this one, can be found in Fitch (1981). There Fitch considers the view that a proper name like ‘Cicero’ is synonymous with a description  the x: Fx in this world, where the demonstrative ‘this world’ is treated as a directly referential term whose content in a context is the possible world of the context. On page 30 he argues that this view is incorrect because it wrongly predicts that speakers in trivially, and irrelevantly, different possible worlds would express different propositions when they utter ‘Cicero denounced Catiline’. I would like to thank David Braun for pointing this out to me. Another useful comment of his concerns the possibility of weakening slightly the formulation of my P1 and P2 above. What the argument requires is the possibility that an agent might believe that Aristotle was a philosopher without believing the content that The actual F was a philosopher has in this world. This will be the case if some agents in other worlds have no access to, or beliefs involving, the content that ‘actual’, or ‘actually’, has in this world. In the text I treat this content as being, or at least including, the actual world Aw itself. What Braun pointed out is that there may be some philosophers who question this precise construal of the content of the actuality operator, while accepting the contention that some agents in other worlds who have beliefs about Aristotle have no beliefs involving the content that the actuality operator has in this world. Such a philosopher could accept the

162 • Essay Five

It must be admitted that there is another possible version of descriptivism that is immune both to Kripke’s original modal argument, and to all further arguments given here. According to this version, proper names have the same contents as descriptions rigidified using David Kaplan’s ‘dthat’-operator.30 This operator combines with a singular definite description D to form a singular term dthat D whose content in a context is just the denotation of D in the context. Rigidified descriptions of this sort designate their referents even in worlds in which the referents do not exist; they also may be used to express propositions that are routinely asserted and believed at alternative possible worlds by agents who have no propositional attitudes about the actual world. This is all to the good. However, the price of this success is too great for any genuine descriptivist to bear. When the ‘dthat’-operator is applied to a description D, it completely obliterates the descriptive content of D, and leaves the rigidified description dthat D with no descriptive content at all. As a result, coreferential ‘dthat’-rigidified descriptions have the same content, and any advantage of distinguishing coreferential names by associating them with different descriptive contents is lost. Conclusion This concludes my discussion of descriptivist challenges to Kripke’s modal argument. If I am right, these challenges are unsuccessful. The original modal argument showed that names do not have the meanings (or contents) of ordinary nonrigid descriptions. The arguments given here show that they also do not have the meanings (or contents) of either widescope, or ‘Actually’-rigidified, descriptions. In addition, these arguments illustrate a general lesson about the importance of paying attention not just to modality, nor just to propositional attitudes, but to the interaction of the two. For example, once it is admitted that a proper name like ‘Aristotle’ rigidly designates a certain man, even in worlds in which he satisfies none of the identifying descriptions we commonly associate with the name, it can scarcely be denied that agents in those worlds may believe that Aristotle “is F” (for various Fs that apply to Aristotle in those worlds), without believing that anyone satisfied those descriptions. This means

conclusion of my argument, even if he maintained that the proposition expressed by Actually S in Aw is distinct from the metaphysically equivalent but epistemically different proposition expressed by In Aw S. I will not investigate this view here, but simply accept the observation that my argument could be formulated, if need be, so as to accommodate it. 30 Kaplan (1989b). See Kaplan (1989a) for a reconsideration of the viability of the ‘dthat’-operator.

The Modal Argument • 163

that we (in the actual world) can truly describe those agents by saying In w, so and so believes that Aristotle is F, even in cases in which the possible beliefs in question do not contain any of the descriptive content that we associate with the name. This fact represents a serious obstacle to any version of descriptivism that assigns substantive descriptive content to propositions expressed by sentences containing proper names.31 Despite these broader implications, my arguments, like the original modal argument, are silent about certain further questions concerning the relationship between names and descriptions. For example, I have not addressed the question of whether names are rigid designators the reference of which is fixed by descriptions (which do not give their meanings, or provide them with descriptive contents). Kripke, of course, allows that this is a genuine possibility, which may be realized by a small number of ordinary proper names. However, he also uses his semantic and epistemic arguments to show that the great majority of proper names do not work this way. When these arguments are combined with the ones discussed here, it is hard to avoid his original conclusion—typically, ordinary proper names aren’t descriptions at all.

References Donnellan, Keith S. 1979. “The Contingent Apriori and Rigid Designators.” In Contemporary Perspectives in the Philosophy of Language, ed. Peter A. French, Theodore E. Uehling, and Howard K. Wettstein, 12–27. Minneapolis: University of Minnesota Press. Dummett, Michael. 1973. Frege: Philosophy of Language. New York: Harper and Row. ———. 1981. The Interpretation of Frege’s Philosophy. London: Duckworth. Evans, Gareth. 1985. “Reference and Contingency.” In Collected Papers, 178–213. Oxford: Clarendon Press. Originally published in The Monist 62, no. 2 (1979): 161–89. Fitch, G. W. 1981. “Names and the ‘De Re–De Dicto’ Distinction.” Philosophical Studies 39:25–34. Jubien, Michael. 1993. “Proper Names.” In Language and Logic, ed. James E. Tomberlin, 487–504. Philosophical Perspectives 7. Atascadero, Calif.: Ridgeview. 31 One such theory for which this fact presents an obstacle holds that the content of a name, which it contributes to propositions expressed by sentences containing it, consists in a pair consisting of its referent plus the descriptive content associated with the name (whether or not that content plays a role in determining the referent of the name). According to one version of such a view, one who believes the proposition expressed by Aristotle is F believes both a singular proposition in which F-hood is attributed to a certain man, and a descriptive proposition, expressed by D is F, where D is the description associated with the name.

164 • Essay Five Kaplan, David. 1973. “Bob and Carol and Ted and Alice.” In Approaches to Natural Language, ed. Jaakko Hintikka, J.M.E. Moravcsik, and Patrick Suppes, 490–518. Dordrecht: Reidel. ———. 1986. “Opacity.” In The Philosophy of W. V. Quine, ed. Lewis E. Hahn and Paul A. Schilpp, 229–89. La Salle, Ill.: Open Court. ———. 1989a. “Afterthoughts.” In Themes from Kaplan, ed. Joseph Almog, John Perry, and Howard Wettstein with the assistance of Ingrid Deiwiks and Edward N. Zalta, 565–614. New York: Oxford University Press. ———. 1989b. “Demonstratives: An Essay on the Semantics, Logic, Metaphysics, and Epistemology of Demonstratives and Other Indexicals.” In Themes from Kaplan, ed. Joseph Almog, John Perry, and Howard Wettstein with the assistance of Ingrid Deiwiks and Edward N. Zalta, 481–563. New York: Oxford University Press. Kripke, Saul A. 1980. Naming and Necessity. Cambridge: Harvard University Press. Originally published in Semantics of Natural Language, ed. Donald Davidson and Gilbert Harman, 253–355 (Boston: Reidel, 1972). Richard, Mark. 1990. Propositional Attitudes. New York: Cambridge University Press. Salmon, Nathan. 1981. Reference and Essence. Princeton: Princeton University Press. ———. 1987–88. “How to Measure the Standard Metre.” Proceedings of the Aristotelian Society, n.s., 88:193–217. Soames, Scott. 1989. Review of Collected Papers, by Gareth Evans. Journal of Philosophy 86, no. 3: 141–56. Sosa, David. 1996. “Representing Thoughts and Language.” Ph.D. diss., Princeton University. Stanley, Jason. 1997a. “Names and Rigid Designation.” In A Companion to the Philosophy of Language, ed. Bob Hale and Crispin Wright, 555–85. Oxford: Blackwell. ———. 1997b. “Rigidity and Content.” In Language, Thought, and Logic: Essays in Honour of Michael Dummett, ed. Richard G. Heck Jr., 131–56. New York: Oxford University Press.

ESSAY SIX

The Philosophical Significance of the Kripkean Necessary A Posteriori

In a recent paper, I discussed Saul Kripke’s two routes to the necessary a posteriori—one correct and far-reaching, the other incorrect and misleading.1 In this essay, I will show how each connects with broader issues and agendas in philosophy. I will argue that Kripke’s first route has led to a distinction between metaphysical and epistemic possibility that is an important advance in analytic philosophy, while his second route has led to an attempted revival of pre-Kripkean orthodoxy which both threatens that advance, and leads to philosophically suspect results. I begin with a broad-brush sketch of the impact of Naming and Necessity, and other seminal works, on the conventional wisdom of the golden age of midtwentieth-century philosophy.2

The Antidescriptivist Revolution The antidescriptivist revolution led by Saul Kripke, Hilary Putnam, and David Kaplan challenged the following central tenets of the dominant philosophy of their time. (i) The meaning of a term is never its referent. Instead, it is a descriptive sense that encodes conditions necessary and sufficient for determining reference. (ii) Since the meaning of a word, as used by a speaker s, is the descriptive sense that s mentally associates with it, meaning is transparent. If two words mean the same thing, then anyone who understands both should be able to figure that out by consulting the sense that he or she associates with them. Word meanings and mental contents are entirely dependent on factors internal to speakers. (iii) A priori and necessary truth amount to the same thing. Both are grounded in meaning. 1 2

Soames (n.d.). Kripke (1980).

166 • Essay Six

(iv) Metaphysical claims about objects having or lacking properties essentially—independently of how they are described—make no sense. Even if a term t designates o and Necessarily t is F (if t exists) is true, there will always be another term t* designating o for which Necessarily t* is F (if t* exists) is false. Since it would be arbitrary to give either sentence priority in determining the essential properties of o, the idea that objects have, or lack, such properties must be relativized to how they are described. (v) Since the job of philosophy is not to come up with new empirical truths, its central task is conceptual clarification, which proceeds by the analysis of meaning. These doctrines and their corollaries provided the framework for the golden age of mid-twentieth-century analytic philosophy, when the necessary, the a priori, and the analytic were one, when all possibility was linguistic possibility, and when proponents of the linguistic turn had made philosophy palatable to the modern mind by giving it a respectable subject matter—language—and rendering it, if not exactly scientific, at least precise and rigorous. This analytic vision had survived family quarrels among the logical positivists, and adapted to the change from logical atomism to the ordinary language school of philosophy. It had even survived Quine’s criticism of the linguistic a priori, and his attack on the analytic-synthetic distinction—the former, mostly by ignoring it, and the latter by infecting it with the presupposition that the necessary, the a priori, and the analytic were one.3 Since it was surely too much to give up all three, Quine’s attack—which indiscriminately lumped them together—didn’t convince. All that changed with the introduction of rigid designation, direct reference, and nondescriptionality in the early 1970s. Kripke’s argument that names and natural kind terms are rigid designators, and so not equivalent to descriptions associated with them by speakers, was the entering wedge. He next used rigid designation to rebut Quine’s famous objection to essentialism, enshrined in (iv). A rigid designator t of an object o is one that picks out o in all possible circumstances in which o exists. This means that when t is rigid, the question of whether o has the property expressed by F essentially—which here means, simply, whether o has that property in all circumstances in which o exists—is equivalent to the question of whether the sentence Necessarily t is F (if t exists) is true. The truth-values of other sentences containing nonrigid designators are irrelevant. Once this was seen, Quine’s objection to the intelligibility of essentialism collapsed. 3

See Soames (2003, vol. 1, chaps. 12, 16, and 17).

The Kripkean Necessary A Posteriori • 167

With both a nondescriptive semantics and a rehabilitated conception of essentialism in place, Kripke next showed how to generate instances of the necessary a posteriori. If n is a name or indexical that rigidly designates o, F expresses an essential property of o, and knowledge that o has this property requires empirical evidence, then the proposition expressed by If n exists, then n is F is both necessary and knowable only a posteriori. All of a sudden, the necessary and the a priori were no longer the same, and the idea that one, or both, might arise from something beyond the linguistic became credible. With this essentialist route to the necessary a posteriori came a sharp distinction between conceivability and genuine possibility—between ways things could conceivably be versus ways things could really be (or have been). It is natural to draw this distinction in terms of possible worlds, or better, possible world-states. For the Kripkean, possible states of the world are not alternate concrete universes, but abstract objects. Metaphysically possible world-states are maximally complete ways the real concrete universe could have been—maximally complete properties that the universe could have instantiated. Epistemically possible world-states are maximally complete ways the universe can coherently be conceived to be—maximally complete properties that the universe can be conceived of as instantiating, and that one cannot know a priori that it doesn’t instantiate. These two sets of properties are different. Just as there are properties that ordinary objects could have had and other properties they couldn’t have had, so there are certain maximally complete properties the universe could have had—metaphysically possible world-states—and other maximally complete properties the universe couldn’t have had— metaphysically impossible world-states. Just as some of the properties that objects couldn’t have had are properties that one can conceive them as having, and that one cannot know a priori that they don’t have, so some maximally complete properties that the universe couldn’t have had are properties that one can conceive it as having, and that one cannot know a priori that it doesn’t have. These states of the world are epistemically, but not metaphysically, possible. On this picture—which Kripke himself didn’t make explicit, but could have—the reason empirical evidence is required for knowledge of necessary truths that are knowable only a posteriori is to rule out metaphysically impossible, but epistemically possible, world-states in which they are false. This is the heart of the philosophical revolution led by Kripke and his allies. By the time the dust had settled, all five of the central theses of the golden age of mid-twentieth-century analytic philosophy had been either decisively rejected, or called into question. As I have stressed, the Kripkean necessary a posteriori played a central role in this. However, his treatment of this topic was not as simple as my broad sketch suggests. As

168 • Essay Six

I indicated at the outset, Kripke’s work contains two routes to the necessary a posteriori. In addition to the essentialist route just noted, there is a second route, both extensionally and philosophically quite different from the first. It is to this distinction that I now turn.4

Kripke’s First, Essentialist, Route to the Necessary A Posteriori (1) Greg Soames ≠ Brian Soames (2) If Saul Kripke exists, then Saul Kripke is a human being. (3) This desk (pointing at the one in my office) was not made out of metal. (4) If this desk exists, then it consists of molecules. Since the propositions expressed by these sentences are true, they are, according to Kripke, necessarily true. However, they are also knowable only a posteriori. How can this be? How can a proposition that is necessary, and known to be so, also be knowable only a posteriori? Kripke’s answer appeals to our knowledge of essential properties and relations.5 We know a priori that being human, being a desk that was not made out of metal, and being a desk consisting of molecules are essential properties of anything that has them. We also know a priori that being nonidentical is a relation that holds essentially of any pair it relates. So, we know a priori that if any objects have these properties, or stand in this relation, then they have, or stand in, them in any genuinely possible circumstance in which they exist. Hence, we know a priori that propositions (1)–(4) are necessary, if true. Still, discovering that they are true requires empirical investigation. Thus, in order to discover what could and could not be, one sometimes must first discover what is. As indicated above, this route to the necessary a posteriori implicitly contains a sharp distinction between epistemic and metaphysical possibility—between ways that the universe could conceivably be (epistemically possible world-states), and ways that the universe could really be (metaphysically possible world-states). On this picture, some things that are coherently conceivable are not genuinely possible. How, then, are conceivability and possibility related? Kripke’s answer is based on the fact that when p is an instance of the necessary a posteriori of the

4

The next three sections summarize the more extensive discussion in Soames (n.d.). When speaking of (Kripkean) “essential” properties and relations, I mean simply properties and relations that hold necessarily of objects (in all genuinely possible world-states in which the objects exist). 5

The Kripkean Necessary A Posteriori • 169

sort just discussed, then although p is knowable only a posteriori, it is knowable a priori that if p is true, then p is necessary.6 The fact one cannot know p a priori means that one cannot know a priori that a world-state in which p is false is not instantiated. Such states are coherently conceivable, and so epistemically possible. The fact that one knows a priori that if p is true, then p is necessary means that one knows a priori that if a world-state in which p is true is instantiated, then no worldstate in which p is false could have been instantiated. Thus, when one finds, empirically, that p is true, one learns a posteriori that epistemically possible world-states in which p is false are metaphysically impossible. On this picture, the things we conceive when trying to determine what is metaphysically possible include not only individual world-states, but entire systems of metaphysical possibility, each with a designated “actual” world-state and a space of related states. Someone seeing my desk for the first time who doesn’t know what it was (originally) made of can conceive of a world-state in which it was made of mahogany, a worldstate in which it was made of oak, and perhaps even a world-state in which it was made of metal. One can conceive of each of these states being instantiated. Accompanying each state, one can conceive of related states that will be genuine metaphysical possibilities, if the initial, designated state is instantiated. So accompanying the designated (actual) state in which the desk was made of reddish-brown mahogany, one can conceive of related world-states in which it was made of mahogany stained another color. But given the supposition that the original state is instantiated, one can conceive of no state possible relative to it in which that very desk was (originally) made of some other material—e.g., oak or metal. A similar point holds for other epistemically possible world-states in which the desk was made of those things. When they play the role of the designated “actual” world-state—i.e., when one considers them as instantiated and asks which states are possible relative to them—one regards world-states in which the desk was made of mahogany as impossible relative to those states. So we have a set of epistemically possible world-states, each of which can coherently be conceived as being instantiated. Along with each such state w1, we have (epistemically possible) world-states w2 which we recognize to be metaphysically possible, if the initial, designated “actual” state w1 is instantiated—i.e., we recognize that if w1 were instantiated, then w2 would be a property that the universe could have had. Moreover, for each such state w2 there are (epistemically possible) world-states w3 6

Kripke (1971, 152–53).

170 • Essay Six

which we recognize to be metaphysically possible, if w2 is instantiated— i.e., we recognize that if w1 were instantiated, then w3 would be (metaphysically) possibly possible. Repeating this process indefinitely, we generate a coherently conceivable system of metaphysical possibility. Collecting all such systems together, we have a set of epistemically possible systems of metaphysical possibility. For a world-state to be genuinely metaphysically possible (or possibly possible) is for it to be a metaphysically possible (or possibly possible) member of some epistemically possible system of metaphysical possibility the designated world-state of which is the state that the world really is in. Though not a definition of metaphysical possibility in nonmodal terms, this is, I believe, an illuminating way of thinking about the relationship between conceivability and possibility. On this picture, conceivability is a fallible, but useful guide to metaphysical possibility. It is fallible because before we know much about what is actual, there are many epistemically possible world-states that appear to be genuinely possible, and so remain candidates for being metaphysically possible. The more we learn about the world, the more we whittle down this field of candidates, and the better able we are to identify the scope of genuine metaphysical possibility. In short, our guide to metaphysical possibility is conceivability plus knowledge of actuality. This relationship between the epistemological and the metaphysical is implicit in the following statement of Kripke’s essentialist route to the necessary a posteriori. ERNA Let p be a true proposition that attributes a property (or relation) F to an object o (or series of objects), conditional on the object (or objects) existing (while not attributing any further properties or relations to anything). Then, p will be an instance of the necessary a posteriori if (a) it is knowable a priori that F is an essential property of o, if F is a property of o at all (or a relation that holds essentially of the objects, if F holds of them at all), (b) knowledge of o that it has F, if it exists (or of the objects that they are related by F, if they exist) can only be had a posteriori, and (c) knowing p involves knowing of o (or of the objects) that it (they) have F, if it (they) exist at all. (o can be an individual or a kind.) The scope of the necessary a posteriori established by this route is determined by which properties and relations can play the role of F in ERNA. With the exception of one significant class of cases, Kripke’s own putative examples of the necessary a posteriori can all be derived by this

The Kripkean Necessary A Posteriori • 171

route, using either his own explicit doctrines about essential properties and relations, or plausible extensions of them. The examples that cannot be so derived are propositions expressed by identity sentences containing variables, names, or demonstratives, plus propositions expressed by corresponding sentences containing simple natural kind terms. Although such examples are standardly taken to be instances of the Kripkean necessary a posteriori, in fact, their status is doubtful. Let o and o* be objects to which the identity relation actually applies, and p be a proposition that (merely) attributes identity to the pair. Then, although conditions (a) and (c) of ERNA are satisfied, condition (b) is not, since knowledge of the pair—i.e., of —that identity truly applies to it can surely be had a priori. Thus, p is an example of the necessary a priori, not the necessary a posteriori. This point is illustrated by (5). (5) [∃x: x = Hesperus] [∃y: y = Phosphorus] it is a necessary truth that x = y. Since (5) is true, the proposition expressed by ‘x = y’, relative to an assignment of Venus to ‘x’ and ‘y’, is a necessary truth. However, since this proposition (merely) predicates identity of Venus and itself, it is knowable a priori, if anything is. Other problematic cases include those expressed by sentences of the form (6a), where m and n are simple, coreferential names, and those expressed by sentences of the form (7a), where K and K* are simple natural kind terms (rigidly) designating the same kind k, and is a K and is a K* are predicates applying to all and only instances of k. (6) a. b. (7) a. b.

n=m Hesperus is Phosphorus ∀x [x is a K ↔ x is a K*] Woodchucks are groundhogs (and conversely)

Since, according to Kripke, names don’t have descriptive senses, it is natural to take a sentence consisting of names plus a relational predicate R to semantically express a proposition which predicates the relation expressed by R of the referents of the names, without any further predication. On this model, the proposition expressed by (6b) merely predicates identity of Venus and itself. Although this proposition is necessary, it seems to be knowable a priori. One could, of course, avoid this conclusion by adopting the assumption (foreign to Kripke) that—in addition to predicating identity of Venus and itself—the proposition expressed by (6b) also predicates the properties of being visible in the evening and being visible in the morning of Venus. But then, the proposition will be

172 • Essay Six

contingent.7 Thus, although Kripke gives (6b), and other instances of (6a), as paradigmatic examples of the necessary a posteriori, one cannot arrive at this result by his standard essentialist route. Analogous remarks apply to instances of (7a).

Kripke’s Second (Attempted) Route to the Necessary A Posteriori The argument for the aposteriority of (6b) given in the last few pages of lecture 2 of Naming and Necessity is based on the observation that the evidence available to a speaker who understands ‘Hesperus’ and ‘Phosphorus’ is insufficient to determine that they are coreferential. Kripke illustrates this by noting that there are possible world-states w in which competent users of ‘Hesperus’ and ‘Phosphorus’ are in evidentiary situations qualitatively identical to ours (prior to the astronomical discovery), and yet, in w, the names refer to different things. This indicates that the evidence available to us, simply by being competent speakers, doesn’t establish (6c) or (6d), and, hence, that these propositions are not knowable a priori. (6) c. ‘Hesperus’ and ‘Phosphorus’ are coreferential. d. ‘Hesperus is Phosphorus’ expresses a truth. However, the lesson Kripke explicitly draws is that the proposition expressed by (6b) is not knowable a priori. So two things are true: first, that we do not know a priori that Hesperus is Phosphorus, and are in no position to find out the answer except empirically. Second, this is so because we could have evidence qualitatively indistinguishable from the evidence we have and determine the reference of the two names by the positions of the two planets in the sky, without the planets being the same. (104) In deriving this conclusion, Kripke seems to implicitly rely on a line of reasoning connecting speakers’ understanding and acceptance of sentences with our ability to use those sentences to report what they believe. On this line of reasoning, before the astronomical discovery speakers understood but didn’t accept sentence (6b), and so didn’t believe that Hesperus was Phosphorus. Since they wouldn’t have been justified in 7 Including these properties in the contents of ‘Hesperus’ and ‘Phosphorus’, and rigidifying using the actuality operator, would preserve the necessity of (6b) (or near enough). However, such an analysis fails on independent grounds. See chapter 2 of Soames (2002).

The Kripkean Necessary A Posteriori • 173

accepting (6b), based on the evidence then, they wouldn’t have been justified in believing that Hesperus was Phosphorus. But then, the proposition that Hesperus is Phosphorus must require empirical justification, in which case it must not be knowable only a priori. Here is the argument: (i) One who understands ‘Hesperus is Phosphorus’ (a) accepts it and believes it to be true iff one believes that Hesperus is Phosphorus, and (b) would be justified accepting it and believing it to be true iff one would be justified in believing that Hesperus is Phosphorus. (ii) In order to be justified in accepting ‘Hesperus is Phosphorus’ and believing it to be true, one needs empirical evidence that the two names refer to the same thing. Given that one knows that ‘Hesperus’ designates the heavenly body visible in the evening and that ‘Phosphorus’ designates the heavenly body visible in the morning, one needs evidence that these are the same. (iii) Since one needs empirical evidence in order to be justified in believing that Hesperus is Phosphorus, it is knowable only a posteriori. When stated in terms of the propositions expressed by sentences, this argument presupposes SDJ. Strong Disquotation and Justification (SDJ) If x understands S, uses S to express p, and knows that S expresses p, then (a) x believes p iff x accepts S (and believes S to be true), and (b) x would be justified in believing p on the basis of evidence e iff x would be justified in accepting S (and believing S to be true) on the basis of e. One who understands ‘Hesperus is Phosphorus’, while associating the names with ‘the heavenly body visible in the evening’, and ‘the heavenly body visible in the morning’, will justifiably accept the sentence and believe it to be true only if one justifiably believes that the heavenly body visible in the evening is the heavenly body visible in the morning. Since justification for this descriptive belief requires empirical evidence, justification for accepting ‘Hesperus is Phosphorus’ does too. SDJ transfers this requirement to one’s belief in the proposition one uses the sentence to express—presumably, the proposition that Hesperus is Phosphorus. Hence, knowledge of this proposition can only be a posteriori. In lecture 3, Kripke generalizes this explanation to all cases of the necessary a posteriori. After summarizing his analysis of natural kind

174 • Essay Six

terms, and illustrating their role in expressing instances of the necessary a posteriori, he takes up a challenge. Up to now, when describing these instances, he emphasizes that although they are necessary, for all we knew prior to empirically discovering their truth, they could have turned out otherwise. Realizing that this may sound puzzling, he tries to defuse a natural objection. Let p be any instance of the necessary a posteriori. Since p is a posteriori, its falsity must be conceivable, and so, it would seem, knowledge of p must require empirical evidence ruling out possibilities in which p is false. Without such evidence, it could turn out that p is false. But, the objector maintains, if p is necessary, there are no such possibilities to rule out, since no matter what possible state the world is in, it is a state in which p is true. Thus, if p is necessary, we don’t require empirical evidence to know p after all, and if p is a posteriori, then p isn’t necessary. Either way, the objector concludes, the necessary a posteriori is an illusion. Kripke’s reply invokes his account of (6b). According to that account, the function of empirical evidence needed for knowledge that Hesperus is Phosphorus is not to rule out possible world-states in which the proposition is false. There are no such states. Rather, evidence is needed to rule out possible states in which we use the sentence (6b) to express something false. Ruling this out involves putting aside our de re beliefs about Venus, and determining whether our justified descriptive beliefs are up to the task. If they fail to rule out the possibility of an epistemic state qualitatively identical to ours in which the names refer to different things, then we can’t rule out the falsity of the sentence we accept, and so, the thought goes, we can’t justify the belief we use the sentence to express. Kripke’s reply to the objection extends this explanation to all instances of the necessary a posteriori.8 He illustrates this extension with an example in which he encounters a table, and comes to know, on the basis of empirical examination, that it is made of wood, not ice. For all he knew prior to the examination, it could have turned out that the table was made of ice. Kripke tells us that this intuition—that it could have turned out that the table was made of ice—is simply the recognition that it is genuinely possible for an agent to be in a situation qualitatively identical to his, prior to the examination, and be facing a table that is made of ice. In pointing at the table and saying ‘This table is not made of ice’, he expresses a necessary truth—since that very table t could not have been made of ice. However, he would not accept, and would not be justified in accepting, the sentence uttered, unless he also believed, and was justified 8

Kripke (1980, 141–43).

The Kripkean Necessary A Posteriori • 175

in believing, the descriptive proposition DP that a unique table over there is not made of ice. It is his justified belief in DP (shared by agents in qualitatively identical states) that rules out possible situations in which his utterance fails to express a truth. DP is, of course, contingent rather than necessary, and hence not to be confused with the (singular) proposition— that t is not made of ice—expressed by the indexical sentence uttered. Still, since Kripke is justified in believing DP only on the basis of empirical evidence, and, since this evidence is required for his utterance to be justified, his justification for accepting the sentence uttered requires empirical evidence. From SDJ, it follows that although it is a necessary truth that t is not made of ice, his knowledge of this truth requires empirical justification, and so is a posteriori. This is Kripke’s second route to the necessary a posteriori. It applies to all his examples—which contain names, natural kind terms, or demonstratives, and semantically express propositions knowledge of which involves de re knowledge of the individuals or kinds those terms designate. The necessity of these propositions is explained by their attribution of essential properties and relations to those individuals or kinds. Their aposteriority is explained—in his first route to the necessary a posteriori—by the fact that the properties and relations can be known to apply to particular individuals and kinds only a posteriori. This explanation, though general, excludes simple identities of the sort (6) and (7). Kripke’s second (attempted) explanation is meant to apply not only to these stragglers, but to the other cases as well. In the second route, knowledge of a necessary proposition p is linked to acceptance of a sentence S used to express p—which in turn is linked to knowledge of a descriptive proposition DP for which empirical evidence is needed. Since justification for accepting S, and believing DP, requires empirical evidence, this evidence is deemed necessary for knowledge of p. The two routes to the necessary a posteriori differ as follows: (i) The first route applies to a proper subset of cases to which the second is meant to apply. (ii) Only the first leads to the recognition of epistemically possible world-states over and above metaphysically possible worldstates. (iii) Only the first takes the empirical evidence needed for a posteriori knowledge of p to rule out epistemic possibilities in which p is false. There is also another important difference. The first route is, as I have indicated, sound. The second is not.

176 • Essay Six

The Unsoundness of Kripke’s Second Route to the Necessary A Posteriori The problem with Kripke’s second route to the necessary a posteriori is that the principle, SDJ, on which it rests, requires an unrealistic degree of transparency in the relationship between sentences and the propositions they express. S1 and S2 may mean the same thing, or express the same proposition p, even though a competent speaker who understands both, and knows of each that it expresses p, does not realize that they express the same proposition. Such an agent may accept S1, and believe it to be true, while refusing to accept S2, or believe it to be true, thereby falsifying SDJ. One such agent is Kripke’s Pierre.9 Although he understands both ‘Londres est jolie’ and ‘London is pretty’, he does not realize that they mean the same thing, and so accepts one while rejecting the other. Since SDJ yields the contradictory result that Pierre both believes and does not believe that London is pretty, it must be rejected, thereby undermining the second route to the necessary a posteriori.10 When applied to Kripke’s examples, SDJ links belief in singular propositions (about individuals or kinds) to acceptance of specific sentences that express them—which, in turn, are linked to belief in certain descriptive propositions related to the original singular propositions. This suggests the possibility of dropping the problematic SDJ, and linking the singular propositions directly to their descriptive counterparts. In the case of (6b) one’s belief that Hesperus is Phosphorus might be linked to (something like) one’s belief that the heavenly body visible in the evening is the heavenly body visible in the morning, while in the case of Kripke’s example about the table, one’s belief that it is not made of ice might be linked to (something like) one’s belief that a unique table over there is not made of ice. The idea, in each case, is that the linked beliefs are related in two ways: (i) one’s coming to have the descriptive belief, in the circumstances in question, is necessary and sufficient for one’s believing the singular proposition, and (ii) one’s justification for believing the singular proposition rests on one’s justification for the descriptive belief. Since in each case, justification of the descriptive belief requires empirical evidence, one’s belief in the putative instance of the necessary a posteriori is taken to require the same evidence. 9

Kripke (1979). The failure of SDJ illustrated by Kripke’s examples is a special instance of a more general point—namely, the failure of A knows that R(a,c) and A knows that R (b,c) to entail  A knows that ∃x (Ra,x & Rb,x) discussed in essay 1, Soames (1987). In the case of SDJ, a and b name sentences, c stands for a proposition, and R relates sentences to what they mean or express. 10

The Kripkean Necessary A Posteriori • 177

The resulting nonmetalinguistic substitute for SDJ is, roughly, the following. The Strong Descriptive Origin and Justification of De Re Belief (SDOJ) If an agent x in a circumstance C is capable of believing a singular proposition p by virtue of believing a certain related descriptive proposition DP, then (a) x believes p in C iff x believes DP in C, and (b) x would be justified in believing p in C on the basis of e iff x would be justified in believing DP in C on the basis of e. Although SDOJ can be used in Kripke’s second route to the necessary a posteriori in essentially the same way as SDJ, it cannot be used to save this route—since the same counterexamples falsify both. In the case of Pierre, a proponent of the idea that belief in singular propositions always arises from belief in associated descriptive propositions must admit that there are several ways that Pierre can come to believe singular propositions about London. He may, for example, come to believe that London is pretty either by believing that the city he lives in is pretty, or by believing that the city on the picture postcards brought from Paris is pretty. SDOJ will then give the results that he believes that London is pretty (i) iff he believes that the city he lives in is pretty and (ii) iff he believes that the city on the picture postcards brought from Paris is pretty. Since, in fact, he believes that the city in the pictures is pretty while failing to believe that the city he lives in is pretty, SDOJ leads to the contradictory conclusion that Pierre both believes and does not believe that London is pretty. Thus, SDJ, SDOJ, and Kripke’s second route to the necessary a posteriori must all be rejected. Fortunately, this rejection does not diminish the correctness of his first route to the necessary a posteriori. The only thing cast into doubt is the aposteriority of (6) and (7).

Mind-Body Identity and the Necessary A Posteriori I now turn to a case study of the philosophical import of a proper understanding of Kripke’s two routes to the necessary a posteriori. The issue is the mind-body problem, which Kripke discusses at the end of Naming and Necessity, where he compares (8) and (9). (8) Heat is mean molecular kinetic energy. (9) Pain is C-fiber stimulation. For simplicity, we treat the nouns as designating kinds (rather than their instances) and ‘is’ as expressing identity. Both sentences are a posteriori, and so may appear contingent. It was an empirical discovery that how hot

178 • Essay Six

something is depends on how fast its molecules are moving. The function of evidence used to show this was to rule out conceivable possibilities in which heat was something else. Thus, if one failed to distinguish epistemic from metaphysical possibility, it might seem that there must be ways that the world genuinely could be in which (8) was false—and, hence that (8) is contingent. Something similar might be said about (9). Kripke argues that the seeming contingency of (8) is an illusion. He takes ‘heat’ and ‘mean molecular kinetic energy’ to be rigid designators, in which case (8) must be necessary, given that it is true. The illusion of contingency is attributed to the fact that we identify heat by the sensations it causes in us. Thus, he thinks, ‘heat’ is associated with “the reference-fixing-description,” ‘the cause of sensation S’. The illusion that (8) is contingent comes from confusing this description with a synonym for ‘heat’, and thereby confusing (8) with (8*). (8*) The cause of sensation S = mean molecular kinetic energy. Of course, ‘heat’ is not really synonymous with ‘the cause of sensation S’—since a world-state in which there are no sentient beings, and hence no cause of sensations, may still be one in which some things are hot. Once the nonequivalence of (8) and (8*) is recognized, the contingency of the latter no longer masks the necessity of the former. Kripke assumes that, like the terms in (8), those in (9) are rigid, and hence that (9) must be necessary, if true. This time, however, he sees no way of dismissing the impression of contingency as an illusion. Unlike ‘heat’, which we use to designate the cause of a certain noticeable sensation, ‘pain’ is used to designate the very sensation we notice. We don’t say to ourselves: “What a horrible sensation! Let’s use ‘pain’ to designate whatever causes it.” Instead, ‘pain’ designates the sensation itself, which we identify directly, without appeal to properties that anything else could have. Because there is no descriptive reference-fixer to confuse with a synonym for ‘pain’, there is no contingent claim to confuse with (9). Since the intuition that (9) is contingent can’t be dismissed as an illusion, Kripke suggests that (9) isn’t necessary, and so isn’t true. Kripke gives this argument as a straightforward application of his second route to the necessary a posteriori. I want to argue that at least the case [of the apparent contingency of (9)] cannot be interpreted as analogous to that of scientific identification of the usual sort, as exemplified by heat and molecular motion. What was the strategy used above to handle the apparent contingency of certain cases of the necessary aposteriori? The strategy was to argue that although the statement itself is necessary, someone could, qualitatively speaking, be in the same epistemic situation as the original, and

The Kripkean Necessary A Posteriori • 179

in such a situation a qualitatively analogous statement could be false. In the case of identities between two rigid designators, the strategy can be approximated by a simpler one: Consider how the references of the designators are determined; if these coincide only contingently, it is this fact which gives the original statement its illusion of contingency. (150) Over the next four pages, Kripke gives the argument involving (8) and (9) summarized above. Though he, cautiously, does not conclude that it demonstrates that (9) is false, he suggests that it constitutes a serious, perhaps insurmountable, obstacle to taking (9) to be true. However, he is wrong about this. The crucial assumption on which the argument is based is that when p is an instance of the necessary a posteriori, any illusion of contingency— any sense that p could have turned out false—results from confusing p with a related, descriptive proposition q that really is contingent. In these cases, p will be expressed by a sentence containing a (rigid) name, demonstrative, or natural kind term, and q will be expressed by a related sentence containing a (nonrigid) description which, in a certain sense, fixes the reference of the rigid designator. The sense of referencefixing involved here is epistemic rather than semantic. For Kripke, most names and natural kind terms do not have their reference fixed, as a matter of linguistic convention, by descriptions associated with them by speakers, and understanding them does not require descriptive, reference-fixing knowledge. Often, however, speakers who use a name, demonstrative, or natural kind term identify the individual or kind designated (on a given occasion) by its possession of certain properties. In these cases, speakers identify the referent by description. For example, I might identify my desk as the reddish desk in my office, supporting my computer. If I do, it will be important to distinguish (10a) from (10b), so as not to allow the contingency of the latter to obscure the necessity of the former. (10) a. This desk (pointing at it) is made of mahogany (if it exists). b. The reddish desk in my office supporting my computer is made of mahogany (if it exists). The point generalizes. Whenever one identifies the referent x of a rigid designator t using a nonrigid description D, there is potential for confusing a world-state in which x is so and so with a world-state in which the thing that satisfies D is so and so. In these cases, the conceivability of (11a) t is so and so (if t exists)

180 • Essay Six

will be a reliable guide to the possibility of x being so and so, only if the claim expressed by (11a) is clearly distinguished from the claim expressed by (11b) D is so and so (if D exists) Applying this lesson to ‘pain’, Kripke observes that although it rigidly designates its referent, we don’t identify the referent via a description that could have designated anything else. Thus, an important source of error in our judgments about the modal profile of ‘heat’ statements is absent from our judgments about ‘pain’ statements. Sometimes, when we take ourselves to be coherently conceiving of a situation in which heat is so and so, what we are in fact conceiving is a situation in which something else—with the descriptive characteristics used to identify heat—is so and so. When this happens, our intuitions of conceivability are inaccurate guides to genuine possibility. If Kripke is right, this doesn’t happen with ‘pain’. Whenever we take ourselves to be conceiving of a situation involving pain, that is what we are conceiving. Presumably, the point extends to similar words for other mental states. So far, so good. However, there is an error to be avoided lurking here which is expressed by the Coherent Conceivability Thesis (CCT). CCT. Apart from confusion about what we are conceiving, coherent conceivability is a reliable guide to genuine (metaphysical) possibility. If we can coherently conceive—without confusion of the sort indicated by (11a, b)—of a world-state in which p is true (false), then there are genuine (metaphysically) possible world-states in which p is true (false). If CCT were true, our ability to coherently conceive of scenarios in which pains are not C-fiber stimulations, or vice versa, would demonstrate that the two kinds are different. However, CCT is inconsistent with Kripke’s first, essentialist, route to the necessary a posteriori. As we have seen, that route is based on the idea that certain properties of objects that they can be known to have only a posteriori, may be known a priori to be essential properties of anything that has them. In any such case, coherent conceivability will not coincide with genuine possibility. If P is such a property of o, then world-states in which o exists without having P will be coherently conceivable, even though they are metaphysically impossible. Thus, Kripke cannot accept CCT. Reasons for rejecting CCT are illustrated by (12). (12) I came from a sperm and egg (if I exist at all). In uttering (12), I identify the referent of ‘I’ directly, without detour through nonrigid descriptions. Hence, no confusion of the sort indicated

The Kripkean Necessary A Posteriori • 181

by (11) threatens. If CCT were correct, this should make for a close connection between conceivability and genuine possibility. But it doesn’t. Though the proposition I use (12) to entertain and assert is (assuming Kripke’s own essentiality of origin thesis) necessary, it is knowable only a posteriori. Putting aside my knowledge of human reproduction, I can coherently conceive of a situation in which I exist without coming from a sperm and egg. Surely, this doesn’t show that it was genuinely possible for me to come into being in some radically different way. For that matter, I can coherently conceive of existing forever, with or without a body, but that doesn’t show that these things really are (metaphysically) possible. Similar remarks apply to my use of (13) to express a necessary a posteriori truth. (13) I am the biological father of Brian Soames (if he exists). Since (13) contains a proper name, a defender of CCT might maintain that—because I use descriptive information to identify Brian—whenever I take myself to be coherently conceiving of him as having a different father what I am really conceiving is not Brian himself having that property, but someone qualitatively similar to Brian. But surely that can’t be right. If it were, it would be hard to see how anyone could ever conceive of anything about a specific individual, or how anyone could ever have de re attitudes. Thus, if a skeptic objects, “How can you be sure that it is Brian in the situation you are imagining, rather than a qualitatively identical duplicate?” the answer is a paraphrase of the one Kripke gives to the skeptic who demands criteria of transworld identification for individuals across metaphysically possible world-states. Don’t ask: how can I identify Brian in another possible world, except by his properties? I have held Brian in my arms, I can point to him, and when I ask whether he might have had a different father, I am talking, by definition, about him. I don’t have to identify him after seeing him through a telescope.11 In short, Kripke’s point about metaphsically possible world-states applies equally to epistemically possible world-states. Thus CCT must be rejected. Once it is, Kripke’s argument against (9), based on its supposedly unfavorable comparison with (8), collapses. We may grant, for the sake of argument, that if they are true, these identities are instances of the necessary a posteriori that fit Kripke’s essentialist paradigm. In each case, a property P that can be known only a posteriori to be possessed by a kind k is claimed to be an essential property of k. In

11

See Kripke (1980, 52–53).

182 • Essay Six

the case of heat, P is the property of having instances with characteristics, involving the motion of molecules, that explain (in the actual world-state) phenomena such as burning and boiling. In the case of pain, P is the property of having instances that are firings of certain neurons. In each case, if P is an essential property of k, then the identity is necessary, even though world-states in which k doesn’t have P are coherently conceivable. This conceivability is itself a potential source of the illusion that the proposition expressed by (9) is contingent—apart from any confusion about identifying, or reference-fixing, descriptions. Since Kripke’s argument does nothing to discredit this explanation of the impression of the contingency of (9), it fails to show that this impression poses any problem for its being necessarily true. Fortunately for Kripke, (9) is likely to be false for independent reasons. If, as he contends, ‘pain’ is a rigid designator, then it designates a kind kp of which any genuinely possible pain of any possible creature—including creatures with physiologies radically different from ours—is an instance. Since these possible pains are likely to outstrip occurrences of the neurological states associated with our actual pains, nothing as parochial as C-fiber stimulation is likely to be identical with kp (though this objection doesn’t rule out that our pain is nothing more than C-fiber stimulation). As for functionalist accounts of the sort offered by David Lewis in “Mad Pain and Martian Pain,” Kripke has a forceful objection.12 According to Lewis, pain (for a given kind k of thing) = the internal physical state which (in normal members of k) plays a certain causal role (which involves arising from, and prompting the avoidance of, injury-causing situations). Lewis is content to deny, on this basis, that ‘pain’ is a rigid designator, since he believes it may designate C-fiber stimulation in the actual world-state, while designating different physical states in other (metaphysically) possible worldstates. Kripke’s observation that being a pain is an essential property of individual pains puts pressure on this view.13 For if Lewis is right, it would seem that a particular C-fiber stimulation might, in the actual world-state, be the headache I had last night, while existing in another possible worldstate, entirely disconnected from the functional role of pain, and so without being a pain in that world-state. Since this is inconsistent with Kripke’s observation—which does have a certain plausibility—his essentialist doctrines do, in the end, generate a worrisome objection to a leading version of the mind-body identity theory. The objection, however, is not the one he belabors in his discussion of (8) and (9).

12 13

Lewis (1980). Kripke (1980, 148–49).

The Kripkean Necessary A Posteriori • 183

Transposing Kripke’s Missteps into the Key of 2D Kripke’s second route to the necessary a posteriori, and his reliance on that route in his discussion of (8) and (9), are rare missteps in a work of monumental achievement. Unfortunately, these missteps have not been without influence. Nowhere has this influence been greater than in guiding the development of fundamentally anti-Kripkean, two-dimensionalist attempts to reinstate descriptivism about names and natural kind terms, to reconnect necessity and apriority to analyticity, and to return philosophy to analytic paradigms of its mid-twentieth-century “golden age.” This influence is evident in David Chalmers’s Kripke-style story of mind-body nonidentity, and in the 2D transformation of Kripke’s flawed, second route to the necessary a posteriori into a dubious semantic theory. Chalmers’s treatment of (8) and (9) is similar to Kripke’s. Whereas Kripke observes that we identify heat as the thing denoted by the nonrigid description ‘the cause of sensation S’, Chalmers takes ‘heat’ to be synonymous with an indexical, rigidified description (which, to simplify the comparison, I will take to be ‘dthat [the cause of sensation S]’). Since ‘heat’ is seen as indexical, (8) is associated with both a primary intension (proposition)—expressed by (8*)—stating the conditions a context must satisfy in order for (8) to express a truth, and a secondary intension (proposition)—that k = k—which S expresses in our present context. For Chalmers, as for Kripke, the illusion that (8) is contingent comes from confusing this necessary proposition, with the contingent proposition expressed by (8*). Kripke and Chalmers also agree that the referent of ‘pain’ is neither fixed nor identified by any description D, and hence that ‘pain’ is not synonymous with any rigidified description. For Kripke, this means that our intuition of contingency is not due to confusing (9) with any other proposition. For Chalmers it means that since the primary and secondary intensions of (9) are identical, the intuition that (9) is contingent doesn’t result from confusing the two. Both men seem to agree that in the absence of this sort of confusion, the conceivability of the distinctness of pain and C-fiber stimulation is a reliable guide to the genuine possibility that they are distinct, and hence to their nonidentity. Since Chalmers’s version of the argument involving ‘pain’ (and other sensation words) is a 2D version of Kripke’s, it is not surprising that it inherits all the difficulties of Kripke’s argument and more. Putting these extra difficulties aside, I focus here on the rejection, by two-dimensionalists, of Kripke’s sound, essentialist route to the necessary a posteriori in favor of his

184 • Essay Six

unsound, second route. The lesson of Kripke’s essentialist route is that there is the gap between coherent conceivability and genuine possibility marked by the existence of epistemically possible world-states that are metaphysically impossible. The presumption that there is no such gap, and there are no such world-states, is the unargued starting point for the 2D systems of both Chalmers and Frank Jackson.14 Given this starting point, they insist, quite understandably, that the necessary a posteriori is a feature of certain sentences, not propositions. On their view, there can’t be any proposition p that is both necessary and knowable only a posteriori, since if there were, there would be an epistemically possible world-state—which could not be known a priori not to obtain—in which p was false, even though this world-state was metaphysically impossible. Since this is ruled out from the beginning, instances of the necessary a posteriori are taken to be sentences S that are associated with pairs of propositions—primary and secondary intensions—one of which is contingent and relevant to the epistemic status of S, the other of which is necessary and relevant to S’s modal status. In this way, the Byzantine formal apparatus of contemporary two-dimensionalism gets its impetus from ignoring, or misunderstanding, Kripke’s essentialist route to the necessary a posteriori.15 In its place, a version of Kripke’s secondary route is embraced. Since this version requires instances of the necessary a posteriori to be indexical (so that they may have different primary and secondary intensions), it requires names and natural kind terms to be semantically equivalent to rigidified descriptions (which do not themselves contain any such terms). As a result, the primary intension of a Kripkean instance S of the necessary a posteriori will be a contingent, descriptive proposition the truth of which rules out metaphysically possible world-states in which (i) the speaker x, or a similar agent, is in an epistemic state qualitatively identical to the one x is actually in, but (ii) S (understood with the meaning it actually has) fails to express a truth. This is what Kripke’s descriptive proposition DP, arising from SDJ or SDOJ, becomes when his second route to the necessary a posteriori is built into 2D semantics. The idea is that when S is used to state an instance of the necessary a posteriori, the (secondary) proposition p expressed by S is necessary, but empirical evidence is needed, not to rule out the possibility that p is false, but to rule 14 Chalmers (1996, 66–67); Jackson (1998, 69–70, 74). For exposition and criticism, see Soames (2005, 152–75, 198–99). 15 The only place in Chalmers’s and Jackson’s two books that I know of which purports to give an argument for the claim that all epistemic world-states are metaphysically possible, rather than simply adopting it by fiat, is the section on “strong metaphysical necessity” in Chalmers (1996, 136–38). Even there, the argument simply ignores, and is refuted by, the passage from “Identity and Necessity” (Kripke 1971, 152–53) cited above in note 6. For discussion, see Soames (2005, 202–9).

The Kripkean Necessary A Posteriori • 185

out the possibility that the utterance expresses something false (because S’s primary proposition is false). The futility of this approach can be illustrated by elaborating our earlier discussion of the principle CCT, suggested by Kripke’s treatment of (8) and (9). There, we considered examples of the form (14), where ‘F’ stands in for predicates expressing essential properties of mine. (14) I am F (if I exist at all). Plausible substitution instances of ‘F’ include: ‘a human being’, ‘a biological father of Brian Soames’, ‘an organism that came from the union of a sperm and egg’, and ‘a being with a body made of molecules’. Since these plausibly express essential properties, instances of (14) are plausibly taken to be necessary, if true. However, since knowing that I have these properties requires empirical knowledge, these instances are knowable only a posteriori. The question is, what is the best account of this? According to Kripke’s essentialist route to the necessary a posteriori, what makes an instance of (14) a posteriori is that the proposition it expresses—which attributes to me the property of being one who “is F if one exists”—requires empirical evidence to rule out coherently conceivable possibilities in which it is false. Since the world-states to be ruled out are ones in which I exist while lacking certain essential properties, they are metaphysically impossible. This is precisely what Chalmers and Jackson deny. According to them, there are no such world-states, so no evidence is needed to rule them out. Rather, empirical evidence is needed to rule out the possibility that the context of my utterance is one in which the sentence is used to express something false. After all, instances of (14) could express falsehoods—if they were uttered by a nonhumans who didn’t arise from unions of a sperm and an egg, and who didn’t have bodies made of molecules. Are there metaphysically possible beings of that sort who are capable of uttering instances of (14)? Although there is room to doubt that there are, let’s suppose, for the sake of argument, that such beings really are possible. On this supposition, there are metaphysically possible contexts in which these beings would express falsehoods, if they were to utter instances of (14). But what does that have to do with the aposteriority of the knowledge I express by an instance of (14)? Well, the two-dimensionalist must argue, in order to know that my utterance expresses a truth, I must rely on empirical evidence to rule out the possibility that my context is one of those in which the agent is a nonhuman who didn’t come from a sperm and egg, and doesn’t have a body made of molecules. But, surely, for empirical evidence to be required to rule out the possibility that my context is one of those is just for empirical evidence to be required to rule out that I am a nonhuman, that I didn’t come from a sperm and egg, and

186 • Essay Six

that I don’t have a body made up of molecules. After all, I know who the agent of my context is; it is not as though I am confused about which member of a class of numerically distinct but qualitatively identical beings uttered the sentence, and need empirical evidence to track him down. But then, the things I need empirical evidence to rule out are just the epistemically possible, but metaphysically impossible, world-states recognized by the essentialist route to the necessary a posteriori. In short, the twodimensionalist has no plausible alternative but to accept the very thing that his elaborate system was designed to avoid. He must grant that there are world-states that are epistemically possible—in the sense of being coherently conceivable, and not knowable a priori not to be instantiated— which are not really—i.e., metaphysically—possible.

The Importance of Nonlinguistic Modalities In my opinion, none of Kripke’s many achievements is more important than his breaking the spell of the linguistic as the source of philosophically important modalities. In other work, I have tried to identify significant arguments of leading figures in the twentieth century that come to grief over the implicit identification of the necessary and the a priori with the analytic.16 However, there is more at stake than a collection of particular arguments. As long as these modalities are seen as varieties of truth in virtue of meaning, while meaning itself is viewed as essentially transparent to competent speakers, there will be no credible alternative to the old, confining orthodoxy of philosophy as linguistic analysis. The problem with this orthodoxy is well-illustrated by the metaphilosophical position of the later Wittgenstein. If all of philosophy is the analysis of meaning, and meaning is fundamentally transparent to competent speakers, then there is little room for philosophically significant explanations and theories. For surely such explanations and theories will be necessary, or a priori, or both. If this renders them analytic, they must either be capable of being seen as trivially transparent by competent speakers, or derivable from trivially transparent truths by trivially transparent steps. But they aren’t. Anyone immersed in the work of a philosopher of any note—including those, like the later Wittgenstein, who espoused this metaphilosophical position—should be able to see that their most interesting philosophical positions do not fit this restrictive model.17 Hence, we need a more expansive model that fits what good philosophers really 16 17

Soames (2003). See, e.g., Soames (2003, vol. 2, chaps. 1–4).

The Kripkean Necessary A Posteriori • 187

do. We will never have it, if we identify the necessary and the a priori with the analytic. This is why it is so important to recognize the lesson of Kripke’s essentialist route to the necessary a posteriori, and why it is worth defending against the revival of linguistically based accounts of these modalities reflected in the following passage from Chalmers.18 If we make the equation [in which primary and secondary intension are taken to be two different aspects of meaning] both of these intensions will back a certain kind of conceptual truth, or truth in virtue of meaning. The primary intension backs a priori truths. . . . Such a statement will be true no matter how the actual world turns out [i.e., in any possible context of utterance], although it need not hold in all nonactual possible worlds. The secondary intension does not back a priori truths, but backs truths that hold in all counterfactual possible worlds. . . . Both varieties qualify as truths in virtue of meaning; they are simply true in virtue of different aspects of meaning. There is, of course, more to be done in constructing a defensible metaphilosophical view than defending, and drawing out the consequences of, Kripkean essentialism. It must also be admitted that Kripke was far more successful in illuminating the nature of necessity, and distinguishing it from both apriority and analyticity, than he was in illuminating the nature of apriority, and distinguishing it from analyticity. But the fact that work remains to be done does not make what has been accomplished any less precious.19

References Chalmers, David. 1996. The Conscious Mind. New York: Oxford University Press. Jackson, Frank. 1998. From Metaphysics to Ethics: A Defence of Conceptual Analysis. Oxford: Clarendon Press. Kripke, Saul A. 1971. “Identity and Necessity.” In Identity and Individuation, ed. Milton K. Munitz, 135–64. New York: New York University Press. ———. 1979. “A Puzzle about Belief.” In Meaning and Use: Papers Presented at the Second Jerusalem Philosophical Encounter, April 1976, ed. Avishai Margalit, 239–83. Dordrecht: Reidel. ———. 1980. Naming and Necessity. Cambridge: Harvard University Press. Originally published in Semantics of Natural Language, ed. Donald Davidson and Gilbert Harman, 253–355 (Boston: Reidel, 1972). 18 19

Chalmers (1996, 62; my emphasis). Thanks to Ali Kazmi and Jeff Speaks for comments on earlier drafts.

188 • Essay Six Lewis, David. 1980. “Mad Pain and Martian Pain.” In Readings in Philosophy of Psychology, ed. Ned J. Block, vol. 1, 216–22. Cambridge: Harvard University Press. Soames, Scott. 1987. “Direct Reference, Propositional Attitudes, and Semantic Content.” Philosophical Topics 15:47–87. ———. 2002. Beyond Rigidity: The Unfinished Semantic Agenda of “Naming and Necessity.” New York: Oxford University Press. ———. 2003. Philosophical Analysis in the Twentieth Century. 2 vols. Princeton: Princeton University Press. ———. 2005. Reference and Description: The Case against Two-Dimensionalism. Princeton: Princeton University Press. ———. n.d. “Kripke on Epistemic and Metaphysical Possibility: Two Routes to the Necessary Aposteriori.” In Saul Kripke, ed. Alan Berger. Cambridge: Cambridge University Press, forthcoming.

ESSAY SEVEN

Knowledge of Manifest Natural Kinds

Manifest kinds are natural kinds designated by terms like water, tiger, gold, green, and electricity. Individual instances of these kinds are objects of our potential acquaintance about which we may have de re knowledge. Natural kinds of a more highly theoretical sort—like photons and neutrons—are not included in this category. Manifest natural kinds, or manifest kinds for short, figure in interesting statements of theoretical identification, many of which are both necessary and knowable only a posteriori. The aim of this essay is to explain why this is so. The statements I will be concerned with are expressed by sentences in which a manifest kind term combines with the copula to form a predicate applying to each instance of the kind. Examples of such predicates are is water, is a tiger, is gold, is green, is electricity, and is hotter than. Theoretical identification sentences containing such terms that express necessary a posteriori truths include the sentences in (1).1 (1) a. For all x, x is water iff x is H2O. (Water is H2O) b. For all x, if x is ice, then x is H2O. (Ice is H2O) c. For all x, if x is lightning, then x is electricity. (Lightning is electricity) d. For all material objects x, x is gold iff it is made up of the element with atomic number 79. (Gold is Au) e. For all x, if x is a tiger, then x is an animal. (Tigers are animals) f. For all x and y, x is hotter than y iff the mean molecular kinetic energy of x is greater than that of y. g. For all objects x, x is green iff x has surface spectral reflectance property SSRgreen—the property of reflecting substantially more light in the middle-wavelength part of the visible spectrum than in the long-wavelength part, and approximately the same amount of light in the shortwavelength part as in the nonshort part. 1 The parenthesized sentences—Water is H2O, Ice is H2O, Lightning is electricity, Tigers are animals, and Gold is Au have natural readings in which they are understood along the lines of the sentences preceding them.

190 • Essay Seven

In discussing these examples, I intend to focus on their most important features, and to sidestep certain controversies surrounding them. For example, some philosophers believe that in order for (1a) and (1b) to be true, is in the predicate is H2O must be understood as what is sometimes called “the is of constitution,” rather than as the normal copula. Their thought is that instances of water are never instances of H2O; rather, instances of water are constituted by, or made up of, numerically distinct instances of H2O. For my own part, I don’t think there is any genuine contrast here, since I believe that just as a chunk of ice can itself be an instance of the kind ice, while also being constituted by a numerically distinct instance of ice that is capable of surviving that chunk’s destruction, so the chunk of ice may be an instance of the kind H2O, while also being constituted by a numerically distinct instance of H2O that is capable of surviving the melting of that ice.2 If this is right, then instances of ice and water not only are constituted by H2O, they are instances of H2O themselves. However, this point of controversy will not matter for my topic. Whichever position involving constitution and identity proves to be correct, all these sentences are examples of the necessary a posteriori. Moreover, the semantic nature of the predicates they contain plays an important role in explaining why this is so.

The Positive Account Let us first examine the way in which the necessity of these sentences is related to their truth. As I see it, the crucial issue involves the nondescriptionality of simple manifest kind terms, and the way in which their reference is determined. What Kripke says about the general term cat is the model for a great many terms for manifest natural kinds. He says, “The original concept of cat is: that kind of thing, where the kind can be identified by paradigmatic instances. It is not something picked out by any qualitative dictionary definition.”3 Although this may be a little cryptic, the point is clear. Just as ordinary proper names are standardly introduced by stipulating that they are to apply to certain objects with which one is already acquainted, so general terms like gold, water, tiger, and green are standardly introduced with the intention that they are to designate certain manifest kinds with which we are already acquainted through their paradigmatic instances. For example, we may imagine these terms introduced by the following stipulations:4 2

See Soames (2002, chap. 11). Kripke (1980, 122). 4 Here I imagine the adjective green and the common nouns gold, water, and tiger being introduced with stipulations about what they are to refer to, or designate. The associated 3

Knowledge of Manifest Natural Kinds • 191

The general term green is to designate the color of all, or nearly all, of certain paradigmatic green-samples (and none, or nearly none, of other paradigmatic nongreen-samples)—i.e., it is to designate the characteristic of object surfaces that is causally responsible for the fact that the paradigmatic green-samples appear the same way to us (and different from the paradigmatic nongreen-samples). Hence, the predicate is green will apply (with respect to any world-state) to all and only those objects the surfaces of which have that characteristic which, in the actual state of the world, causally explains why the paradigmatic green-samples look the same to us (and look different from the paradigmatic nongreen-samples.)5 The general term gold is to designate the unique substance of which all, or nearly all, members of the class of paradigmatic gold-samples are instances (and of which none, or nearly none, of the class of paradigmatic nongold-samples are instances). Substances are understood to be physically constitutive kinds—i.e., kinds instances of which share the same basic physical constitution. Hence, the predicate is gold will apply (with respect to any world-state) to all and only those objects that share the basic physical constitution that nearly all the paradigmatic gold-samples actually have (and that none, or nearly none, of the paradigmatic nongold-samples actually have). The general term water is to designate the kind instances of which share with all, or nearly all, members of the class of paradigmatic watersamples those properties that “make them what they are” (and that distinguish them from certain paradigmatic nonwater-samples). These are properties that explain their most salient characteristics—e.g., the

predicates—is green, is gold, is water, and is a tiger are then understood as applying to all and only instances of the kinds designated by the adjectives, or common nouns they contain. (For a more detailed account of the distinction between kind terms and the predicates formed by combining them with the copula see (i) my reply to Bernard Linsky in Soames [2006b], and (ii) my reply to Gómez-Torrente in Soames [2004].) Strictly speaking, the designata of the predicates, their extensions, are the sets of objects they apply to, rather than the kinds designated by the adjectives or nouns they contain. However, since the extension of one of these predicates (at a world-state) will just be the set of instances of the kind designated (at the world-state) by the adjective or noun it contains, the kind plays a crucial role in the semantics of the predicate. More on this later. Until then it will do no harm to sometimes speak loosely of the kinds designated by these predicates. 5 It will be noticed that I allow the stipulation introducing the term green to make reference both to a set of paradigmatic positive instances (the items in the green-sample) and to a set of paradigmatic negative instances (the items in the nongreen-sample). As indicated in Soames (1999, chap. 7), this seems natural with color terms. How far this point extends to other terms, like gold, water, and tiger, is an open question. For that reason, I have formulated the stipulations introducing those terms so as to be compatible with this possibility. Thanks to David Manley for bringing this to my attention.

192 • Essay Seven

fact that they boil and freeze at certain temperatures, that they are clear, potable, necessary to life, and so on. Hence, the predicate is water will apply (with respect to any world-state) to all and only those quantities of matter that have the properties that actually explain the salient features of all, or nearly all, of the paradigmatic water-samples (and that are lacking in all, or nearly all, the paradigmatic nonwatersamples).6 The general term tiger is to designate the species of animal of which all, or nearly all, the members of the class of paradigmatic tigersamples are instances (and of which none, or nearly none, of the members of the class of paradigmatic nontiger-samples are instances). Hence, the predicate is a tiger will apply (with respect to any world-state) to all and only those individuals that are members of the species of which nearly all the paradigmatic tiger-samples are actually members (and of which none, or nearly none, of the paradigmatic nontiger-samples are actually members). These stipulations are, of course, idealized. The manifest natural kind terms could have been introduced in this way, and they behave pretty much as they would if they had been so introduced. However, they need not have been introduced by any formal stipulation. It is enough if at some point speakers started calling relevant things green, gold, water, and tigers, with the intention that the predicates were to apply not only to the particular objects speakers happened to encounter, but also to all and only instances of the relevant kinds to which those objects actually belonged. An analogous point holds for proper names. Although formal baptisms are common, there are also cases in which a proper name is introduced more informally, as when people start calling a certain body of water Green Lake and the habit catches on. In all of these cases, both the formal and the informal, we may speak of a manifest natural kind term or a proper name as being introduced ostensively. When a manifest kind predicate P is introduced ostensively, and Q is another predicate related to P in a certain ways, there are often linguistic explanations of why the corresponding sentences with the logical forms (2a) and (2b) are necessary if true. (2) a. ∀x (Px ⊃ Qx) b. ∀x (Px ≡ Qx) 6 This represents a slight change from the discussion of water in Soames (2002). There I treated it as governed by a stipulation analogous to the one for gold given above, rather than one mentioning the causal explanation of salient properties. See, however, pages 285–86 for a brief discussion of the possibility of treating water as standing for an explanatory kind of the sort indicated here.

Knowledge of Manifest Natural Kinds • 193

I will consider two types of example. Case 1 is illustrated by the predicates is gold and is a tiger; case 2 is illustrated by is water and is green. In case 1, we let P be the ostensively introduced predicate is gold or is a tiger, and we let Q be a natural kind predicate that “designates” a kind of the same type as the kind “designated” by P—a substance, i.e., a physically constitutive kind, in the case of is gold, and a species of animal in the case of is a tiger. In speaking of the kind “designated” by a predicate, I mean to include two types of cases—those in which the kind is the semantic content of the general term from which the predicate is formed by adding the copula, and those in which the semantic content of the general term is property that determines the kind. Later I will say more about which predicates fall into which class. For now, suffice it to say that in both sorts of cases the extension, at an arbitrary world-state w, of a predicate that “designates” a kind k is the set of instances of k at w. We now go through the reasoning, for case 1, of the necessity of true sentences of the forms (2a) and (2b): (i) From the assumption that the ostensive manifest kind predicate P has successfully been introduced it follows that there is a unique natural kind kP of a given type T—a substance in the case of is gold, a species of animal in the case of is a tiger—of which nearly all members of the P-sample are instances (and nearly no members of the non-P-sample are instances), and P applies, with respect to a world-state w, to all and only instances of kP at w. (ii) From the assumption that the natural kind predicate Q also “designates” a natural kind, it follows that there is a kind kQ which is such that Q applies with respect to a world-state w to all and only members of kQ at w. (iii) By hypothesis, the two predicates designate kinds of the same type; in the case of is gold, kP and kQ are both substances, while in the case of is a tiger they are both species of animal. (iv) Now suppose that (2a) is true. Since nearly all objects in the P-sample are P’s, nearly all of the objects in the P-sample are also Q’s, and hence instances of kind kQ as well as kind kP (v) But, by hypothesis, there is a single kind of the given type T—a single substance or species—of which nearly all members of the P-sample are instances (and nearly no members of the nonP-sample are instances). Assuming that we find that nearly no members of the non-P-sample are Q’s, we may conclude that kind kP = kind kQ. (vi) This means that in addition to (2a), (2b) must also be true. (vii) Moreover, both must be necessary, since the extension of P with respect to a world-state w = the set of instances of kP with respect to w = the set of instances of kQ with respect to w = the extension of Q with respect to w. Since sentence (1d) (about gold) is of the form (2b), it is necessary. (1e), which is about tigers and is of the form (2a), is slightly different. Here the term tiger is introduced with the stipulation that it is to designate whatever species of animal is the one of which paradigmatic members of the tiger-sample are instances (and of which paradigmatic members of

194 • Essay Seven

the nontiger-sample are not). Thus, if the term is successfully introduced— in accord with the presupposition that there is one and only one species of animal satisfying this condition—then it will follow that the sentence Tigers are animals is true. It will even follow that it is a necessary truth, provided that it is an essential property of any species of animal that instances of it are themselves required to be animals. Given this, we conclude that (1e) is also necessary. Next consider case 2, which is illustrated by the predicates is water and is green. As I have imagined their ostensive introduction, both stand for what might be called explanatory kinds. In the case of is water, the kind is one that is determined by the properties possessed by paradigmatic water-samples that both distinguish them from paradigmatic nonwater-samples and (causally) explain, in the actual state of the world, such salient characteristics of the members of the water-sample as their boiling point, freezing point, their properties as solvents, and so on. In the case of is green, the kind is one that is determined by the properties possessed by the surfaces of objects among the paradigmatic greensamples that distinguish them from paradigmatic nongreen-samples and that actually explain why the members of the green-sample look alike to us, and different from members of the nongreen-sample. When P is a simple manifest kind predicate, like is water or is green, which stands for an explanatory kind, the necessity of (2a) and (2b) is accounted for a little differently than before. (i) We begin with the ostensive introduction of P by a stipulation that it is to apply (with respect to any possible state of the world) to all and only instances of the kind determined by certain properties—namely those possession of which by all, or nearly all, members of the P-sample in the actual state of the world distinguishes them from members of the non-P-sample, and (causally) explains the salient characteristics of the P-sample. (ii) It is then discovered scientifically that possession of the property expressed by Q distinguishes the members of the P-sample from the members of the non-P-sample, and (causally) explains the salient characteristics of the P-sample. (iii) From this it follows that the kind designated by the simple manifest kind predicate P is the kind determined by the property expressed by Q. (iv) This is sufficient to establish the necessity of sentences like (1a) and (1g) (about water and green objects respectively), which have the logical form (2b). Sentence (1f) (about heat and kinetic energy) is essentially the same, except for containing the two-place explanatory kind term is hotter than.7 The explanation of the necessity of (1b), which is of the form (2a), is derivative from that of (1a).8 (1c) (about electricity) is analogous to (1b). 7 8

See Soames (2002, 275). See Soames (2002, 294–97).

Knowledge of Manifest Natural Kinds • 195

In giving these linguistic explanations, I have appealed to the manifest natural kinds “designated” by different predicates without saying anything about what these kinds are. Although a number of fundamental questions about their metaphysical natures can be left open, some features of kinds are central to the linguistic model I have sketched. For one thing, the model presupposes that, whatever they are, manifest kinds are things that exist and have instances in different possible world-states. The color green, though not a green object itself, has green objects as instances. Moreover, since different objects are green in different possible world-states, the color green has different instances in different worldstates. The color remains the same from state to state, even if all its instances vary. The same is true of the substances gold and water, as well as the species tiger. In giving the model, I have also presupposed that manifest kinds are rather coarse grained, and are individuated by their possible instances in an interesting sense. If manifest kinds a and b have precisely the same instances in all possible world-states, then the kind a is the kind b; alternatively put, if two kinds are different, then they differ in at least some possible instances. Intuitively this seems plausible. It is hard to imagine two distinct species of animal, two distinct substances, or two distinct colors which have precisely the same instances in every possible world-state. The reason this is important for my linguistic model is that it is important that kinds not be individuated as finely as the properties that determine them. Consider, for example, the color green. Physicalists about color tell us that the object-color green is determined by a certain type of surface spectral reflectance property—one which specifies proportions of light reflected at different wavelengths.9 Let Q be a complex predicate of English explicitly mentioning proportions of light reflected at different wavelengths that expresses this property. The predicate is green is clearly not synonymous with Q. The same can be said for other descriptive predicates. Suppose there is a further complex predicate Q' that applies to surfaces on the basis of a specification of their minute physical structure, which turns out to be necessarily equivalent to Q. Then, although the predicate is green is necessarily equivalent to both Q and Q', it is synonymous with neither. The different complex properties expressed by these predicates both have equal claim to determining the natural kind green, but neither is identical with the kind—that is, the color—itself. What is said here about the predicate is green, and the color it designates, applies to linguistically simple manifest kind predicates generally, and the kinds they designate.10 9 10

See Byrne and Hilbert (1997). See Soames (2002, 278–79).

196 • Essay Seven

The picture that emerges is one in which the meanings, or semantic contents, of linguistically simple manifest kind predicates like is gold, is water, is a tiger, and is green are complexes , consisting of the meaning of the copula (roughly the relation of being an instance of) and the kind designated by the manifest kind term from which the predicate is constructed. Once this is seen, talk of the kind designated by such a predicate may be viewed as a loose approximation of something more precise. Normally in semantics, the designation of an expression is taken to be its extension, relative to a context and circumstance of evaluation. Since the extension of a predicate is the set of things to which it truly applies, simple manifest kind predicates don’t, in this strict sense, designate kinds, though the simple manifest kind terms they contain do. Since, these terms are directly referential, the kinds designated are also their semantic contents. Thus, a simple predicate is TS containing such a term semantically expresses, as its meaning, the complex indicated above, in which K is designated by TS. What the predicate designates (its extension), at any possible world-state, is the set of instances of K at that state. By contrast, a manifest kind predicate is (a /the) TC containing a complex general term TC semantically expresses , where D is the complex semantic content of TC. When TC is nonrigid, like substance that falls from the sky and fills the lakes and rivers, different kinds are determined by D (instances of which are in the extension of the predicate) at different world-states. When TC is rigid, like substance molecules of which have two hydrogen atoms and one oxygen atom, the same kind is determined at every world-state. However, even when D determines the same kind in every world-state, it is not identical with that kind. Thus, when P is a simple manifest kind predicate the semantic content of which is , and Q is a semantically complex predicate the semantic content of which is , and D determines K with respect to every world-state, the sentences (2a) and (2b), though necessary, do not express the very same propositions as (3a) and (3b). (3) a. ∀x (Px ⊃ Px) b. ∀x (Px ≡ Px) This point is crucial to blocking what would otherwise be a troubling argument that a number of the truths in (1) are not only necessary, but also knowable a priori. However, blocking this argument is one thing, establishing that these truths really are knowable only a posteriori is another. Example (1e), Tigers are animals, is particularly useful in illustrating the problem. Since I imagined the term tiger being introduced by a stipulation that it designate the animal species of which all or nearly all paradigmatic samples are instances, it may seem that it is part of the meaning of the predicate is a tiger that the things it applies to are animals, and,

Knowledge of Manifest Natural Kinds • 197

hence, that it is knowable a priori that tigers are animals. However, this, I believe, is a mistake. In order to see that it is a mistake, it is important to resist the temptation of an all-too-common line of argument. The tempting line goes something like this: (i) To introduce a name or natural kind term n by stipulating that it is to stand for that which satisfies a certain condition, is to use a description D expressing that condition to semantically fix the reference of n. (ii) When D semantically fixes the reference of n, competent speakers associate D with n and know that the semantic rules governing n guarantee that it refers to whatever, if anything, satisfies D. (iii) Because of this, competent speakers know a priori that which is expressed by n is D (if there is a unique thing that is D) when n is a proper name, and For all x, x is an n iff x is an instance of the kind D (if there is such a thing as the kind D) when n is a general term. Here, we may let (ii) serve as a definition of what it is for a description to semantically fix the reference of a simple proper name or natural kind term. The idea it expresses is, essentially, the one behind Kripke’s weak, fix-the-referent, version of the descriptivism defined by the first five theses listed at the beginning of lecture 2 of Naming and Necessity. Examples of expressions in the semantic literature that accord with the definition are descriptions rigidified using David Kaplan’s dthat operator.11 Although the semantic content of dthat D is simply the denotation, if any, of D, in order to understand the rigidified description, one must know that its referent is, by definition, whatever satisfies D. The same is true of names and natural kind terms that have their reference fixed by descriptions in the sense of (ii). Given this understanding of what it is for a description to semantically fix the referent of a term, we can isolate two mistakes in the reasoning from (i) to (iii). (i) is in error because it is possible to use a description as a tool to introduce a name or natural kind term without the description semantically fixing the referent of the term, and hence becoming part of what a competent speaker must master in order to understand it. For example, when looking at my first-born son and naming him Greg Soames, his mother and I did not intend the name to have the force of any dthat rigidified description incorporating the content we used in singling him out. Although our stipulation relied on descriptive information to initially endow the name with meaning, that information was not incorporated 11

Kaplan (1989).

198 • Essay Seven

into either the content of the name, or the conditions required to understand it. (iii) is in error because—whether or not a description used to introduce a term also semantically fixes its reference—the description associated with the term does not give rise to a priori knowledge. In stipulating that Greg Soames was to be the name of our son, his mother and I did not thereby come to know a priori that if we had a son, he was Greg Soames. On the contrary, it was because we were already acquainted with him, and knew him to be our son, that we were able to name him with our stipulation. The knowledge we relied on was, of course, a posteriori. After the name had been introduced, it was possible for us to express our knowledge in a new way—with the sentence Greg Soames is our son, and even more cautiously with the sentence If we have a son, Greg Soames is our son. But the knowledge expressed was just the old a posteriori knowledge we had before the name was introduced. The same can be said for all cases in which one uses a description to introduce a name, since, as I have argued elsewhere, it is a necessary condition on all such introductions that one believe, of the object to be named, that it is the denotation of the description used to introduce it.12 Applying these lessons to manifest kind terms leads one to several significant conclusions. First, although my imagined, idealized stipulations introducing green, gold, water, and tiger make use of descriptions of kinds in terms of certain paradigmatic instances (and noninstances), typically the descriptions do not semantically fix the reference of these terms. The descriptions usually don’t enter into their meanings, and competence with the terms, and the predicates containing them, doesn’t require speakers to understand anything about the paradigmatic instances (or noninstances) used to introduce them. Second, even those introducing the terms green, gold, water, and tiger don’t know a priori of the particular objects mentioned in the introductions that they are green, gold, water, or tigers—any more than Greg’s mother and I knew a priori that Greg Soames was our son. We did know, in virtue of knowing of our own stipulation, that he was named Greg Soames (if he indeed was our son). However, even that metalinguistic knowledge was not a priori, resting as it did on our knowledge of the empirical facts that endowed the name with its meaning. By the same token, one who introduces the general term gold with the stipulation imagined thereby knows the metalinguistic truth that nearly all of the objects mentioned in the stipulation are ones to which the predicate is gold applies (if nearly all of them share the same physical constitution). However, such a person does not know a priori that nearly all those sample objects are gold, since that would require knowing that they are of the same physically constitutive kind, and that is something 12

See Soames (2003, vol. 2, chap. 16).

Knowledge of Manifest Natural Kinds • 199

one can know only a posteriori.13 Is anything weaker known a priori? Does such a person know a priori that if nearly all of those sample objects are of the same physically constitutive kind, then they are gold? To ask this is to ask whether such a person knows a priori, of the kind gold, that nearly all those sample objects are instances of it, if nearly all of them are instances of any physically constitutive kind at all. In answering this question one must distinguish two related claims. (4) a. The gold-stipulator knows a priori that if there is a physically constitutive k of which nearly all gold-samples are instances, then nearly all those samples are instances of k. b. If there is a physically constitutive kind k of which nearly all gold-samples are instances, then the gold-stipulator knows a priori that nearly all those samples are instances of k, if they are instances of any physically constitutive kind. Although (4a) is trivially true, it has nothing to do with the semantics of reference-determination for natural kind terms. Although (4b) is relevant to the semantics of reference-determination, it is not true. However, seeing this takes a little work. There are two main cases to consider. In the first case, the stipulator already knows a posteriori that nearly all the samples are instances of the same physically constitutive kind, even though he is not able to describe the kind in any very informative way except by reference to the samples themselves.14 It is, I think, reasonable to regard this as a case in which the 13 Here I disagree with Kripke. See, for example, Kripke (1980, 135), where he says the following: “the present view asserts, in the case of species terms as in that of proper names, that one should bear in mind the contrast between the a priori but perhaps contingent properties carried with a term, given by the way its reference was fixed, and the analytic (and hence necessary) properties a term may carry, given by its meaning. For species, as for proper names, the way the reference of a term is fixed should not be regarded as a synonym for the term. . . . If we imagine a hypothetical (admittedly somewhat artificial) baptism of the substance [gold], we must imagine it picked out as by some such ‘definition’ as, ‘Gold is the substance instantiated by the items over there, or at any rate, but almost all of them’. Several features of this baptism are worthy of note. First, the identity in the ‘definition’ does not express a (completely) necessary truth: though each of these items is, indeed, essentially (necessarily) gold, gold might have existed even if the items did not. The definition does, however, express an a priori truth, in the same sense as (and with the same qualifications applied as) ‘1 meter = length of S’: it fixes a reference.” For a critique of Kripke’s discussion of the a priori in connection with the meter stick see Soames (2003, vol. 2, chap. 16). As to the parenthetically mentioned “qualifications” mentioned in the passage just quoted, these are dealt with below. 14 The corresponding point for green and for tiger—namely that speakers may already be presumed to know, prior to introducing the term, that (nearly) all green-samples share some characteristic feature of their surfaces that explains their appearance and that (nearly) all tiger-samples are members of the same animal species—is quite plausible. Whether or not the same might be said for gold may be more controversial.

200 • Essay Seven

agent is acquainted with the kind by virtue of being acquainted with some of its instances, and knowing of them that they are instances of a single physically constitutive kind. In this case, the agent’s knowledge of the kind gold is a posteriori, and remains a posteriori when he introduces the predicate is gold to apply instances of it. The case is analogous to one in which I see one and only one man standing in front of me, and I think to myself, He is standing in front of me. My knowledge, of the man in question, that he is standing in front of me is based on, and justified by, my perceptual experience. Hence it is a posteriori. This fact would not change if I were to introduce the name Saul with the stipulation that it is to refer to the man standing in front of me. If I were to do that, the sentence Saul is standing in front of me would not express a proposition that I knew a priori. It would simply express a proposition that I already knew a posteriori, and that can be known only in that way. The case of the gold-stipulator who already knows that the items in his sample are instances of a single physically constitutive kind is similar. He already knows of the kind that his samples are instances of it. This knowledge is a posteriori, and remains so even after he has introduced the term gold to designate it. But what about the knowledge mentioned in (4b)—knowledge of the kind gold that nearly all of the stipulator’s paradigmatic samples are instances of it, if they are instances of any one physically constitutive kind at all? Isn’t that something that the gold-stipulator knows a priori? No, it isn’t. Think again about the man standing in front of me, of whom I know a posteriori that he is standing in front of me. My knowledge, of this man m, that if one and only one man is standing in front of me, then he, m, is standing in front of me is a posteriori, not a priori. Although this knowledge is based exclusively on the perceptual experience that presents the man to me, and hence allows me to grasp the proposition known to be true, my knowledge is also justified by that experience. As Jim Pryor has usefully reminded us, the fact that perceptual experience may play a crucial role in allowing me to entertain a certain proposition does not negate the fact that it may also play a crucial role in justifying my knowledge of that proposition.15 Hence, the knowledge I express by saying If one and only one man is standing in front of me, then he [demonstrating m] is standing in front of me is a posteriori, as is the knowledge I express by saying If one and only one man is standing in front of me, then Saul is standing in front of me [if I have introduced the name Saul to stand for that man]. The same is true of the gold-stipulator who says If nearly all those samples are instances of a single physically constitutive kind, then nearly all of them are 15

Pryor (n.d.).

Knowledge of Manifest Natural Kinds • 201

instances of it [demonstrating the kind, gold], or If nearly all those samples are instances of a single physically constitutive kind, then nearly all of them are gold [if he has introduced the term gold]. What justifies this knowledge is that it is instances of a certain particular kind—gold—that he is acquainted with, and has empirically justified beliefs about. Hence, his knowledge is a posteriori. For this gold-stipulator (4b) is false. The second case to be considered is one in which we imagine the goldstipulator as not knowing in advance that the items in the sample are of the same physically constitutive kind—even though in fact they are. The stipulation will be a little strange, if the stipulator doesn’t at least believe that they are of the same kind, and take himself to have some evidence for this. However, if his evidence falls short of knowledge, then stipulatively introducing the term won’t put him in a better epistemic position than before—any more than introducing the name Saul when I am not sure anyone is in front of me would improve my epistemic situation in that case. The gold-stipulator will, presumably, assent to the sentence If nearly all these items in the sample are of the same physically constitutive kind, then they are all gold, and to the extent that he is justified in believing that the sample does uniquely determine such a kind, he will be justified in believing the proposition expressed by the sentence.16 He may even know this weaker proposition to be true. But if so, his knowledge is justified by the fact that the items he is perceptually acquainted with, and has beliefs about, are instances of one particular kind—gold—as opposed to any other. Hence, his knowledge is a posteriori, and (4b) is false. As I see it, then, the situation is this. In order to successfully introduce a name or manifest kind term one must be acquainted with the object to be named or the natural kind to be designated. Standardly, this will involve being perceptually acquainted with, and believing certain things of, the object or the kind. In the case of manifest kinds, the normal way of being acquainted with, and believing things of, them is by being acquainted with, and believing things of, some of their instances. In order to successfully introduce a general term designating a manifest kind k, and to use it to express propositions of which k is a constituent, which one knows to be true, one must be acquainted with some particulars that are instances of k, and one must believe, or assume, with at least some justification, that they are instances of a unique kind of the type that k is. In virtue of this, one is counted as knowing, a posteriori, of the particulars that they are instances of k, if they are instances of any relevant kind at all. Having said this, I need to add three clarifying qualifications. First, I have not said that in order to have beliefs about the color green, the 16 The same is true of the sentence For all x, x is gold iff the predicate ‘is gold’ applies to x, and the proposition it expresses.

202 • Essay Seven

substances gold or water, or the species tiger, one must believe of some particular instances of these kinds that they are green, gold, water, or tigers (if they are instances of any relevant kind). Someone introducing these terms with the stipulations I have sketched must have such beliefs, but once the terms have been successfully introduced, they can be picked up by other competent speakers of the language, provided these speakers intend to use the terms with the semantic contents they have already acquired. These speakers need have no beliefs about particular instances of the kinds. Second, some account must be given of what happens when a speaker introduces a term with a stipulation like the one I have given for tiger, without realizing that the supposedly paradigmatic items in the tigersample don’t determine a species of animal at all. Borrowing from Putnam, we may imagine a world-state in which speakers stipulate that the predicate is a tiger is to designate all members of the same species of animal as the tawny, striped, cat-like individuals they have seen in various zoos, as well as in the wild—even though, unknown to them, these socalled tiger specimens are not animals at all, but cleverly disguised robots controlled by space aliens.17 Putnam’s intuition, which I share, is that in this fantastic scenario—in which speakers are under a monumental misimpression—the predicate is a tiger nevertheless turns out to be meaningful, and to truly apply to paradigmatic members of the tigersample. However, its meaning in the imagined world-state is not , where I is the semantic content of the copula and K is the animal species which is the meaning of the term tiger for us, in the world as it actually is. The imagined world-state is not one in which tigers fail to be animals; there are no tigers in that scenario (in our sense of tiger), even though there are things that speakers in the scenario correctly call tigers. Nor is the scenario one in which speakers wrongly believe that tigers are animals; speakers in the scenario have no beliefs about tigers, or the kind tiger, in our sense—even though they have beliefs which they express using the word tiger. Given all this, we can only conclude that there must be a process by which a word introduced with the intention that it is to designate a manifest kind of a certain sort may acquire quite a different meaning. I suspect that what is going on is something like this: one who introduces the predicate is a tiger with the stipulation I have suggested intends (i) that it apply to nearly all specimens in the paradigmatic tiger-sample, (ii) that it apply to other things iff they bear a certain important relation of similarity to specimens in the sample, and (iii) that this similarity 17

Putnam (1962).

Knowledge of Manifest Natural Kinds • 203

relation be the relation of being-an-instance-of-the-same-animal-speciesas. In Putnam’s fantastic scenario, these intentions cannot all be fulfilled, and the predicate acquires a different meaning by default—one which conforms to the first two intentions, but not the third. Depending on the beliefs and intentions of speaker-hearers, plus further empirical facts about the world-state, a new similarity relation comes to be the salient one—with the result that the predicate is a tiger acquires a meaning that is as close as is reasonably possible, given the situation, to the one intended by speakers. The third point of clarification to be added to the picture presented above involves the question of whether it is part of the meaning of the predicate is a tiger that it applies only to animals. Nothing we have said so far settles this question. Since, in the Putnam scenario, the predicate means something other than what it actually means to us, the fact that it applies to nonanimals in that scenario has no bearing on whether its actual meaning involves reference to animals.18 Neither does the fact that Tigers are animals does not express an a priori truth. It is conceivable that it should be part of the meaning of a predicate P that it applies only to individuals with the property expressed by F, even though the proposition expressed by P’s are F is not a priori. Consider the following analogy with names. Suppose I introduce the name Philosopher-Saul with the stipulation that it is to be synonymous with the rigidified description dthat [the x: x is a philosopher and x = Saul Kripke]. In order to understand the name, one must know that it applies to an individual iff that individual is both Saul Kripke and a philosopher.19 Hence, it is part of the meaning of the name that it refers to a philosopher, if it refers to anything at all. However, the proposition semantically expressed by (5)

Philosopher-Saul is a philosopher (if Philosopher-Saul has a referent)

18 By the same token, Putnam’s scenario is essentially irrelevant to the question of whether the propositions we actually use sentences containing the word tiger to express are knowable a priori. If we hadn’t already concluded that it is not knowable a priori that tigers are animals, Putnam’s scenario would not justify drawing this conclusion. 19 It is worth noting that this shows that character in David Kaplan’s sense—a function from contexts of utterance to contents—cannot, in general, be identified with the meaning of an expression, in the sense of that knowledge of which is necessary and sufficient for understanding the expression. If meaning were identified with character, then in a language containing dthat-rigidified descriptions, dthat [D1] and dthat [D2] would be synonymous whenever D1 and D2 were necessarily codesignative, and grasping the meanings of the unrigidified descriptions would not be required for understanding the rigidified descriptions. These results are clearly unacceptable.

204 • Essay Seven

is just the singular proposition that says of Saul Kripke that he is a philosopher (if the name Philosopher-Saul has a referent). Since this proposition can be known only a posteriori, it is not (strictly speaking) knowable a priori that Philosopher-Saul is a philosopher (if Philosopher-Saul has a referent). This is true, even though the a posteriori knowledge needed to understand (5) is, arguably, sufficient for knowledge of the proposition (5) expresses. This illustrates the larger point that in order for a proposition p to be knowable a priori, it is not enough that there be some sentence S which both expresses p and is such that understanding S provides one with all the justification one needs to know p. Such sentences and propositions do have an interesting epistemological status; the sentences might well be termed analytic, and the propositions they express can be known to be true without any empirical justification beyond that required to understanding sentences that express them. Nevertheless, these propositions are not knowable a priori.20 Applying this lesson to the predicate is a tiger, we get the result that if it is part of the meaning of the predicate that it applies only to animals, then, even though the proposition that tigers are animals is not knowable a priori, it is knowable solely by virtue of the knowledge needed to understand the (analytic) sentence Tigers are animals, which expresses it. Whether or not this is part of the meaning of the predicate is a question that I leave open.21 This completes my explanation of the necessity and aposteriority of sentences containing simple natural kind predicates like those in (1). As I see it, the necessity of many of these statements follows from their truth, plus the way in which the reference of the terms they contain is standardly fixed. The explanation of their aposteriority is based on the idea that our knowledge of manifest kinds parallels our knowledge of individuals. Just as our de re knowledge of individuals standardly depends either on our own acquaintance with them, or on the acquaintance of others who pass important parts of their knowledge on to us, so our de re knowledge of manifest kinds standardly depends either on own acquaintance with members of these kinds, or on the acquaintance of others who pass aspects of their knowledge on to us. Because of this requirement on acquaintance, most of our knowledge of individuals, and of manifest kinds, is a posteriori. It is not possible to circumvent this result by using descriptions to introduce or to semantically fix the reference of names or manifest kind terms. In both cases, the requirement that we antecedently believe of the object to be named, or the kind to be designated, that it is denoted by the description used to introduce 20 21

For further discussion, see Soames (2003, vol. 2, chap. 16). For further discussion see Soames (2005, chap. 4).

Knowledge of Manifest Natural Kinds • 205

the term renders our knowledge of the propositions expressed by relevant sentences containing the term a posteriori, rather than a priori.22

Contrast with Two-Dimensionalism The account I have offered of natural kind predicates for manifest kinds is both Millian and nondescriptional. It is Millian in that it holds that the semantic contents of the general terms out of which simple manifest kind predicates are constructed are the kinds they designate. It is nondescriptional in recognizing that although descriptions may be used to introduce these general terms, the terms themselves are standardly not synonymous with rigidified descriptions. Moreover, even when descriptions are involved, their use does not give rise to a priori knowledge, but rather must be grounded in empirical, de re knowledge of their denotations. These features of my account stand in marked contrast to competing twodimensionalist accounts, all of which treat simple general terms like green, gold, water, and tiger as fundamentally descriptive, in one way or another. Since my space is limited, I will here say just a word about one particular type of two-dimensionalist account—something I call weak two-dimensionalism.23 Its central tenets are: Tenets of Weak Two-Dimensionalism T1. Each sentence is semantically associated with a pair of semantic values—primary intension, and secondary intension. The primary intension of S is its Kaplan-style character. The secondary intension of (or proposition expressed by) S at a context C is the proposition assigned by its primary intension to C. T2. Understanding S consists in knowing its primary intension— i.e., its meaning, or character. Although this knowledge, plus complete knowledge of the context C, would give one knowledge of the proposition expressed by S in C, one often does not have such knowledge of C. Since we never know all there is to know about the designated world-state of C, sometimes we don’t know precisely which proposition is expressed by S in C. However, this does not stop us from correctly using S in C. 22 One can, of course, know a priori that for all x if x is gold, then x is gold, even though simply entertaining the proposition involves some kind of acquaintance with the kind, which in turn may require some empirical knowledge. The reason such empirical grounding does not prevent knowledge of this proposition from being a priori is that it plays no role in justifying the proposition we apprehend. 23 For a thorough critique of all major forms of two-dimensionalism, see Soames (2005).

206 • Essay Seven

T3a.

T3b.

T4a.

T4b. T5a.

T5b.

T6a.

Examples of the necessary a posteriori are sentences the secondary intensions of which are necessary, and the primary intensions of which assign false propositions to some contexts. Examples of the contingent a priori are sentences the secondary intensions of which are contingent, and the primary intensions of which assign true propositions to every context. All proper names and natural kind terms have their reference semantically fixed by purely descriptive properties, which can, in principle (given a rich enough vocabulary), be expressed by descriptions not containing proper names or natural kind terms. These names and natural kind terms are synonymous with descriptions rigidified using actually or dthat.24  It is a necessary truth that S is true w.r.t. a context C iff the secondary intension of S in C is true w.r.t. all world-states that are possible relative to C. Standardly, an attitude ascription x v’s that S, when taken in a context C, is true of an agent i iff there is some meaning (character) M such that (i) i bears R to M, and (ii) M assigns to i’s context the secondary intension of S relative to C. So propositions are objects of the attitudes, and attitude verbs are two-place predicates of agents and their objects. However, this two-place relation holds between an agent a and a proposition p in virtue of a three-place relation holding between a, a character, and p. To believe p is to accept a character M that expresses p (and to believe that M expresses a truth). To know a true proposition p is to justifiably accept a character M that expresses p (and to know that M expresses a truth). For all propositions p, p is both necessary and knowable only a posteriori iff (i) p is necessary, (ii) p is knowable in virtue of one’s justifiably accepting some meaning M (and knowing that it expresses a truth), where M is such that (a) it assigns p to one’s context, (b) it assigns a false proposition to some other context, and (c) one’s justification for accepting M (and believing it to express a truth) requires one to possess empirical evidence, and (iii) p is knowable only in this way.

24 The character of dthat [the D] is a function from contexts to the denotation o of the D in the context; propositions expressed by sentences containing dthat [the D] are singular propositions about o. The character of the x: actually Dx is a function from contexts C to the property of being the unique object which “is D” in Cw (the world of C); propositions expressed by sentences containing the description are singular propositions about Cw. 

Knowledge of Manifest Natural Kinds • 207

T6b. For all propositions p, p is both contingently true and knowable a priori iff in addition to being contingently true, p is knowable in virtue of one’s justifiably accepting some meaning M (and knowing that it expresses a truth), where M is such that (a) it assigns p to one’s context, (b) it assigns a true proposition to every context, and (c) one’s justification for accepting M (and believing it to express a truth) does not require one to possess empirical evidence. I will close by using terms for manifest kinds to highlight certain problems with weak two-dimensionalist theories of this sort. As is indicated in thesis T4a, a crucial feature of these theories is their analysis of natural kind terms as having their reference semantically fixed by descriptions. For example, it is common for two-dimensionalists to maintain that the word water has its reference semantically fixed by a description expressing widely shared knowledge about water—something like the description the clear, potable stuff that fills the lakes and oceans.25 Accordingly, these two-dimensionalists take water to be synonymous with a rigidified version of this description, and they take the proposition expressed by (6) Something is water iff it is an instance of the clear, potable stuff that fills the lakes and oceans, . . . to be an example of the contingent a priori. On my view, this is incorrect. Since the proposition expressed by (6) contains the kind water as a constituent, knowledge that it is true requires de re knowledge of the kind, which cannot be had purely descriptively, but rather requires grounding in a posteriori knowledge of some instances of water.26 The two-dimensionalist’s failure to see this is, I think, rooted in the unwavering descriptivism expressed by T4a and T4b. Both are problematic. Contrary to T4a, it is very hard—I believe impossible—to come up with any plausible description to play the role of a semantic referencefixer for a manifest kind term like water, just as it is very hard to come up with such a reference-fixer for a name like Bill Clinton. The reason is the same in both cases: although virtually everyone who uses the terms will associate them with some descriptive information, that information may vary widely from speaker to speaker, and none of it is required in order for someone to understand the terms. Rip Van Winkle, who knew Clinton by name when Clinton was a preteen and who wakes up today knowing only that he has slept for a long time, understands the sentence 25 26

See, for example, Chalmers (1996, 57) and Jackson (1998, 212). Jim Pryor (n.d.) develops a similar critique of two-dimensionalism.

208 • Essay Seven

Bill Clinton is not very honest perfectly well, and he surely counts as a competent user of the name, even though there is little overlap between the information he associates with it and the information we do. The same is true of the term water. Imagine someone who has lived in a very restricted environment—someone who has been confined to a basement, who has never been outside, who knows nothing of oceans or lakes, who has heard the word water and knows that it applies to cloudy stuff running through a drain in his basement from a nearby laundry, but doesn’t know that water is normally clear or drinkable. Such a person may nevertheless correctly use the word water to refer to water, just as we do, and he may understand many sentences containing it, despite not knowing their truth values. What is important is that he has been in contact with the stuff and knows that the word applies to it—just as with Rip Van Winkle and Clinton. Observations like these cast doubt on T4a. However, that thesis is not the only problem for weak two-dimensionalism. Even if one could find plausible candidates for descriptions semantically associated with natural kind terms, the rigidification required by T4b would itself be problematic. Here, weak two-dimensionalists face a dilemma. On the one hand, if they use the actuality operator to rigidify, and analyze the secondary intension of Water is H20 as equivalent to that expressed by (7) Instances of the kind which is actually D are instances of H2O, then they will wrongly characterize the truth of attitude ascriptions like (8) Even if it had been the case that so and so, John would have believed that water was H20 as requiring agents like John in other, merely possible, world-states to have beliefs about the actual world-state in which (8) is assertively uttered.27 On the other hand, if they rigidify using the dthat operator, and analyze the secondary intension of Water is H2O as the proposition expressed by (9) Instances of the kind dthat [D] are instances of H20, then that proposition will contain the kind water as a constituent—in which case, knowledge of its truth must be grounded in some acquaintance with the kind water. Since such acquaintance is not required (for arbitrary D) in order to accept the meaning of (9) and know that it expresses a truth, thesis T5b’s account of what it takes to know that water is H2O is liable to be incorrect, as is the general account of the necessary a posteriori given in 27 This is a straightforward extension to natural kind terms of the argument found in chapter 5 of this volume, Soames (2002).

Knowledge of Manifest Natural Kinds • 209

T6a. In short, the weak two-dimensionalist theory faces apparently intractable difficulties no matter which form of rigidification is chosen.28 So much for my brief critique of the weak two-dimensionalist treatment of manifest kind terms. It is worth noting that this last problem, involving the rigidification required by T4b, does not arise for versions of what is probably a more familiar form of two-dimensionalism—one which I call strong two-dimensionalism. The chief distinguishing characteristic of this view is its rejection of the analysis of attitude ascriptions given in T5b in favor of a more radical thesis that takes belief and knowledge ascriptions—including those involving operators like it is knowable a priori that and it is knowable only a posteriori that—as operating on the primary intensions, rather than the secondary intensions, of their complement clauses.29 According to this view, what is reported by the knowledge ascription x knows that water is H2O is not that the agent knows the secondary intension of Water is H2O, but rather that the agent knows its primary intension—something he will know just in case he knows of the meaning of Water is H2O that it expresses a truth. Since rigidifying operators in the complement clauses of these knowledge ascriptions make no significant contributions to their primary intensions, none of the problems with T4b carry over to versions of strong two-dimensionalism. In the end, however, this is no help, since, as I have argued elsewhere, strong two-dimensionalism is independently refutable on other grounds.30 Although this is a bad result for two-dimensionalist semantic theories, it does not affect our ability to explain instances of the necessary a posteriori involving natural kind terms that designate manifest kinds. As I have argued above, by adhering to the fundamentals of Kripke’s nondescriptionalist account of these terms we can explain both the necessity and the apriority of examples like those in (1), without making unrealistic semantic or metaphysical assumptions.31 28 These (and other) arguments are developed at much greater length in Soames (2005, chap. 10). 29 There are several other differences between strong and weak two-dimensionalism that accompany this one—including a somewhat different characterization of primary intension (which must, on the strong two-dimensionalist view, be a proposition). In addition, strong two-dimensionalists are often inclined to take propositions to be sets of possible worldstates. For a full discussion see Soames (2005). 30 See Soames (2005, chap. 10), as well as Soames (2006a). 31 This paper is a slightly updated version of a talk given on June 5, 2003 at the Third Barcelona Workshop on Issues in the Theory of Reference. The general point of view it presents—especially the critique of two-dimensionalism—is developed more fully in Soames (2005). As for the positive view of natural kind terms developed here, a substantial portion of the middle part of this essay—on natural kind terms, reference-fixing, and the a priori—has been incorporated into a section, with that title, of chapter 4 of the same book.

210 • Essay Seven

References Byrne, Alex, and David R. Hilbert. 1997. “Colors and Reflectances.” In Readings on Color, vol. 1, The Philosophy of Color, ed. Byrne and Hilbert, 263–88. Cambridge: MIT Press. Chalmers, David. 1996. The Conscious Mind. New York: Oxford University Press. Jackson, Frank. 1998. “Reference and Description Revisited.” In Language, Mind, and Ontology, ed. James E. Tomberlin, 201–18. Philosophical Perspectives 12. Oxford: Blackwell. Kaplan, David. “Demonstratives: An Essay on the Semantics, Logic, Metaphysics, and Epistemology of Demonstratives and Other Indexicals.” In Themes from Kaplan, ed. Joseph Almog, John Perry, and Howard Wettstein with the assistance of Ingrid Deiwiks and Edward N. Zalta, 481–563. New York: Oxford University Press. Kripke, Saul A. 1980. Naming and Necessity. Cambridge: Harvard University Press. Originally published in Semantics of Natural Language, ed. Donald Davidson and Gilbert Harman, 253–355 (Boston: Reidel, 1972). Pryor, James. n.d. “Thinking about Water.” In preparation. Putnam, Hilary. 1962. “It Ain’t Necessarily So.” Journal of Philosophy 59:658–71. Salmon, Nathan. 2003. “Naming, Necessity, and Beyond.” Mind 112:475–92. Soames, Scott. 1999. Understanding Truth. New York: Oxford University Press. ———. 2002. Beyond Rigidity: The Unfinished Semantic Agenda of “Naming and Necessity.” New York: Oxford University Press. ———. 2003. Philosophical Analysis in the Twentieth Century. 2 vols. Princeton: Princeton University Press. ———. 2004. “Reply to Ezcurdia and Gómez-Torrente.” Crítica 36:83–114. ———. 2005. Reference and Description: The Case against TwoDimensionalism. Princeton: Princeton University Press. ———. 2006a. “Kripke, the Necessary Aposteriori, and the Two-Dimensionalist Heresy.” In The Two-Dimensional Semantics, ed. Manuel García-Carpintero and Josep Macia, 272–92. Oxford: Clarendon Press; New York: Oxford University Press. ———. 2006b. “Reply to Critics.” Philosophical Studies 128:711–28.

ESSAY EIGHT

Understanding Assertion

Introduction In his groundbreaking 1978 article, “Assertion,” Robert Stalnaker presents an elegant model of discourse designed to solve philosophical problems arising, in part, from his identification of propositions with functions from possible world-states to truth-values, and his restriction of the epistemically possible to the metaphysically possible. Among these problems are those posed by Kripkean examples of the necessary a posteriori. Being necessary, all such examples are seen by Stalnaker as semantically expressing the same trivial, universally known, a priori truth. Nevertheless, assertive utterances of them often result in the assertion of propositions that are both highly informative and knowable only a posteriori. A central task of “Assertion” is to explain how this can be so. Stalnaker’s model is based on the insight that considerably more goes into determining what is said by an assertive utterance than the meaning of the sentence uttered. Additional assertion-determining factors include (i) objective features of the context of utterance, such as the speaker, audience, time, place, and world-state of the context, (ii) general conversational rules, including those against asserting what is already known or presupposed to be true (or to be false), and (iii) salient beliefs and assumptions known to be shared by conversational participants. These latter encompass beliefs and assumptions about who is speaking to whom, what words are being uttered and what they mean, what is happening in and around the speech situation, the topic of conversation, what has already been established or taken for granted, and what remains on the conversational agenda. Imagine, for example, a speaker who utters She is late in response to the entrance of a woman who comes into a meeting after it has already begun, attracting everyone’s attention. The speaker relies on the fact that everyone will recognize him to be saying, of the woman who just entered, that she is late for the meeting. In this case, obvious facts about the conversational context interact with the meaning of the sentence uttered to determine the proposition asserted. Other cases are more indirect, and may involve reinterpretation of what the speaker might, at first, appear to have said. For example, after listening to the remarks of a well-known campus orator, Mary might

212 • Essay Eight

turn to her companion and say “Norman really is God’s fountain pen, isn’t he?” knowing full well that her companion won’t take her to be asserting the proposition semantically expressed by the sentence she utters— one which predicates of Norman the property of being a certain kind of artifact (used to write by depositing ink on paper) possessed by God. Since that proposition is egregiously false, Mary realizes that her conversational partner will recognize that she is not committing herself to it, and so will look for an alternative interpretation of her remark. Depending on what else is taken for granted in the conversation, Mary might be taken to have asserted (i) that God really is using Norman to communicate his thoughts and desires, (ii) that (God aside) Norman really does have the truth on the matters about which he is speaking, or (iii) that Norman is a blowhard who takes himself to be an authority, even though he really isn’t (irony). This variation is not a sign of indexicality. Although Mary’s utterance would result in assertions of different propositions in different contexts, this is not a matter of semantics. It is a matter of the way in which the meaning of the sentence uttered interacts with both principles governing discourse and the background beliefs and assumptions of conversational participants. Similar principles govern the reinterpretation of utterances of trivially obvious, literal truths. For example, a candidate awaiting the results of an election after the polls have closed might respond to an early bit of unfavorable news by saying “What will be will be.” In such a case, the speaker understands that his audience will not take him to have asserted a trivial tautology, but will instead interpret him to have said something significant—typically something to the effect that since the outcome is out of his control, he is prepared to accept whatever the result proves to be. These examples are simple and relatively uncontroversial. However, as Stalnaker correctly recognizes, they are part of a larger, more systematic picture. A central message of “Assertion” is that there is often a substantial gap between the propositions assigned to sentences by a correct semantic theory and those asserted by utterances of these sentences in different contexts. Because of this, continued progress in solving problems in semantics and philosophical logic depends on our coming to have a better understanding of the ways in which semantic and nonsemantic factors interact in determining what is asserted and conveyed by utterances, and in guiding rational discourse and inquiry. This is the central insight, and seminal contribution, of Stalnaker’s inquiry— and one with which I fully agree. Nevertheless, I am skeptical about some of the burdens taken up in the article—in particular, the attempt to render familiar examples of the necessary a posteriori compatible with the restriction of epistemically

Understanding Assertion • 213

possible to metaphysically possible world-states, and the identification of the objects of assertion with functions from such states to truth-values. Thus, my first task, after reconstructing Stalnaker’s discourse model, will be to demonstrate why it cannot be used to reconcile the necessary a posteriori with his antecedent conceptions of possibility and of propositions. Instead, I will argue, a proper understanding of the necessary a posteriori requires the restriction of epistemic possibility to metaphysical possibility to be abandoned, and the existence of epistemically possible world-states that are metaphysically impossible to be recognized. Having reached this point, I will next isolate inherent limitations of the discourse model itself by showing that de re belief and the nontransparency of meaning lead to insoluble problems—even after the model has been improved by substituting the broader class of epistemically possible world-states for the narrower class of metaphysically possible world-states. Finally, I will draw lessons from these problems, distinguish aspects of Stalnaker’s discourse model that need to be revised from those that should be retained, and suggest how further progress can be made in understanding the ways in which semantics and pragmatics interact in assertion.

Stalnaker’s Model of Discourse According to the model, conversations take place against a set of background assumptions shared by the conversational participants which rule out certain possible world-states as not obtaining, or “being actual.”1 As the conversation proceeds, and assertions are made and accepted, new propositions are admitted into the set of shared background assumptions, and the set of world-states that remain compatible with what has been assumed or established shrinks. This set is called the context set (at any given point in the conversation). The aim of further discourse is to further narrow down this set of possibilities, within which the actual state of the world—the maximally complete property that the universe really instantiates—is assumed to be located. When one asserts p, the 1 Here, and throughout, I use the term possible world-state instead of the more familiar possible world to reflect my view, shared with Stalnaker, that the items under discussion are not alternate concrete universes, but ways the world could have been—i.e., maximally complete properties the universe could have instantiated. To say of such a state that it “is actual” is to say that the universe instantiates it. I also limit myself to contexts of utterance that Stalnaker calls nondefective (those in which the possible world-states that the speaker takes to be compatible with everything believed and assumed by conversational participants at a given point in the conversation are the same as those that the hearers take to have this property), plus contexts that he calls close enough (those in which differences between speakers and hearers on this point don’t arise, or affect the course of the conversation.)

214 • Essay Eight

function of one’s assertion is to shrink the context set by eliminating from it all world-states in which p is not true. Stalnaker postulates three rules governing assertion.2 R1. A proposition asserted should always be true in some but not all of the possible world-states in the context set. R2. Any assertive utterance should express a proposition, relative to each possible world-state in the context set, and that proposition should have a truth-value in each possible world-state in the context set. R3. The same proposition should be expressed relative to each possible world-state in the context set. The rationale for R1 is that a proposition true in all world-states of the context set would be uninformative, and so would fail to perform the essential function of assertion, which is to narrow down the range of world-states that conversational participants take to be candidates for being the way the world actually is. By the same token, a proposition false in all world-states in the context set would contradict what has already been conversationally established. Since it would eliminate the entire context set, it would also fail to narrow down the range in which the actual world-state is to be located. Of course, this rule, like the others, allows for some flexibility in how it applies. If someone seems to say something that violates it, one may sometimes conclude that no violation has really taken place because the context set isn’t quite what one originally thought, or because the speaker didn’t really assert, or mean, what he at first seemed to assert or mean. This is not to say that violations never occur, but it is to say that common knowledge of the rule can sometimes be exploited for conversational purposes—as when a speaker deliberately says something the literal interpretation of which would violate the rule, knowing full well that he will be reinterpreted in a certain obvious way so as to be seen as conforming to it.3 Stalnaker’s rationale for R2 is that if an utterance violates it, then for some world-state w in the context set, the assertive utterance won’t determine whether it should remain in the set, or be eliminated. If the sentence uttered does not express a proposition at w, or if it does express a proposition, but one for which no truth-value—truth or untruth—is defined at w, then no verdict on whether w stays or goes will, Stalnaker thinks, be forthcoming. This is to be avoided.4 In explaining the rationale for R3 Stalnaker employs his notion of the propositional concept associated with an assertion. A propositional 2

Stalnaker (1999, 88). Stalnaker (1999, 89). 4 Stalnaker (1999, 89–90). 3

Understanding Assertion • 215

concept is very much like one of David Kaplan’s characters. For Stalnaker, it is a function from world-states, considered as possible contexts of utterance, to propositions—where propositions are taken to be nothing more than assignments of truth-values to world-states, considered as circumstances of evaluation. The propositional concept associated with an utterance of a sentence S at a certain moment m in a conversation is a function that maps each world-state w in the context set at m onto a proposition—which is simply an assignment of truthvalues to all world-states in the context set. This assignment of truthvalues is (implicitly) identified with the proposition that would be expressed by S at m, if the actual context of utterance were to turn out to be w. Propositional concepts can be given pictorial representations, as is indicated by Stalnaker’s matrix D. D i j k

i T F F

j T F T

k T T T

D represents the propositional concept associated with the use of S at a moment m in which the context set consists of the world-states i, j, and k. D tells us (i) that if i is the state the world is actually in at m, then the proposition (semantically) expressed by the speaker’s utterance of S is the proposition that assigns truth to every world-state of the context set, (ii) that if j is the state the world is actually in at m, then the proposition (semantically) expressed assigns truth to k and falsity to i and j, and (iii) that if k is the state the world is actually in at m, then the proposition (semantically) expressed assigns falsity to i and truth to the other two world-states.5 Stalnaker uses D to give the following rationale for R3. To see why the principle must hold, look at the matrix for the propositional concept D. Suppose the context set consists of i, j, and k, and that the speaker’s utterance determines D. What would he be asking his audience to do? Something like this: If we are in the world i, leave the context set the same; if we are in the world j, throw out worlds i and j, and if we are in world k, throw out just world i. But of course 5 Stalnaker (1978) is not fully clear about what status the propositions “expressed by” S at the world-states of the context set are supposed to have. Although they are not always the propositions that would be asserted, if those world-states were to obtain, they are often propositions semantically expressed in those eventualities—hence the parenthetical “semantically” above. However, there are exceptions to this—to which I will return—which prevent any such general identification.

216 • Essay Eight

the audience does not know which of those worlds we are in, and if it did the assertion would be pointless. So the statement, made in that context, expresses an intention that is essentially ambiguous. Notice that the problem is not that the speaker’s utterance has failed to determine a unique proposition. Assuming that one of the worlds i, j, or k, is in fact the actual world, then that world will fix the proposition unambiguously The problem is that since it is unknown which proposition it is that is expressed, the expression of it cannot do the job that it is supposed to do.6 The idea is that if R3 is violated, the conversational participants won’t know which proposition is (semantically) expressed by the sentence uttered, because they won’t know which world-state “is actual.” But if the proposition asserted is always the one (semantically) expressed by the sentence uttered (in the context), then the conversational participants won’t know what is asserted, and so will be at a loss as to how to update the context set and proceed with the conversation. This is the rationale for R3.

Presuppositions of the Model Before going further it is worth pointing out certain presuppositions of the model, and dealing with obvious worries that might arise. The model presupposes that speakers have a great deal of knowledge—(i)–(iii)— about possible world-states. (i) For every world-state w, conversational participants at a time t in the conversation know whether w is compatible with everything believed, established, or assumed in the conversation at t, and hence whether w is in the context set at t. (ii) For any sentence S that might be uttered at t, any world-state w in the context set at t, and any proposition p, if an utterance of S at t would express p, were it to turn out that w were actual (i.e., were w to turn out to be the world-state that is actually instantiated), then conversational participants know that this is so. (iii) For any proposition p and world-state w, conversational participants know the truth-value of p in w—i.e., they know what the truth-value of p would be were w to be actual. It is natural to wonder whether speaker-hearers really have all this knowledge of world-states. 6

Stalnaker (1978, 90–91; my emphasis).

Understanding Assertion • 217

One worry concerns what might be called the size of world-states. Each world-state encodes a massive amount of information about the universe—far too much for our minds to encompass. But if that is so, how are we able to know anything significant about such entities? The answer is that the knowledge of world-states required by the model is not very extensive. Although each world-state encodes a massive amount of information, only a tiny fragment of it will relevant in any given conversational setting. Because of this, we can ignore differences among world-states that are irrelevant to our conversational purposes. For example, if our conversation has been exclusively about the 2004 American League Championship Series between the Boston Red Sox and the New York Yankees, we can form equivalence classes of worldstates that agree on their accounts of the series, while differing arbitrarily on extraneous matters. In representing the context set and propositional concept for an utterance u, we can then take each ‘w’ as standing for one of these equivalence classes. Knowing of each equivalence class (and thereby of each member in it) that it is compatible with everything assumed, established, or believed at the time of u can then be assimilated to knowing, of the account of the series on which all members of the class agree, that it is compatible with all this background information. If this is correct, then the worry about size disappears. Although nothing I have said guarantees that speaker-hearers can, in general, be relied upon to have knowledge of types (i)–(iii), presupposed by the model, the sheer quantity of information encoded in world-states is not itself an obvious barrier. The next thing to notice is that the model presupposes systematic de re knowledge of world-states. For each relevant world-state w, sentence S, and proposition p, speakers in a conversation C are said to know (i) that w is (or is not) compatible with the background assumptions of C, (ii) that an utterance of S would (or would not) express p, if w were to turn out to be actual, and (iii) that p would be true (false), if w were to obtain. In each case, an occurrence of the variable ‘w’ appears inside the content clause of the knowledge ascription, while being bound by a quantifier outside the clause. Since this is the mark of de re knowledge-ascriptions, the model attributes far-reaching de re knowledge of world-states to conversational participants. How should we think of this? Consider the following example: A is speaking to B at a conference on the philosophy of language. A points at a man across the room and says “He teaches at UCLA.” Suppose, for whatever reason, that the following world-states are members of the context set. w1: A is pointing at David Kaplan, and Kaplan teaches (exclusively) at UCLA, and . . .

218 • Essay Eight

w2: A is pointing at David Kaplan, and Kaplan teaches (exclusively) at USC, and . . . Stalnaker’s discourse model presupposes that conversational participants know of w1 and w2 that if they “are actual” (i.e., if either one obtains or is instantiated), then A’s utterance will express the proposition pk that David Kaplan teaches at UCLA. In addition, the model presupposes that conversational participants know of w1 that if it “is actual,” then pk will be true, while knowing of w2 that if it “is actual,” then pk will be false. Since the only relevant aspects of w1 and w2 are those indicated above, this de re knowledge of world-states amounts, essentially, to knowledge of the following propositions: a. b.

that if A is pointing at David Kaplan, then A’s utterance expresses the proposition that David Kaplan teaches at UCLA. that the proposition that David Kaplan teaches at UCLA is true, if Kaplan teaches (exclusively) at UCLA, while it is false, if Kaplan teaches (exclusively) at USC.

On this account, the de re knowledge of world-states presupposed by the model is pretty easy to come by. In the case of (a), it involves knowledge of David Kaplan, and of the meaning of the sentence uttered. In the case of (b), it is a priori knowledge that every speaker acquainted with Kaplan can be expected to have. Seen in this light, the presuppositions of the model may seem to be readily satisfiable. There is, however, cause for concern. Typically, the de re knowledge of world-states presupposed by the model will, as in the previous example, bottom out in ordinary de re knowledge of individuals (natural kinds, or other constituents of the world). In our example, knowledge of the world-states w1 and w2, and of the proposition pk, that the latter would be expressed if either of the former were “actual,” as well as knowledge of the truth-values the proposition would have in those eventualities, is really nothing more than knowledge of David Kaplan (a) that if the speaker is pointing at him, then the speaker’s utterance will express pk, and (b) that if he teaches at UCLA, then pk will be true, whereas if he teaches at USC, pk will be false. However, de re attitudes of this sort are notorious for resisting the neat logical transitions presupposed by Stalnaker’s model. For example, it is well known—from the discussion of puzzling Pierre in Kripke (1979), as well as from the discussions of other examples—that one can know of one and the same individual i that he is F and that he is G, without knowing (or being in a position to know) of i that he is both F and G. Similarly, one can know of i that he is F and that if he is F, then he is G, without knowing (or being in a position to know) of i that he is G; and one can know of i that he is F, while also knowing

Understanding Assertion • 219

that S expresses the proposition that he is F, without knowing (or being in a position to know) that S expresses a truth. Cases like these pose a threat to Stalnaker’s model of discourse. This threat will be examined in due course. For now, we simply note it. Before moving on, we need to understand another aspect of the de re knowledge of world-states presupposed by the model—de re knowledge of the actual world-state. According to Stalnaker, the point of rational inquiry and conversation is to reduce, as much as possible, the space of possible world-states within which the actual world-state is believed, or known, to be located. On this picture, the idealized goal of these activities is to eliminate all possible world-states but one, which can then be correctly identified as actual. Why is this desirable? Well, it is natural to think, an agent who correctly identifies the actual world-state @ is thereby in a position to know everything. This will be so if (i) coming to know of the world-state @ that it is actual inevitably involves coming to know, for each genuine truth p, that p is true in @,and (ii) knowing that p is true in @ involves knowing, or being in a position to know, p. However, this reasoning is incorrect. First consider (ii). A fundamental presupposition of the model is that for any world-state w in the context set, and any proposition p that might be asserted in the conversation, speakers and hearers know the truth-value of p in w. In many conversations—those without false presuppositions or assertions—the actual world-state @ will be a member of the context set. It follows that, in these conversations, there will be many propositions p that agents know to be true in @, without knowing (or having any way of coming to know) p. Thus, knowing that p is true in the world-state @, which actually obtains, is not sufficient for knowing p. This brings us to (i). It is tempting to think that the reason an agent can know, of @, that p is true in it without knowing p is that, in cases like this, the agent does not know, of @, that it really obtains, or is instantiated. Were the agent to know this, the thought continues, the agent could not know that p is true in @ without thereby knowing p, as well. In fact, however, this is highly dubious. Imagine an agent who, at a certain point, says or thinks to himself “this world-state, the one I find myself in now (which I know to be such and such, and so and so) is the one that really obtains, or is instantiated.” It certainly seems that such an agent demonstratively refers to the actual world-state @, with which he is acquainted, and truly says, of @, that it obtains, or is instantiated. If he is sincere, the agent should qualify as knowing, of @, that it is the actual world-state. Most likely, he and his fellow conversationalists have known this all along. But then, if identifying the actual world-state is coming to know, of that state, that it “is actual,” then identifying the actual world-state cannot be the idealized goal of rational inquiry, or conversation.

220 • Essay Eight

This conclusion could be resisted, if it could be shown that the familiarity of ordinary agents with the way things are in the universe is inevitably too limited and fragmentary to provide them with the sort of acquaintance with @ needed to acquire de re knowledge of it at all. Although the idea that such knowledge is unattainable is not without force, it is also not easy to accept. For one thing, accepting it would render that staple of indexical semantics, the actuality operator, essentially useless. Since the proposition expressed by Actually S is a proposition which says, of the world-state Cw of the context, that p is true in Cw (where p is the proposition expressed by S in C), an inability to have de re knowledge of @ would prevent speakers from ever knowing the propositions expressed by utterances of sentences containing the actuality operator— thereby depriving such sentences of any normal use. Since these sentences do seem to have such a use, there is reason to believe that de re knowledge of @ is possible. Such knowledge is also defensible on other grounds. Although de re knowledge of individuals normally requires some sort of contact with them, it does not require extensive or systematic knowledge of the totality of facts involving them. For example, even though my knowledge of my city, my country, my planet, my solar system, and my universe is an infinitesimal fraction of all there is to be known about these things, I am surely able to acquire some de re knowledge of them. If de re knowledge of states of individuals (including states of the universe) is similar in this respect to de re knowledge of individuals themselves, then it too is compatible with extreme limitations on the extent and systematicity of such knowledge. Thus, the limited and fragmentary nature of our knowledge of the actual world-state, @, presents no obvious bar to our having some de re knowledge of it. Finally, it should be noted that a proponent of Stalnaker’s model of discourse is in no position to deny this. Since the model routinely attributes de re knowledge of world-states to speakers on the basis of a much slenderer acquaintance with those states than any of us have with @, the proponent of the model ought to accept the idea that conversational participants do have de re knowledge of @. Once this is accepted, there is, as I have argued, no plausible grounds for denying that we know of @ that it is actual, or instantiated. This brings us back to the goal of rational inquiry and discourse presupposed by the model. We have seen that the goal cannot be that of identifying the actual world-state, in the sense of coming to know, of the actual world-state @, that it obtains, or is instantiated. What, then, should we take the goal to be? The answer that the proponent of the model ought give is, I think, that the goal is to “identify the actual world-state” in the sense of arriving at maximally complete, descriptive knowledge of the form, the state of the world that actually obtains, or is instantiated, is one

Understanding Assertion • 221

in which p, q, r, . . . where ‘p, q, r, . . .’ are filled in with a comprehensive list of the facts of @. There are two things to notice about this answer. First, it is compatible with Stalnaker’s discussion of the model—since approaching the goal involves learning, or coming to accept, more and more truths, which has the effect of shrinking the set in which the world-state that actually obtains is, and is assumed to be, located. Second, on this way of understanding the goal, world-states have no priority over propositions. The goal of identifying the actual world, in the sense in which we have now come to understand it, is simply that of learning as many (relevant) truths as we can. It is hard to quarrel with that.

A Lesson about the Necessary A Posteriori We are almost ready to tackle Stalnaker’s attempt to use his discourse model to explain Kripkean examples of the necessary a posteriori. Before we do, however, it is worth pausing to tease out an important consequence of the model regarding how the necessary a posteriori should not be understood. The consequence involves sentences containing the actuality operator that are often taken to be paradigmatic instances of this category of truths. Although any sentence of this sort will do, we will focus on those constructed from contingently codesignative descriptions—the x: x is F and the x: x is G—that are rigidified using ‘actually’. This gives us two descriptions—the x: actually x is F and the x: actually x is G—which designate the same object o in every possible world-state in which o exists, and designate nothing in any world-state in which o doesn’t exist. These are used to construct (1). (1) If [the x: actually x is F] exists, then [the x: actually x is F] = [the x: actually x is G] (1) is necessary, since it is true by falsity of antecedent in any world-state in which o doesn’t exist, and true by truth of the consequent in any world-state in which o does exist. Is (1) knowable a posteriori? Well, one might come to know it is by first coming to know the contingent truth (2) If [the x: x is F] exists, then [the x: x is F] = [the x: is G] and inferring (1) from (2). Since (2) can be known only a posteriori, anyone who comes to know (1) by this route knows it a posteriori. However, since all a priori truths can also be known a posteriori, there is nothing significant about this. In order to show that (1) is a genuine instance of the necessary a posteriori, one must show that it cannot be known a priori, and so is knowable only a posteriori. However, if the lessons we have

222 • Essay Eight

drawn from Stalnaker’s model about our knowledge of possible worldstates are correct, then this cannot be shown. Consider a scenario in which we imagine a possible state of the world to ourselves, or perhaps a class of such states. We say to ourselves, Let w be a possible world-state in which o is the unique thing which is F, and o is the unique thing which is G, and . . . and . . . . We go on enumerating the aspects of w for awhile, and then ask Is w a world-state with respect to which (2) is true? We answer that, of course, it is. On Stalnaker’s model this counts as knowing of w that (2) is true with respect it—which is knowing that which is expressed by (3a). (3a) In w: if [the x: x is F exists], then [the x: x is F] = [the x: x is G] Since (3a) is a priori equivalent to (3b) and (3c), knowing the former, on the basis of our a priori imagining, provides a sufficient basis for coming to know the latter in the same way. (3) b. If, in w, [the x: is F] exists, then, in w, [the x: x is F] = [the x: is G] c. If [the x: in w, x is F] exists, then [the x: in w, x is F] = [the x: in w, x is G] Now let’s suppose something else—namely that the state of the world w we have been imagining is, unknown to us, its actual state. In other words, the state the universe actually is in has precisely the characteristics we were imagining, even though we didn’t realize this at the time. If this is so, then in knowing (3a), and hence, (3c), a priori, we knew (4) a priori as well.7 (4) If [the x: in @, Fx] exists, then [the x: in @, x is F] = [the x: in @, x is G] But then, since (1) expresses the very same thing as (4), it too is knowable a priori, and so is not an instance of the necessary a posteriori. A similar conclusion holds for every purported instance of the necessary a posteriori that makes essential use of the actuality operator. This is significant, since for a number of philosophers, particularly those who attempt to explain the necessary a posteriori by appeal to so-called two-dimensionalist semantics, such sentences have provided the template for understanding a necessary a posteriori truths.8 7 We may, of course, have been imagining a class of world-states satisfying our stipulations, of which @ is a member, rather than imagining @ by itself. However, if, as Stalnaker’s model presupposes, knowing of this class that (2) is true with respect to its members counts as knowing of each member that (2) is true with respect to it, then the argument is not affected. Note, the argument does not depend on the model’s problematic identification of propositions with functions from world-states to truth-values. 8 See Davies and Humberstone (1980) and Soames (2005b), plus the references cited there.

Understanding Assertion • 223

Stalnaker’s Account of the Necessary A Posteriori We return to the three rules governing assertion in Stalnaker’s model of discourse. R1. A proposition asserted should always be true in some but not all of the possible world-states in the context set. R2. Any assertive utterance should express a proposition, relative to each possible world-state in the context set, and that proposition should have a truth-value in each possible world-state in the context set. R3. The same proposition should be expressed relative to each possible world-state in the context set. Having motivated these rules, Stalnaker uses them to explain assertive utterances of Kripkean examples of the necessary a posteriori. As with the other principles, one may respond to apparent violations [of R3] in different ways. One could take an apparent violation as evidence that the speaker’s context set was smaller than it was thought to be, and eliminate possible worlds relative to which the utterance receives a divergent interpretation. Or, one could reinterpret the utterance so that it expresses the same proposition in each possible world. Consider an example: hearing a woman talking in the next room, I tell you, That is either Zsa Zsa Gabor or Elizabeth Anscombe. Assuming that both demonstrative pronouns and proper names are rigid designators—terms that refer to the same individual in all possible worlds—this sentence comes out expressing either a necessary truth or a necessary falsehood, depending on whether it is one of the two mentioned women or someone else who is in the next room. Let i be the world in which it is Miss Gabor, j the world in which it is Professor Anscombe, and k a world in which it is someone else, say Tricia Nixon Cox. Now if we try to bring the initial context set into conformity with the third principle [R3] by shrinking it, say by throwing out world k, we will bring it into conflict with the first principle [R1] by making the assertion trivial. But if we look at what is actually going on in the example, if we ask what possible states of affairs the speaker would be trying to exclude from the context set if he made that statement, we can work backward to the proposition expressed. A moment’s reflection shows that what the speaker is saying is that the actual world is either i or j, and not k. What he means to communicate is that the diagonal proposition of the matrix E exhibited below, the proposition expressed by ⇑E, is true.9 9

Stalnaker (1999, 91).

224 • Essay Eight

E i j k

i T T F

j T T F

k T T F

⇑E i j k

i T T T

j T T T

k F F F

In this example, the propositional concept E associated with the sentence S uttered by the speaker tells us two things: (i) we don’t know which proposition is (semantically) expressed by S in the actual context, because which proposition is expressed depends on which world-state actually obtains, and we don’t know which state does obtain; (ii) none of the possible propositions expressed would serve any useful purpose in the conversation. To assert a necessary truth is to assert something which is of no use in narrowing down the location of the actual world-state within the context set; and asserting a necessary falsehood is even worse. Thus, E violates R3, and any attempt to avoid this violation by excluding one or more of the world-states will violate R1. So, if we are to avoid violation entirely, and to regard the speaker’s utterance as useful and informative, we must take it as asserting some proposition other than the proposition it (semantically) expresses at i, j, or k. Which proposition? Since whatever the actual world-state turns out to be, the speaker will be committed to the utterance of S expressing a truth in the context, that is what we should take to be asserted. The proposition asserted is the proposition that is true (false) at a world-state w (of the context set) just in case the proposition (semantically) expressed by S in w is true (false) at w—it is the assignment of truth-values that arises from E by looking along the diagonal and selecting the truth-value that appears in row w of column w, for each w. Stalnaker calls this the diagonal proposition. Since, in this example, the diagonal proposition is neither true in all world-states of the context set nor false in all those states, it can do the job that asserted propositions are supposed to do—shrink the set. Hence, he maintains, this is the proposition that is really asserted by the speaker’s utterance—no matter which member of the context set turns out actually to obtain. This is what ⇑E represents, where ‘⇑’ (pronounced DAGGER) is an operator that maps a propositional concept C1 onto the propositional concept C2 that arises from C1 by taking each of the rows of C2 to be the diagonal proposition determined by C1. This is the prototype for Stalnaker’s treatment of the necessary a posteriori, which—extrapolating and generalizing his explicit remarks—we may take as suggesting T1. T1.

Although no necessary propositions are knowable only a posteriori, a sentence S, as used in a particular conversation C, is an example of the necessary a posteriori iff the proposition

Understanding Assertion • 225

(semantically) expressed by S at the world-state that really obtains in the speaker’s context is necessary, but the diagonal proposition asserted by a use of S in C is contingent, and hence knowable only a posteriori. Given T1, plus Stalnaker’s discussion of the example—That is either Zsa Zsa Gabor or Elizabeth Anscombe—motivating it, one might get the mistaken impression that he thought that all genuine examples of the necessary a posteriori are indexical, in the sense of semantically expressing different propositions in different contexts of utterance. However, he didn’t believe this. How, then, were instances of the necessary a posteriori involving names and natural kind terms to be treated? He addresses this point in the following passage. I suggest that a common way of bringing utterances into conformity with the third principle [R3] is to interpret them to express the diagonal proposition, or to perform on them the operation represented by the two-dimensional operator DAGGER. There are lots of examples. Consider: Hesperus is identical with Phosphorus, it is now three o’clock, an ophthalmologist is an eye doctor. In each case, to construct a context which conforms to the first principle [R1], a context in which the proposition expressed is neither trivial nor assumed false, one must include possible worlds in which the sentence, interpreted in the standard way, expresses different propositions. But in any plausible context in which one of these sentences might reasonably be used, it is clear that the diagonal proposition is the one that the speaker means to communicate. The two-dimensional operator DAGGER may represent a common operation used to interpret, or reinterpret, assertions and other speech acts so as to bring them into conformity with the third principle [R3] constraining acts of assertion.10 Let us focus on (5a) and (5b). (5) a. Hesperus is identical with Phosphorus. b. An ophthalmologist is an eye doctor. Since (5a, b) don’t contain indexicals, their meanings—i.e., their Kaplanstyle characters—will be constant functions. Each expresses the same (necessary) proposition in every context of utterance. If the propositional concepts associated with them in these conversations were simply their meanings, then the application of the dagger operation would have no effect, and Stalnaker’s explanation of their informative use wouldn’t get off the ground. Thus, in these cases, he must not have been taking the needed 10

Stalnaker (1999, 92).

226 • Essay Eight

propositional concepts to be the meanings (characters) of the sentences uttered.11 Instead, it is natural to interpret him as taking the propositional concept associated with an utterance of S in a conversation to be that which speaker-hearers ( jointly) believe the meaning of S to be. In cases in which they know all the relevant semantic facts, this will simply be the meaning of S. In cases in which they are ignorant of, or confused about, some of these facts, the propositional concept associated with S may be something less than the actual meaning of S. For example, in the case of (5a), the propositional concept may be given by the formula x is identical with y— with different possibilities regarding the constant functions from worldstates to objects which are candidates for the meanings of the names to be substituted for the ‘x’ and ‘y’ being reflected in different world-states of the context set. The case of (5b) is similar, except that the different possibilities for filling in the content of ‘O’ in the relevant formula—An O is an eye doctor—are meanings of general terms, rather than meanings of proper names. On this interpretation, the context set for an utterance of (5a) will contain some world-states in which one or both of the names Hesperus and Phosphorus stand for something other than what they both actually stand for, and the context set for an utterance of (5b) will contain some world-states in which ophthalmologist means something other than what it actually means. Presumably, the justification for this way of looking at things is the idea that (5a) and (5b) will be used only if (some) conversational participants are ignorant about what these words actually mean, or stand for—with the result that world-states in which the words mean, or stand for, something different from what they actually mean, or stand for, will be among the genuine possibilities left open by the conversation prior to the utterances. But then, the thought continues, different propositions will be expressed when the sentences are “interpreted in the standard way,” at these world-states, considered as contexts. This, I think, is how Stalnaker intended to generalize his explanation beyond genuinely indexical sentences.12 At this point, however, we run into a problem. Although there may be some sentences and conversations that fit the picture, some do not. For 11 Unlike later two-dimensionalists, Stalnaker never subscribed to the general thesis that names and natural kind terms are indexical, rigidified descriptions. For discussion, see the introduction to Stalnaker (1999, especially 14–19), and also Stalnaker (2003, especially 199–200). 12 On this interpretation, propositional concepts map each world-state w in the context set onto the proposition that speaker-hearers believe would be semantically expressed if w were to obtain. When they know all relevant semantic facts, these are the propositions that really would be expressed if w obtained.

Understanding Assertion • 227

example, it is not true that (5a) would be used only in a conversation in which (some) conversational participants are ignorant of what Hesperus and Phosphorus stand for, in the sense most relevant to Stalnaker’s model. Each participant may know perfectly well that ‘Hesperus’ refers to this object [pointing in the evening to Venus] and that ‘Phosphorus’ refers to that object [pointing in the morning to Venus]. They may even have done the pointing themselves. Clearly, such speakers know of the referent of each name that it is the referent of that name. Hence the (contingent) propositions expressed by ‘Hesperus’ refers to x and ‘Phosphorus’ refers to x relative to an assignment of Venus to ‘x’ should be among those that have already been assumed or established in the conversation. But then, metaphysically possible world-states in which the names mean and refer to different things will already have been eliminated from the context set as incompatible with what has been assumed or established. Since (5a) can, nevertheless, be used in these circumstances perfectly intelligibly, Stalnaker’s explanation cannot successfully be applied to this case. This is an instance of the general problem noted earlier. When de re attitudes are involved, speakers cannot always determine the compatibility relations presupposed by the model.

Failure of the Model A related problem is posed by a different example. Imagine you are sitting across from me in my office, you point to a paperweight in plain view on my desk, and ask What is that paperweight made of?, and I respond It is made of wood. Although you don’t know, prior to my utterance, what the paperweight is made of, we both assume that, whatever it is made of, it is an essential property of that paperweight that it be made of that stuff. Since, in fact, the paperweight is made out of wood, my remark is an example of the necessary a posteriori. How would this conversation be represented in Stalnaker’s model of discourse? Prior to the utterance there would be different possible world-states in the context set that were compatible with everything assumed or established in the conversation up to that point. We may take these to include a context/world-state i in which the thing that, in i, is the one and only one paperweight on the desk is made of wood, a context/world-state j in which the paperweight on the desk in j is made of something else, e.g., plastic, and a context/world-state k in which a paperweight in front of us in k is made out of something else again—say, metal. In short, in Stalnaker’s model, the propositional concept PW would be associated with my utterance.

228 • Essay Eight

PW i j k

i T F F

j T F F

k T F F

The rules R1–R3 for assertion would then yield two conclusions: (i) that on hearing my utterance you had no way of knowing which proposition was (semantically) expressed by my sentence, because which proposition was expressed depended on which world-state—i, j, or k— actually obtained, and you didn’t know, in advance of accepting my remark, which world-state did obtain; and (ii) that none of the propositions that might have been expressed would have served a useful purpose. To have asserted a necessary truth would have been to have asserted something uninformative, and of no use in narrowing down the location of the actual world-state in the context set; and to have asserted a necessary falsehood would have been a nonstarter. So, if you were to regard my utterance as successful, you had to take it as asserting some proposition other than any of the candidates for being the one it (semantically) expressed. Which proposition might that have been? Since you knew that whatever the world-state of the context turned out to be, I would be committed to my remark being true, the proposition you must have taken me to have asserted is a proposition that is true (false) at a world-state of the context set iff the proposition expressed by my sentence at that worldstate is true (false) at that world-state. This is the diagonal proposition associated with PW. Since it is neither true at all world-states in the context set, nor false at them all, asserting it does the job that assertions are intended to do. Implicitly recognizing this, we both rightly understood the diagonal proposition to be the proposition I asserted. That is the explanation provided by Stalnaker’s model. There are two things wrong with it. First, it is wrong to suppose that you had any relevant doubt about what proposition was (semantically) expressed by my utterance of It is made of wood in response to your question, What [pointing at the paperweight] is that made of? The proposition I expressed is one that predicates being made of wood of that very paperweight—the one we both were looking at, and saw clearly sitting on the edge of my desk. You knew that it was the object you had asked about, and about which I had given an answer. Since you also knew what wood is, you knew precisely which property was predicated of which object by my remark. Surely, then, you did know the proposition my sentence expressed. In short, there was a proposition p such that you and I both knew that my utterance expressed p, even though you didn’t know, in advance of accepting my remark, whether or not p was

Understanding Assertion • 229

true, and so didn’t know whether or not p was necessary. Of course, given his identification of propositions with functions from metaphysically possible world-states to truth-values, Stalnaker can’t say this—since the fact that p is necessary would require him to say (i) that you knew p all along, and (ii) that you knew that my utterance expressed a trivial truth, simply by virtue of understanding it. Since this is absurd, he is forced to the patently counterintuitive conclusion that upon hearing my utterance, you didn’t know that it expressed p (where p is the proposition it actually did express). The second thing wrong with Stalnaker’s explanation is that the world-states j and k in the context set must either be (a) ones that are not really metaphysically possible, or (b) ones that are not compatible with all the shared assumptions of the conversational participants prior to my utterance—both of which are contrary to the dictates of the model. What are the world-states i, j, and k? They are total possibilities regarding how the world might be in which one and only one paperweight is sitting on my desk, seen by us, and the subject of our discourse. The paperweight satisfying these conditions in i is made of wood, whereas the paperweights satisfying them in j and k are made of plastic in one case and metal in the other. What paperweights satisfy these conditions in j and k? If j and k are really metaphysically possible, as Stalnaker insists, then the paperweights in j and k can’t be the paperweight that is really on my desk. Since that paperweight is made of wood in every genuinely possible world-state in which it exists, it is not made of plastic in j or metal in k. It follows that j and k must be world-states in which some other paperweight is between us on the desk, seen by us, and the subject of our conversation. But how can that be? Surely, one thing that was part of the shared conversational background prior to my remark was the knowledge that this very paperweight [imagine me demonstrating it again] was between us on the desk, seen by us both, and the subject of our conversation. To deny this would be tantamount to denying that we ever know, of anything we perceive or talk about, that it has one property or another. Even if we put the question of knowledge aside, surely we both believed these things about this very paperweight, which is all the model requires. But if we did have this de re knowledge, or these de re beliefs, then the discourse model’s requirement that the world-states in the context set be compatible with everything assumed and established in the conversation must have eliminated all metaphysically possible worldstates in which other paperweights, not made out of wood, were the one and only paperweight under discussion. But then, there is no room for the diagonalization required by Stalnaker’s explanation. This is the fundamental problem. Unless some persuasive defense can be found for excluding obvious, shared de re belief and knowledge from

230 • Essay Eight

the conversational model, Stalnaker’s explanation cannot succeed. I will argue that no such defense can be given. First, however, I will improve the model by liberalizing the notion of possibility it employs. Having strengthened the model so that it can accommodate cases involving essential properties of objects, like my paperweight, I will return to the problems posed by de re knowledge and belief, and investigate why they are intractable.

Improvements and Problems The model can be improved by dropping Stalnaker’s antecedent philosophical commitment to restricting epistemic possibility to metaphysical possibility. To drop this commitment is to recognize world-states that are metaphysically impossible but epistemically possible—i.e., maximally complete properties that the universe couldn’t really have had, but which we cannot know a priori that it doesn’t have (on analogy with properties that ordinary objects couldn’t have had, but which we cannot know a priori that they don’t have). When we allow context sets to include such world-states, the propositional concept associated with my utterance about the paperweight turns out to be different from the one we earlier took it to be. On this way of looking at things, i, j, and k are different epistemic possibilities involving the very same object, o—where o is the paperweight that we actually see on my desk, are talking about, and know that we are talking about. In world-state i, o is made of wood; in j, o is made of plastic; and in k, o is made of metal. The resulting matrix is PW*. PW* i j k

i T T T

j F F F

k F F F

Since the same proposition is expressed with respect to each epistemologically possible world-state, and since it is neither trivially true nor trivially false, no diagonalization is needed. This improvement encourages a certain thought. Perhaps Stalnaker’s model of inquiry can be divorced from the philosophically contentious motivations that partially inspired it. The idea is to give up the identification of epistemic possibility with metaphysical possibility, to give up the goal of explaining away the necessary a posteriori, and to give up the analysis of propositions as functions from metaphysically possible world-states to truth-values. We retain the idea that utterances are associated with propositional concepts or matrices, plus the general model

Understanding Assertion • 231

of discourse that makes use of these matrices. We also retain the idea that the point of a discourse is to narrow the set of the possibilities—now thought of as including both epistemic and metaphysical possibilities— within which the actual world-state is presumed to be located. As before, an assertion is supposed to shrink the set of possibilities compatible with everything that has previously been assumed or established in the conversation. On this new picture, the conversational rules R1–R3 remain intact. The model can be illustrated using the following example. I say, He is John Hawthorne (demonstrating a man sitting at the end of the table) in a conversation in which it is common knowledge that this man—the one I am talking about—is either John Hawthorne or Ted Sider. The utterance takes place in a context in which everyone knows a few facts about John and Ted already, but not everyone knows what they look like. Perhaps everyone has talked to each of them on the phone, or read the work of each, or corresponded with each, or some combination of the three, even though many would not recognize John or Ted by sight. Let us stipulate that everyone already knows of John that his name is ‘John’, that he is a Rutgers professor, and that he is not Ted—similarly for everyone’s antecedent knowledge of Ted. Moreover, this shared knowledge is known to be shared, and so the propositions known are part of the presupposed conversational background. In this situation I utter the sentence, He is John Hawthorne, demonstrating John, who is sitting at the end of the table. The sentence uttered contains a name, which, like the demonstrative he, is a rigid designator with respect to all possible worldstates—epistemic and metaphysical alike. What are the epistemic possibilities prior to my utterance? It might seem that the two most obvious possibilities—j and t—could be described as follows: in j there is a unique person sitting at the end of the table and that person is John, and in t there is a unique person sitting there and that person is Ted. This gives us the following matrix. j t

j T F

t T F

R1–R3 dictate that we perform the diagonalization operation, which gives us an asserted proposition that is true just in case John is sitting at the end of the table, and false otherwise. That is a good result, since it, or something quite like it, would normally be regarded as having been asserted by such an utterance. If you were to report my remark by saying Scott said that John Hawthorne was sitting there (gesturing to the place at the end of the table), I think most people would judge what you said to be true.

232 • Essay Eight

Nevertheless, the way we reached this result is problematic. Worldstates j and t are supposed to be epistemic possibilities compatible with everything taken for granted in the conversation prior to my remark. But, as in the earlier example about the paperweight, I left out of the specifications of j and t certain things known by all conversational participants. I ignored the fact that it was known (prior to my remark) that he [imagine me pointing again at John] was sitting there and also the fact that it was known (prior to my utterance) that since there weren’t two people sitting there, and because John and Ted are different people, if John was sitting there, then Ted wasn’t. When these things are added to the conversational background, t becomes incompatible with what is known or assumed by conversational participants, and so is excluded from the context set. Why? First, since it is known (prior to my utterance) that he [pointing at John] is sitting there, it follows that he, John, is an x, such that it is known that x is sitting there. This is just to say that the singular proposition p which says of John that he is sitting there is known to be true by the conversational participants, and so must be true with respect to t, if t is to be compatible with everything commonly known or assumed. Second, since it is known (prior to my remark) that if John is sitting there, Ted isn’t, it again follows that John is an x such that it is known that if x is sitting there then Ted isn’t. But it has already been stipulated that the proposition q that Ted is sitting there is true in t. Hence, t can be compatible with everything which is known or assumed in the conversation (prior to my utterance) only if the trio of propositions—p, q, and the conditional proposition the antecedent of which is p and the consequent of which is the negation of q—is consistent. Since this trio is inconsistent, t must be excluded from the context set—in which case our revised, Stalnaker-style explanation of what I asserted fails in a way similar to the way the original explanation of my assertion about the paperweight failed. How, then, is it that my utterance of He is John Hawthorne was informative? Since I discuss this sort of issue in considerable detail in chapters 3 and 4 of Soames (2002), I will deal with it only briefly here.13 We know that prior to the utterance my audience already believed of John that he was John. So the new belief acquired by virtue of accepting my utterance wasn’t that one. What might it have been? One such belief was surely that he, the person sitting there, was John Hawthorne. Every13 The account initially given in Soames (2002) of the relationship between the semantic content of a sentence S in a context C and what is asserted by uttering S in C is modified and extended in Soames (2005a); essay 9 in volume 1 of this collection.

Understanding Assertion • 233

one in the audience could see—without any appeal to propositional concepts or diagonalization—that I was attributing the property of being John Hawthorne to the guy sitting there, at whom I was pointing. So naturally I was committed to that being true. Moreover, if someone in the audience were to describe what I said to a third party who hadn’t been present, he might say, At first several of us didn’t know who was sitting at the end of the table, but then Scott said that John Hawthorne was the one sitting there. In ordinary life, such a report would be taken to be completely correct. If it is correct, then not only did I convey this informative proposition, I actually said (i.e., asserted) it. This is evidence that what I asserted went a little beyond the strict semantic content of the sentence I uttered in the context. In this respect, I agree with Stalnaker; in cases like this the speaker does assert a proposition which is not the proposition semantically expressed by the sentence he utters. But the mechanism by which this occurs is a rather ordinary one, and typically doesn’t involve any forced two-dimensionalist diagonalization.

The Nature of the Problem If what I have just argued is correct, then Stalnaker’s elegant model of discourse can be improved, but not saved, by liberalizing it to allow for epistemically possible world-states, over and above those that are metaphysically possible. The fundamental, mistaken assumption embedded in the model leading to its failure is that conversational participants can do two things: (i) identify, at the time of each utterance, precisely which possible world-states are compatible with everything previously assumed or established in the conversation; and (ii) determine which of these possible states are compatible with propositions expressed by the sentence we utter under different assumptions about which possible world-state actually obtains. In reality, we can’t always do these things—no matter whether the possible world-states in question are metaphysical or epistemic. We can’t do them because the relationship between sentences and the propositions they express is nontransparent in an important way. There are pairs of sentences S1 and S2, and contexts C, such that in C a.

b.

S1 expresses a proposition p1, S2 expresses p2, and speakerhearers understand both sentences, while knowing that to accept S1 is to believe p1 and to accept S2 is to believe p2, p1 bears some intimate “logical” relation to p2—e.g., p1 is the negation of p2, or p1 is identical with p2, or p1 is a conditional and p2 is its antecedent,

234 • Essay Eight

even though c.

speaker-hearers have no way of knowing that the relation mentioned in (b) holds between the proposition believed in virtue of accepting S1 and the proposition believed in virtue of accepting S2.

Because of this, there are cases in which speaker-hearers believe p1, and yet are in no position to recognize that in believing p2 they are believing something inconsistent with this, which—in terms of the model—rules out all epistemically possible world-states. In other cases, in which p1 and p2 are consistent, but some different relation holds between them, the fact that speaker-hearers believe both p1 and p2 may rule out some but not all possible world-states, without their being able to recognize which. Because of this nontransparency in the relationship between sentences, the propositions we believe (assert) in virtue of accepting (uttering) them, and the world-states in which these propositions are true, our beliefs and assertions cannot always interact with one another in the way the model presupposes. So, the model fails. The assumptions that lead to this result are modest. In order to reach our conclusion, one may, but need not, endorse the contentious, but I believe correct, doctrine that the semantic contents of names and indexicals (relative to contexts) are their referents, or the similarly contentious, but correct, doctrine that the semantic contents of natural kind terms are the kinds they designate. One reason these semantic assumptions are not needed is that we can generate corresponding problems for the model using pairs of synonymous expressions of other sorts—for example catsup/ketchup and dwelling/abode—where in each case a speaker can understand both expressions without realizing that they are synonymous.14 Another reason that contentious semantic assumptions are not necessary is that what generates problems for the model are not so much semantic facts about the sentences involved, as cognitive facts about speakers who use them. When an agent looks directly at the paperweight on my desk, and sincerely utters That [pointing at the paperweight] is the paperweight I am talking about, he is correctly described as believing, of the paperweight, that he is talking about it—where the proposition believed is also expressed by x is the paperweight I am talking about, relative to an assignment of the object itself to ‘x’. It is believing this proposition that creates trouble for the model, whether or not we identify it with the semantic content of the sentence uttered. Similar points hold for examples in which the sentence uttered contains a proper name or natural kind term. In all these cases, the propositions that prove problematic for 14

See Salmon (1990) and Rieber (1992).

Understanding Assertion • 235

the model are among those that conversational participants come to believe and assume at later stages of the conversation. Since these assumptions determine the context set for later utterances, the only hope of saving the model is to exclude beliefs of this sort from playing this role. However, there seems to be no reasonable way of doing this.

The Ubiquity of the De Re The fundamental reason that de re belief can’t be excluded from the model is that the model itself presupposes such belief. As I have stressed, it presupposes de re knowledge (or belief) of world-states, which, we have seen, is founded in de re knowledge (or belief) of individuals or kinds. This knowledge (or belief) of world-states is of three sorts: (i)

knowledge (belief) of world-states that they are, or are not, compatible with propositions previously assumed or established in the conversation, (ii) knowledge (belief) of world-states w in the context set, sentences S, and propositions p, that if w obtains then S expresses p, and (iii) knowledge of the truth-value of p in w, for each w in the context set. In each of these cases, the de re knowledge (belief) of world-states required by the model is inextricably linked to ordinary de re knowledge (belief) of individuals (or kinds). Hence there is no excluding the latter. Regarding (i), the propositions previously assumed or established in the conversation will standardly include singular propositions—knowledge of which amounts to de re knowledge of individuals (or kinds)—about the speaker and other conversational participants, the salient items in the context of utterance, and the various words in use plus their meanings. In many cases, the propositions previously assumed or established will also include singular propositions about the individuals (or kinds) which are topics of the conversation. It is not unusual for these singular propositions to be more readily available as commonly held assumptions of conversational participants than many of their purely descriptive counterparts. For example, there surely are cases in which conversational participants discussing an individual i each knows of i that they all know that i is the individual being talked about, and each know that it has been assumed or established that i has one or another property P—even though the descriptive information about i possessed by conversational participants varies so much from one participant to next that there may be few, if any, (purely qualitative) descriptions D that uniquely identify i which are known by each participant to be associated by all of them with any of the terms used

236 • Essay Eight

in the conversation. In such cases, the descriptive differences between the parties will wash away, and the most salient proposition about i known to be commonly assumed or established in the conversation may well be a singular proposition that predicates P of i. For reasons like these, there seems to be no way for a viable model of discourse to exclude singular propositions from the set of propositions commonly assumed or established in a conversation. The consequence of this for the de re knowledge of world-states of type (i) presupposed by Stalnaker’s discourse model is easy to see. If the propositions assumed or established include a proposition that says of o that it “is F”, then to know of an arbitrary world-state w whether it is compatible with what has been assumed or established in the conversation (and hence to know whether w is in the context set), one must know of w whether it is a world-state with respect to which o “is F”. It is not enough to know that w is a world-state in which whatever object satisfies a certain description (in the world-state in which the conversation actually takes place) “is F”; one must know of o itself that it “is F”, or that it “is not F”, in w. A similar point holds for the combination of (ii) and (iii), which generates the propositional concepts, or matrices, on which Stalnaker’s model is based. To generate these matrices, speaker-hearers must know, of each sentence S and pair of world-states w1 and w2 in the context set, whether the proposition that an utterance of S would express, were w1 to obtain, would be a true, or a false, description of w2. With this in mind, suppose, as Stalnaker does in several of his own examples, that S contains a rigid designator α which rigidly designates o1 if w1 obtains, and o2 if w2 obtains. Suppose further that S says of whatever is designated by α that it “is F”. Then, in order for the background knowledge and beliefs of speakerhearers to generate the propositional concept employed by Stalnaker, speaker-hearers must know of both o1 and o2 whether they “are F” in w1 and w2. This is ordinary de re knowledge and belief of those objects. Since the discourse model presupposes knowledge and belief of this sort, it cannot relegate this knowledge and belief to the sidelines.

Lessons For this reason, Stalnaker’s model of assertion fails. However, it is important not to overreact. Although certain aspects of the model, and the uses to which it was put, must be abandoned, other features of it can be retained. Among the former are Stalnaker’s revisionary account of the necessary a posteriori, his restriction of possible world-states to the metaphysically possible, and his identification of propositions with functions from such states to truth-values. Among the latter are versions of his

Understanding Assertion • 237

rules for assertion, R1–R3, suitably reformulated to avoid the model’s problematic features. The correct account of the necessary a posteriori is illustrated by the homely example of the paperweight on my desk, which I show you. You see it, pick it up and feel it, but can’t tell what it is made of. You imagine that it might be made of plastic, or metal, or wood. In imagining this, you are imagining the very object itself having the property of being made of plastic, being made of metal, or being made of wood. In so doing, you are also imagining different properties the universe might have—the property of containing this very object being made of plastic, the property of containing this very object being made of metal, and so on. You don’t know which, if any, of these properties the universe really does have. Since you can’t find this out by a priori reasoning alone, these properties, or more complete versions of them, are conceivable ways the world might be that are epistemically possible. When you finally learn that the paperweight is, in fact, made out of wood, you realize that it couldn’t have existed without being made out of wood, and so you realize that certain epistemically possible ways the world might be are not ways that it could genuinely have been, and so are metaphysically impossible. Since these ways are just world-states, this elementary Kripke-style example of the necessary a posteriori shows that certain epistemically possible world-states are metaphysically impossible. This example relies on a potentially contentious metaphysical doctrine— the essentiality of constitution. However, there is nothing special about the particular essential property chosen. Other essential properties or relations (e.g., the relation of nonidentity) would serve equally well. The important thing is simply that there be such properties (and relations). Given that there are, we may reason that just as there are properties (relations) that ordinary objects could possibly have had (stood in) and other properties (relations) they couldn’t possibly have had (or stood in), so there are certain maximally complete properties that the universe could have had—metaphysically possible states of the world—and other maximally complete properties that the universe could not have had— metaphysically impossible states of the world. Just as some of the properties that ordinary objects couldn’t have had are properties that one can coherently conceive them as having, and that one cannot know a priori that they don’t have, so some maximally complete properties that the universe could not have had are properties that one can coherently conceive it as having, and that one cannot know a priori that it doesn’t have. Given this, one can explain the informativeness of certain necessary truths as resulting (in part) from the fact that learning them allows one to rule out certain impossible, but nevertheless coherently conceivable, states of the world. Moreover, one can explain the function played by empirical

238 • Essay Eight

evidence in providing the justification needed for knowledge of necessary a posteriori propositions. Empirical evidence is required to rule out certain metaphysically impossible world-states which cannot be known a priori not to be instantiated, with respect to which these propositions are false. Thus, by expanding the range of epistemically possible states of the world to include some that are metaphysically impossible, one can accommodate Kripkean examples of the necessary a posteriori. This—rather than any Stalnaker-style diagonalization—is the correct account of the necessary a posteriori.15 Finally, there is the matter of salvaging what we can from Stalnaker’s rules governing conversation and inquiry. Stalnaker’s R1 is equivalent to R1NC. R1NC. A proposition asserted should never be a necessary consequence of, or necessarily inconsistent with, the set of propositions already assumed or established in the conversation. (A proposition p is a necessary consequence of a set S of propositions iff there is no metaphysically possible world-state w in which the members of S are all true, while p is not; p is necessarily inconsistent with S iff there is no metaphysically possible world-state in which p and the members of S are jointly true.) Since necessary consequences that are not a priori consequences are not, in general, discernable to agents, R1NC won’t do as a conversational maxim. Although R1AC is an improvement over R1NC, it won’t quite do, either. R1AC. A proposition asserted should never be an a priori consequence of, or a priori-inconsistent with, the set of propositions already assumed or established in the conversation. An important problem with R1AC is that some a priori consequences are highly unobvious, requiring intricate and arduous a priori reasoning to reach. For this reason, the assertion of such a consequence of propositions already assumed or established may be highly informative, and effective in furthering the purposes of the conversation. Thus, it makes sense to replace R1AC with R1OAC. R1OAC. A proposition asserted should never be an obvious a priori consequence of, or obviously a priori–inconsistent with, the set of propositions already assumed or established in the conversation. 15 For more on the distinction between metaphysically and epistemically possible worldstates, and its role in explaining the necessary a posteriori, see Soames (2005b, 198–209).

Understanding Assertion • 239

At this point, however, the effects of the nontransparency of belief (and assertion) must be faced. Imagine a case in which conversational participants accept a pair of utterances—This A is B and That A is C (separated by a brief span of time)—accompanied by a pair of demonstrations, each clearly indicating a given object o in full view of each of the parties, without it being recognized that the same object is demonstrated on both occasions. In such a case, the propositions assumed or established in the conversation shortly after the remarks have been accepted will include a singular proposition p that says of o that it “is both A and B”, and a similar proposition q that says of o that is “is both A and C”. With p and q in the conversational background, it would be a violation of the intent of R1OAC for any of the conversational participants to assert that nothing “is both A and B”, or that nothing “is both A and C”, since to do so would be to assert something obviously inconsistent with what has already been established. Similarly, it would not do to assert—without further ado—that something “is both A and B”, or that something “is both A and C”, since these are obvious a priori consequence of p, and q, respectively. However, it might be very informative, and not in the least inappropriate, for someone to assert of o that it “is A, B, and C”, even though the proposition thereby asserted is equivalent to the conjunction of p and q, or to assert that something “is A, B, and C”—even though the truth of that proposition is a trivial consequence of the conjunction of p and q. Similar remarks hold for certain assertions obviously inconsistent with the conjunction of p and q— assertions of o that it “is not A, B, and C”, or that nothing “is A, B and C”. Although such assertions would render the set of propositions accepted by conversational participants inconsistent, in cases in which such inconsistency is undetectable, there is no culpable violation of conversational rules. These considerations suggest reformulation R1OAC so as not to focus exclusively on the proposition asserted, the propositions assumed or established, and the logical relations between them. When singular propositions are involved, the ways in which propositions are presented or entertained are as important as which propositions are asserted, believed, or accepted. This can be accommodated by reformulating R1OAC along the following lines. RO.

An assertive utterance U should allow the conversational participants to correctly identify the proposition asserted, but U should never be such that it is obvious to conversational participants that the proposition asserted by U is a consequence of, or is inconsistent with, the propositions already assumed or established in the conversation.

240 • Essay Eight

Since a violation of RO will occur only when the speaker’s utterance presents the proposition asserted in a way that allows conversational participants to see that it is an obvious consequence of, or obviously inconsistent with, the propositions already assumed or established in the conversation, RO accommodates the nontransparency of belief and assertion that caused problems for R1OAC. Thus, RO, or something like it, offers the best hope we have for salvaging what was correct about Stalnaker’s rule R1. In requiring that conversational participants be able to identify the proposition asserted, RO also incorporates what was correct about R3. The insight embedded in the latter was, essentially, that one should be able to recognize what is asserted without settling the open questions remaining in the conversation. The reason RO is able to accommodate this point is that the sense in which it requires the proposition asserted to be identified, or recognized, is quite weak. What is required is simply that for some p, the conversational participants be able to correctly recognize that U is an assertion of p. It is not required that U present p to conversational participants in a way that allows them to determine the truthvalue of p in all genuine metaphysically possible world-states compatible with propositions already assumed or established, or even to accurately assess whether the truth of all those propositions would guarantee the truth, or the falsity, of p. It is only because Stalnaker built these unreasonable requirements of cumulative, global transparency into the discourse model that R3 was made to seem to incorporate a truth more farreaching than it really does. Similar remarks apply to R2. Is it important that U express a proposition, and that speaker-hearers be able to recognize that it does, without answering all the questions still open in the conversation? Of course it is, but this is already implicitly incorporated into RO. No reference to the different world-states in the context set is needed, since when the nontransparency of knowledge, belief, and assertion is accommodated, and propositions are no longer identified with functions from world-states to truth-values, these states fall away and the rules of discourse can best be stated directly in terms of structured propositions and assertive utterances. In many cases, there may be no set of world-states compatible with everything assumed or established, because that which has been assumed or established contains nontransparent inconsistencies. Even when there are no such inconsistencies, compatibility relations may be obscured by pockets of nontransparency among the propositions already accepted. To treat the set of world-states that are compatible with all these propositions as if it were a central component of the discourse model guiding the computations of speaker-hearers is to assume a global

Understanding Assertion • 241

and cumulative transparency in our beliefs, assertions, and knowledge that is simply not a part of our cognitive or conversational lives.16 If I am right about all this, then Stalnaker’s model of discourse must be drastically modified. Its central idea—that what is asserted in literal, nonmetaphoric speech often differs substantially from the semantic content of the sentence assertively uttered—is both true and important. However, the basic structure of the model—involving metaphysically possible world-states, propositions as functions from such states to truth-values, propositional concepts, and speaker-hearer calculations involving these items—cannot accommodate many of the facts about language use for which any acceptable theory of discourse must be responsible. I have indicated that an important core of truth can be salvaged from Stalnaker’s rules, R1–R3, governing assertion. However, this is only the beginning. There is much more to be said about the ways in which the propositions commonly assumed in the conversational background, together with the meaning and semantic content of the sentence uttered, contribute to the proposition, or propositions, asserted by the utterance. One important factor which I have not been able to talk about here differs sharply from anything in the Stalnaker model. This is the phenomenon of routine pragmatic enrichment of the semantic content of the sentence uttered, explored in Soames (2002) and (2005a). In the Stalnaker model, the assertion of a proposition other than the one semantically expressed by the sentence S that is assertively uttered is always forced by a conflict between the conversational background, general conversational rules governing assertive utterances, and what speaker-hearers take the meaning of S (and the context) to be. By contrast, I believe that the semantic content of S can be intimately related to the proposition asserted, without there being any general but defeasible presumption that the aim of a literal, assertive utterance of S is the assertion of the proposition that S semantically expresses (in the context). How fruitful this idea will prove to be remains to be seen. However, whatever success may be in store for us in the future will come on top of the progress made by the pioneering work done by Robert Stalnaker. No one has done more than he to open up this important field of investigation. 16 The second clause of R2, requiring propositions asserted to have truth-values, also goes by the board. Here there is no significant truth to be salvaged, since there is no conversational rule of the sort Stalnaker imagines against asserting propositions which cannot be assigned a truth-value—as opposed to those that are simply false or untrue. For discussion see Soames (1989, 583–89; essay 2 in volume 1 of this collection).

242 • Essay Eight

References Davies, Martin, and Lloyd Humberstone. 1980. “Two Notions of Necessity.” Philosophical Studies 38:1–30. Kripke, Saul A. 1979. “A Puzzle about Belief.” In Meaning and Use: Papers Presented at the Second Jerusalem Philosophical Encounter, April 1976, ed. Avishai Margalit, 239–83. Dordrecht: Reidel. Rieber, Steven. 1992. “Understanding Synonyms without Knowing That They Are Synonymous.” Analysis 52:224–28. Salmon, Nathan. 1990. “A Millian Heir Rejects the Wages of Sinn.” In Propositional Attitudes: The Role of Content in Logic, Language, and Mind, ed. C. Anthony Anderson and Joseph Owens, 215–47. Stanford, Calif.: Center for the Study of Language and Information. Soames, Scott. 1989. “Presupposition.” In Handbook of Philosophical Logic, vol. 4, Topics in the Philosophy of Language, ed. D. Gabbay and F. Guenthner, 553–616. Dordrecht: Reidel. ———. 2002. Beyond Rigidity: The Unfinished Semantic Agenda of “Naming and Necessity.” New York: Oxford University Press. ———. 2005a. “Naming and Asserting.” In Semantics vs. Pragmatics, ed. Zoltán Szabó, 356–82. Oxford: Clarendon Press; New York: Oxford University Press. ———. 2005b. Reference and Description: The Case against TwoDimensionalism. Princeton: Princeton University Press. Stalnaker, Robert. 1999. “Assertion.” In Context and Content: Essays on Intentionality in Speech and Thought, 78–95. New York: Oxford University Press. Originally published in Syntax and Semantics, vol. 9, Pragmatics, ed. Peter Cole, 315–22 (New York: Academic Press, 1978). ———. 2003. “On Considering a Possible World as Actual.” In Ways a World Might Be, 188–200. Oxford: Oxford University Press. Originally published in Aristotelian Society Supplementary Volume 75:141–74.

ESSAY NINE

Ambitious Two-Dimensionalism

Three decades ago, a group of philosophers led by Saul Kripke, Hilary Putnam, and David Kaplan ushered in a new era by attacking presuppositions about meaning that occupied center stage in the philosophy of the time. Among these presuppositions were the following: (i) The meaning of a term is never identical with its referent. Instead, its meaning is a descriptive sense that encodes conditions necessary and sufficient for determining its reference. (ii) Understanding a term amounts to associating it with the correct descriptive sense. Different speakers who understand a predicate of the common language, or a widely used proper name such as London, associate it with essentially the same sense. However, for many names of lesser-known things, the defining descriptive sense associated with the name varies from one speaker to the next. (iii) Since the meaning of a word, as used by a speaker s, is the descriptive sense that s mentally associates with it, meaning is transparent. If two words mean the same thing, then anyone who understands both should be able to figure that out by consulting the sense that he or she associates with them. For similar reasons, word meanings and mental contents are entirely dependent on factors internal to speakers. (iv) A priori and necessary truth amount to the same thing. If they exist at all, both are grounded in meaning. (v) Claims about objects having or lacking properties essentially— independent of how they are described—make no sense. Even if a term t designates o and Necessarily t is F is true, there will always be another term t* designating o for which Necessarily t* is F is false. Since it would be arbitrary to give either sentence priority in determining the essential properties of o, the idea that objects have, or lack, properties essentially, must be relativized to how they are described. (vi) Since the job of philosophy is not to come up with new empirical truths, its central task is that of conceptual clarification, which proceeds by the analysis of meaning.

244 • Essay Nine

These doctrines and their corollaries provided the framework for much of the philosophy in the analytic tradition prior to the 1970s. Of course, not every analytic philosopher accepted all tenets of the framework, and some, like Quine, rejected the traditional notions of meaning, necessity, and apriority altogether. However, even Quine—the framework’s most severe critic—believed that if the traditional notions make sense at all, then they must be related more or less along the lines indicated. What was, for the most part, absent was a recognition that all of these notions do make sense, and are important for philosophy, even though they are mischaracterized by the traditional framework. That changed with Kripke, Putnam, Kaplan, and the line of research growing out of their work. Today, all doctrines of the framework have been challenged, and replacements have been suggested. However, no new consensus has been reached. Although everyone recognizes the need to take into account the arguments of Kripke and his fellow antidescriptivists, some continue to believe that the traditional paradigm contained much that was correct, and that a new, more sophisticated version of descriptivism should be put in its place. Even those who reject the idea of a descriptivist revival, and want to push the antidescriptivist revolution further, have found the task of constructing a positive, nondescriptivist conception of meaning to be daunting. In short, the struggle over the legacy of the original challengers to descriptivism is far from over. Here, I will discuss an important part of that struggle. In the last 25 years a systematic strategy has grown up around a technical development called two-dimensional modal logic, for reviving descriptivism, reconnecting meaning, apriority, and necessity, and vindicating philosophy as conceptual analysis along recognizably traditional lines. Although the logical and semantic techniques are new, the motivating ideas are old. Since many of these ideas were not without plausibility, it is not surprising that an attempt has been made to reinstate them. But there is more to the attempted revival than this. Antidescriptivism has brought with it problems of its own. I will begin by explaining how these problems have motivated a cluster of views I call ambitious two-dimensionalism. I will then identify what I take to be the shortcomings of these views, and explain why I believe that no version of ambitious two-dimensionalism can succeed.

The Antidescriptivist Revolution First a word about the antidescriptivist revolution. In Naming and Necessity, Kripke offers two main arguments against the view that the meanings of names are given by descriptions associated with them by speakers,

Ambitious Two-Dimensionalism • 245

plus a further class of arguments against the view that the reference of most names is fixed in this way.1 The modal argument holds that since sentences containing proper names are true in different possible worldstates than corresponding sentences containing descriptions, names can’t mean the same as descriptions. The epistemological argument holds that if names really meant the same as descriptions, then certain sentences containing them would express a priori truths. Since these sentences don’t express such truths, names aren’t synonymous with descriptions. Finally, Kripke’s semantic arguments show that the descriptions speakers would volunteer in answer to the question To whom, or what, do you use ‘n’ to refer? sometimes fail to denote the object to which n really refers, and sometimes denote something to which n doesn’t refer. From this he concludes that, except in relatively rare cases, the linguistic rule mastered by speakers for determining the referent of a name is not that it is to refer to whatever is denoted by a set of associated descriptions. Instead, reference is determined by historical chains of reference transmission connecting later uses to earlier ones, and ultimately to initial baptisms introducing the name. A similar story is told for natural kind terms. To this David Kaplan added an account of indexicals.2 On this account, to know the meaning the first person, singular pronoun, is to know that one who uses it in a sentence— I am F —refers to oneself, and says of oneself that one “is F”. Similar rules govern other indexicals. These rules both tell us how the referents of indexicals depend on aspects of contexts in which they are used, and implicitly identify the semantic contents of indexicals with their referents. What is semantic content? The semantic content of a sentence is the proposition it expresses. Sentences containing indexicals express different propositions, and so have different contents, in different contexts. Nevertheless, the meaning of such a sentence is constant; it is a function from contexts to contents. Kaplan’s word for this is character. The picture is recapitulated for subsentential expressions. The character of the pronoun I is a function that maps an arbitrary context C onto the agent of C, which is its semantic content of I in C. There are two antidescriptivist implications here. First, the referents of at least some indexicals are not determined by descriptions speakers associate with them. One example involves Kaplan’s identical twins, Castor and Pollux, raised in qualitatively identical environments to be molecule for molecule identical and so, presumably, to associate the same purely qualitative descriptions with the same terms.3 Despite this, each twin refers to himself, and not the other, when he uses I. Although this 1

Kripke (1980); originally given as lectures in 1970, and originally published in 1972. Kaplan (1989). 3 Kaplan (1989, 531). 2

246 • Essay Nine

leaves open the possibility that some indexicals may have their referents semantically fixed by descriptions containing other indexicals, it precludes the possibility that all indexical reference is determined in this way. The second antidescriptivist implication is that since the semantic content of an indexical in a context is its referent, its content is not that of any description. The underlying picture is one in which the proposition expressed by S is a structured complex, the constituents of which are the semantic contents of the words and phrases of S. For example, the proposition expressed by the sentence I am F is a complex in which the property expressed by F is predicated of the agent of the context. This is the same proposition that is expressed by the formula x is F, relative to an assignment of o to ‘x’. A similar story is told for other indexicals. As Kaplan tells it, this story has consequences for propositional attitude ascriptions. Suppose, to adapt Russell’s famous example, that on some occasion in which George IV spied Walter Scott, he gave voice to his conviction, saying He [gesturing at Scott] isn’t the author of Waverly. Had he done this, the attitude ascription The author of Waverley, namely Scott, is such that George IV said that he wasn’t the author of Waverley would have been true—as would the ascriptions George IV said that you weren’t the author of Waverley. (said addressing Scott) George IV said that I wasn’t the author of Waverley. (said by Scott) On Kaplan’s picture, these reports are true because the semantic content of the sentence George IV uttered (in his context), and so a proposition he asserted, is the same as the content of the complement clauses in the reports of what he said. Whatever descriptions speakers who utter these indexical sentences may happen to associate with the indexicals are irrelevant to the semantic contents of the sentences they utter. When used in attitude ascriptions, indexicals, like variables in cases of quantifying-in, are used to report an agent’s attitude toward someone, or some thing, abstracting away from the manner in which the agent thinks of that person, or that thing. All they contribute to the proposition George IV is reported as asserting is the individual Scott. We may express this by saying that for Kaplan indexicals are not only rigid designators, but also directly referential. Some, including Nathan Salmon and me, extend this to proper names.4 So far, I have talked only about semantics. However, the new view of semantics is closely linked to a view that recognizes the contingent a 4

Salmon (1986); Soames (2002).

Ambitious Two-Dimensionalism • 247

priori and the necessary a posteriori. Since the latter will be most important for us, we will focus on it. Kripke’s route to the necessary a posteriori was simple. He first used the concept of rigid designation to rebut Quine’s objection to essentialism.5 Then, with both a nondescriptive semantics and a rehabilitated conception of essentialism in place, he showed how to generate instances of the necessary a posteriori. If n is a name or indexical that rigidly designates o, and P expresses an essential property of o which is such that knowledge that o has it requires empirical evidence, then the proposition expressed by If n exists, then n is P is both necessary and knowable only a posteriori.

Reasons for Descriptivist Revival So much for antidescriptivism. We now turn to what some take to be the grounds for a descriptivist revival. First is the conviction that antidescriptivists have not adequately addressed Frege’s puzzle about substitution in attitude ascriptions, and Russell’s problem of negative existentials. There is still a widespread belief that these problems show that names can’t be directly referential.6 Although Kripke never asserted that they were, it is hard to see how, if his doctrines are correct, they could be otherwise. According to him, the meaning of a name is never that of any description, and the vast majority of names don’t even have their referents semantically fixed by descriptions. If these names are so thoroughly nondescriptional, it is hard to see how their meanings could be other than their referents. Consequently, one who takes this Millian view to have been refuted by Frege and Russell may naturally suspect the power of Kripke’s arguments to have been exaggerated, and may be motivated to find a way of modifying descriptivism so as to withstand them. The second factor motivating descriptivists is their conviction that critics like Kripke have focused on the wrong descriptions. To be sure, it will be admitted, for many speakers s and names or natural kind terms n, the descriptions most likely to be volunteered by s in answer to the question  To whom, or what, do you refer when using n? neither give the meaning of n, nor semantically fix its reference. However, the referents of these terms must be determined in some way, and surely, whatever way that turns out to be is one which speakers have some awareness of, and which can be described. So, for each n, there must be some description that correctly picks out its referent—perhaps one encapsulating Kripke’s own historical chain picture of reference transmission. Some descriptivists even 5 6

For discussion of Quine’s objection and Kripke’s rebuttal see Soames (2003, 2:347–54). See Soames (2005, chap. 1).

248 • Essay Nine

go so far as to suggest that descriptive theories of reference are, for all intents and purposes, irrefutable.7 The idea is that any refutation would require an uncontroversial scenario in which n refers to some object o not satisfying the description D putatively associated with n by speakers (or in which n fails to refer to the thing that is denoted by D). However, the very judgment that n refers to o in this scenario (or doesn’t refer to what D denotes) is taken by these descriptivists to demonstrate the existence of a different, implicit, description in our minds that does determine reference in the scenarios—even though we can’t articulate it. The third factor motivating descriptivist revival involves the inability of some to see how any single proposition could be either both necessary and a posteriori, or both contingent and a priori, as antidescriptivists maintain. Again, we focus on the necessary a posteriori. How, some philosophers ask, can empirical evidence about the actual world-state be required to know p, if p is true in every possible state? Surely if such evidence is required, it must have the function of ruling out possible ways in which p could be false. But if p is true in every possible world-state, then there are no such ways to rule out. So, if p is knowable at all, p must be knowable a priori. The idea that p is both necessary and knowable only a posteriori is incoherent, and any nondescriptive semantics that says otherwise must be incorrect.8 What gives this reasoning force is a commitment to metaphysical possibility as the only kind of possibility. On this view, there are different metaphysically possible ways the world could be, but there are no further, epistemically possible ways that the world might be. There are no world-states which, though metaphysically impossible, cannot be known by us a priori not to obtain. This restriction of epistemic possibility to metaphysical possibility renders the necessary a posteriori problematic— since it precludes seeing it as involving metaphysically necessary propositions for which empirical evidence is needed to rule out metaphysically impossible, but epistemically possible, world-states in which they are false. When one adds to this the popular analysis of knowing p as having evidence that rules out all relevant possible ways of p’s being untrue, one has, in effect, defined propositions that are both necessary and knowable only a posteriori out of existence. A different philosophical commitment that leads to the same result identifies propositions with sets of metaphysically possible world-states.9 On this view, there is only one necessary proposition, which is known a priori. But then, the antidescriptivist 7

See, for example, Jackson (1998b, especially at 212). An analysis of knowledge with this consequence is given by Lewis (1999a). 9 See Stalnaker (1984). 8

Ambitious Two-Dimensionalism • 249

semantics that leads to the view that there are necessary a posteriori propositions must be mistaken.

The Two-Dimensionalist Strategy for Reviving Descriptivism I now turn to the strategy of descriptivist revival, which consists of three main elements: (i) the attempt to find reference-fixing descriptions withstanding Kripke’s semantic arguments, (ii) the rigidification of those descriptions to avoid the modal argument, and (iii) the use of twodimensional semantics to explain away putative examples of the necessary a posteriori and the contingent a priori. The most popular strategy for finding reference-fixing descriptions is causal descriptivism, which involves extracting a description from Kripke’s historical account of reference transmission. This idea is illustrated by the following passage from David Lewis. Did not Kripke and his allies refute the description theory of reference, at least for names of people and places? . . . I disagree. What was well and truly refuted was a version of descriptivism in which the descriptive senses were supposed to be a matter of famous deeds and other distinctive peculiarities. A better version survives the attack: causal descriptivism. The descriptive sense associated with a name might for instance be the place I have heard of under the name “Taromeo”, or maybe the causal source of this token: Taromeo, and for an account of the relation being invoked here, just consult the writings of causal theorists of reference.10 The second part of the descriptivists’ strategy is to rigidify referencefixing descriptions. The idea is to explain apparent instances of substitution failure involving coreferential names in attitude ascriptions by appealing to descriptive semantic contents of names; while using rigidification to guarantee substitution success when one such name is substituted for another in modal constructions. The final weapon in the descriptivists’ arsenal is ambitious two-dimensionalism, which may be illustrated using a putative example of the necessary a posteriori. (1)

If the x: actually Fx exists, then the x: actually Fx = the x: actually Gx.

Here, we let the F and the G be contingently codesignative, nonrigid descriptions. The semantics of actually guarantees that since (1) is true, it 10

Lewis (1999b, n. 22).

250 • Essay Nine

is a necessary truth. Nevertheless, the knowledge reported by (2) is seen as being, at bottom, nothing over and above the knowledge reported by (3). (2) y knows that if the x: actually Fx exists, then the x: actually Fx = the x: actually Gx (3) y knows that if the x: Fx exists, then the x: Fx = the x: Gx Since the latter is a posteriori, so is the former. How can this be?11 The two-dimensionalist answer is based on the relationship between the complement sentence, (1), of the ascription (2), and the complement sentence, (4), of the ascription (3). (4) If the x: Fx exists, then the x: Fx = the x: Gx (1) and (4) are nonequivalent in that they express different propositions, but equivalent in that they express truths in the same contexts of utterance. Since (4) expresses the same proposition p in all contexts, the two sentences express truths in all and only those contexts in which p is true. By contrast, (1) is semantically associated with two propositions—one which it expresses in our present context, and one which states the conditions a context must satisfy if (1) is to express a truth. The former, called the secondary proposition, is necessary, while the latter, called the primary proposition, is the contingent, a posteriori truth p that is also expressed by (4). How does this provide a two-dimensionalist explanation of the putative a posteriori status of (1)? Two main possibilities suggest themselves (between which informal discussions of two-dimensionalism often do not distinguish). The first arises from a semantic theory I call strong twodimensionalism. It holds that although the proposition expressed by (1) is a necessary truth, the knowledge reported by (2) is knowledge, not of this truth, but of the conditions under which (1) expresses a truth. On this view, there is no puzzle explaining how the proposition expressed by (1) can be both necessary and knowable only a posteriori, because it isn’t. Instead, the secondary proposition associated with (1) is relevant to its modal status, while its primary proposition is relevant to its epistemic status. All sentences that express different propositions in different contexts are seen as semantically associated with two propositions relative to any single context C: the proposition the sentence expresses in C (its 11 Although it is evident that the proposition expressed by (1) is knowable a posteriori, it is far less clear that it is knowable only a posteriori. For an argument that it is in fact possible to know it a priori (without knowing the proposition expressed by (4)), see Soames (2006), essay 8 of this volume. For more detail, see essay 10. Although I will return to this point below when assessing ambitious two-dimensionalism, in presenting the view I will temporarily take it for granted that examples like (1), containing the actuality operator, are genuine instances of the necessary a posteriori.

Ambitious Two-Dimensionalism • 251

secondary proposition relative to C) and the proposition that states the conditions that must be satisfied by any context in which the sentence expresses a truth (its primary proposition). The primary proposition associated with S provides the argument to the operators it is knowable a priori (or a posteriori) that, as well as Jones knows (a priori or a posteriori) that. The secondary proposition provides the argument to modal operators like it is a necessary truth that. This is the basis for the strong twodimensionalist’s claim that every instance of the necessary a posteriori can be explained along the same lines as (1). A slightly different explanation arises from a semantic theory I call weak two-dimensionalism. As before, sentences that express different propositions in different contexts are semantically associated with primary and secondary propositions. However, according to weak twodimensionalism, the argument provided by such a sentence to the operators it is knowable a priori (or a posteriori) that, and Jones knows (a priori or a posteriori) that is not its primary proposition, but its secondary proposition. On this view, (2) reports knowledge of (1)’s necessary, secondary proposition. However, this knowledge is counted as a posteriori because, it is claimed, (1)’s secondary proposition can be known only by virtue of knowing (1)’s primary proposition. As before, what makes the knowledge reported by (2) a posteriori is that (1)’s primary proposition is a posteriori. But whereas the strong two-dimensionalist claims that knowledge of the primary proposition is reported instead of knowledge of the secondary proposition, the weak two-dimensionalist claims that knowledge of the primary proposition counts as knowledge of the secondary proposition. A similar explanation is envisioned for all cases of the necessary a posteriori.

Varieties of Descriptive Two-Dimensionalism That is the basic idea behind the two-dimensionalist revival of descriptivism. I will now outline several different versions of two-dimensionalism in more detail. I begin with benign two-dimensionalism, which is the view that there are two dimensions of meaning—character and content. Character is a function from contexts of utterance to content, which in turn determines a function from circumstances of evaluation to extensions. Characters, which are sometimes called two-dimensional intensions, are, as David Kaplan taught us, crucial to the semantics of context-sensitive expressions. It is Kaplan who gave us benign two-dimensionalism. In his sense, “we are all two-dimensionalists now.” In recent years, however, the term two-dimensionalism has come to stand for something more ambitious—a cluster of views that attempt to use something like the

252 • Essay Nine

distinction between content and character to explain, or explain away, all instances of the necessary a posteriori and the contingent a priori. I will sketch four versions of ambitious two-dimensionalism—the pragmatic version of Robert Stalnaker’s 1978 paper, “Assertion,” the strong semantic version suggested in the mid-nineties by Frank Jackson’s From Metaphysics to Ethics and David Chalmers’s The Conscious Mind, a weak semantic version which is a natural retreat from strong two-dimensionalism, and a hybrid version suggested by Chalmers in 2002.12 Stalnaker’s Pragmatic Two-Dimensionalism In “Assertion,” Stalnaker accepted that Kripke’s examples of the necessary a posteriori express necessary truths that predicate essential properties of objects. He also recognized that a speaker who assertively utters one of them asserts something that is knowable only a posteriori. However, he maintained that in every such case the proposition asserted is contingent, and so not identical with the proposition semantically expressed by the sentence uttered. He believed he could show this by appealing to a pragmatic model of discourse. According to the model, conversations take place against a background of shared assumptions which rule out certain possible world-states as not obtaining, or “being actual.” As the conversation proceeds, new assertions acquire the status of shared assumptions, and the set of world-states compatible with what has been assumed or established shrinks. The aim of further discourse is to continue to shrink this set (called the context set) within which the actual state of the world is assumed to be located. The function of assertion is to eliminate from the context set all world-states in which the proposition asserted isn’t true. Stalnaker postulates three rules governing assertion. R1. A proposition asserted should always be true in some but not all members of the context set. R2. Any assertive utterance should express a proposition relative to each world-state in the context set, and that proposition should have a truth-value in each such state. R3. The same proposition should be expressed relative to each world-state in the context set. The rationale for R1 is that a proposition true in all world-states of the context set, or false in all such states, fails to perform the function of assertion—namely, to circumscribe the range of possibilities within which the actual world-state is located. Of course, this rule, like the others, allows for some flexibility in how it applies. If someone seems 12

Stalnaker (1999); Jackson (1998b); Chalmers (1996, 2002).

Ambitious Two-Dimensionalism • 253

to say something that violates it, one may sometimes conclude that no violation has really taken place because the context set isn’t quite what one originally thought, or because the speaker didn’t really assert, or mean, what he at first seemed to assert or mean. In such a case, a speaker may say something the literal interpretation of which would violate the rule, knowing full well that he will be reinterpreted in a way that conforms to it. Stalnaker’s rationale for R2 is that if an utterance violates it, then for some world-state w, the utterance won’t determine whether w should remain in the context set or not. In explaining the rational for R3, he employs his notion of the propositional concept associated with an assertion, which is related to Kaplan’s notion of the character of a sentence. For Stalnaker, a propositional concept is a function from world-states, considered as possible contexts of utterance, to propositions—where propositions are nothing more than assignments of truth-values to such states. The propositional concept associated with an utterance of S is a function that maps each world-state w of the context set onto such an assignment. This assignment can be thought of, roughly, as the proposition that would be expressed by S, if the context of utterance were to turn out to be w. D i j k

i T F F

j T F T

k T T T

Example D represents the propositional concept associated with a use of a sentence at a moment in which the context set consists of the worldstates i, j, and k. D tells us that if i is the state the world is actually in, then the proposition expressed by the speaker’s utterance is the proposition that assigns truth to every world-state of the context set, if j is the actual world-state, then the proposition expressed assigns truth to k and falsity to i and j, and if k obtains, the proposition expressed assigns falsity to i and truth to the other two. Which proposition is expressed depends on which world-state actually obtains. Since the conversational participants haven’t agreed on this, they won’t know which proposition the speaker is asserting, and so they will be at a loss as to how to update the context set. In this sort of case, we have a violation of R3. With these rules in place, Stalnaker is ready to explain assertive utterances of Kripkean examples of the necessary a posteriori. The key involves apparent violations of R3. He gives an example of hearing a woman speaking in the next room. I tell you, That is either Ruth Marcus or Judy Thomson. Since demonstrative pronouns and proper names are rigid designators, this sentence expresses either a necessary truth or a

254 • Essay Nine

necessary falsehood, depending on who is in the next room. Let i be a world-state in which the woman speaking is Ruth, j be a state in which it is Judy, and k be a state in which it is Mrs. Clinton. The propositional concept associated with this utterance is then E. E i j k

i T T F

j T T F

k T T F

E tells us two things: (i) we don’t know which proposition is actually expressed by the sentence uttered, because we don’t know which context actually obtains; (ii) none of the possible propositions expressed would serve a purpose in the conversation. Thus, E violates R3, and any attempt to avoid violation by excluding i, j, or k from the context set would violate R1. So, if we are to avoid violation entirely, and to regard the speaker’s utterance as informative, we must take it as asserting something else. What else? Well, whichever world-state turns out to obtain, the speaker will be committed to his utterance expressing a truth. Thus, we take the proposition asserted to be one that is true at any world-state w (of the context set) just in case the proposition expressed by the sentence in w is true at w. This is the assignment of truth-values that arises by looking along the diagonal in E to find the value that appears in row w of column w, for each w. Stalnaker calls this the diagonal proposition. Since it is neither true in all world-states of the context set nor false in all those states, it can do what asserted propositions are supposed to do. Hence, it is what is really asserted by the speaker’s utterance—no matter which member of the context set actually obtains. This is what ⇑E represents, where ‘⇑’ is an operator that maps a propositional concept C1 onto the propositional concept C2 that assigns each potential context the diagonal proposition of C1. ⇑E i j k

i T T T

j T T T

k F F F

This example is the prototype for Stalnaker’s treatment of the necessary a posteriori. Since the example is supposed to generalize to all such instances, one might get the impression that he thought that all genuine cases of the necessary a posteriori involve indexical sentences, which semantically express different propositions in different contexts. Since many examples of the necessary a posteriori involve only names or natural kind terms, such a view would require analyzing these terms as descriptions

Ambitious Two-Dimensionalism • 255

rigidified using either the actuality operator, or David Kaplan’s dthat operator. However, Stalnaker didn’t accept any such analysis. This created a problem for his pragmatic model. In order to get the desired results, he needed propositional concepts associated with instances of the necessary a posteriori to assign different propositions to different contexts. However, if the sentences aren’t indexical, the needed propositional concepts can’t be their characters, since these will be constant functions. But then, it is not evident where the needed propositional concepts will come from, or what interpretation they should be given once we have them. Although Stalnaker tried to finesse this issue, it was seen as a problem.13 Thus, it is not surprising that later two-dimensionalists took the step of analyzing names and natural kind terms as indexical, rigidified descriptions—thereby transforming the model from a pragmatic to a semantic one. The standard argument for taking this step goes like this: Imagine finding out that chemists have been wrong about the stuff that falls from the sky as rain, and fills the lakes and rivers. It is not really H2O but XYZ. What would water refer to, if this scenario were to turn out to be actual? XYZ, of course. Although in the world as it really is, we use water to rigidly refer to H2O, in a possible context in which the scenario described is actual, water rigidly refers to XYZ. Since the reference of water varies in this way from context to context, even though its meaning remains the same, the ambitious two-dimensionalist concludes that it is indexical. Thus, there must be some description implicitly associated with it by speakers that determines its reference in different contexts, which is then rigidified. It may be difficult to determine precisely what this description is, but that there is descriptive content to be rigidified is beyond question. Ditto for every name and natural kind term.14 Strong and Weak Two-Dimensionalism With this we move from Stalnaker’s pragmatic model to contemporary semantic versions of ambitious two-dimensionalism. I begin with two versions, which I will call strong and weak. Central Tenets of Strong Two-Dimensionalism ST1. Each sentence S is semantically associated with a primary intension and a secondary intension. Its primary intension is a proposition which is true with respect to all and only those contexts C to which the character of S assigns a proposition 13

See Soames (2006), essay 8 of this volume, for a discussion of this point. David Chalmers’ss use of this line of reasoning is discussed at length in Soames (2005, 209–28). 14

256 • Essay Nine

ST2.

ST3a.

ST3b. ST4a.

ST4b.

ST5a.

ST5b.

that is true at C. The secondary intension of S at C is the proposition assigned by the character of S to C. Understanding S consists in knowing its character and primary intension. Although this knowledge, plus relevant knowledge of a context C, would give one knowledge of the secondary intension of—i.e., the proposition expressed by—S in C, one does not always have such knowledge of C. However, this does not prevent one from using S correctly in C. All names and natural kind terms have their reference semantically fixed by descriptive properties which can, in principle, be expressed by descriptions not containing any (ineliminable) names or natural kind terms. These terms are synonymous with descriptions rigidified using dthat or actually.  It is a necessary truth that S is true with respect to a context C and world-state w iff the secondary intension of S in C is true with respect to all (metaphysically possible) world-states w* that are possible relative to w. Similarly for other modal operators.  It is knowable a priori that S is true w.r.t. C and w iff the primary intension of S in C is knowable a priori in w; x knows/believes (a priori) that S is true of an individual i w.r.t. C and w iff in w, i knows/believes (a priori) the primary intension of S in C. S is an example of the necessary a posteriori iff the secondary intension of S (in C) is necessary, but the primary intension of S is contingent and knowable only a posteriori. Such a sentence expresses a necessary truth in our actual context, while expressing falsehoods in other contexts. Its primary intension is not knowable a priori because we require empirical information to determine that our context is not one to which the character assigns a falsehood. S is an example of the contingent a priori iff the secondary intension of S (in C) is true, but not necessary, while the primary intension of S is necessary and knowable a priori. Such a sentence expresses a proposition which is false at some world-states, even though it expresses a truth in every context. The primary intension of such a sentence is knowable a priori because no empirical information is needed to determine that its character assigns one’s context a truth.

These theses carry with them certain more or less inevitable corollaries, including ST6a.

Ambitious Two-Dimensionalism • 257

ST6a. There is no proposition that is both necessary and knowable only a posteriori; nor is there any proposition that is contingent yet knowable a priori. If there were such propositions, then it should be possible to express them using nonindexical sentences, the primary and secondary intensions of which are identical (or at any rate equivalent). Since this is ruled out by ST5, the strong two-dimensionalist has reason to accept ST6a. A similar point holds for the corollary ST6b. ST6b. The necessary a posteriori and the contingent a priori are, in effect, linguistic illusions born of a failure to notice the different roles played by primary and secondary intensions in modal and epistemic sentences. In what follows, I will take strong two-dimensionalism to include ST6a and ST6b. One final thesis is ST7, which adds the claim that all necessary truths are knowable. ST7. A proposition is necessary iff it is knowable a priori. In systems that identify propositions with sets of possible world-states, ST7 is trivial, and already presupposed. In fact, the acceptance by a strong two-dimensionalist of ST7 would seem to go hand in hand with a possible-world-state analysis of propositions. Since it is difficult to imagine another conception of propositions that would justify the claim that all necessary truths are knowable, it is difficult to understand why a strong two-dimensionalist would adopt ST7, unless the theorist wished to adopt that analysis of propositions. Why would such a theorist find such an analysis congenial? Well, consider the standard strong twodimensionalist explanation of the contingent a priori, which hinges on the necessity of the sentence’s primary intension. How does the necessity of this proposition guarantee that it is knowable a priori, unless it is guaranteed that every necessary truth is knowable? And how is this guaranteed unless propositions are just sets of possible world-states? To the extent that two-dimensionalists want simply to take it for granted that necessity of primary intension is enough for a priori truth, they have reason to adopt the possible-world-state analysis of propositions, and with it ST7. In what follows, I will call systems incorporating ST1–ST6 strong two-dimensionalist, and those that also include ST7 very strong two-dimensionalist. I will take these latter systems to identify propositions with sets of possible world-states. Before giving a précis of weak two-dimensionalism, I pause over a point of interpretation. Although the analysis of attitude ascriptions given in

258 • Essay Nine

ST4 is crucial to strong two-dimensionalism, my best examples of strong two-dimensionalism—David Chalmers’s The Conscious Mind and Frank Jackson’s From Ethics to Metaphysics—do not explicitly include ST4b, or any other semantic analysis of attitude ascriptions. Worse, Chalmers repudiates ST4b in “Components of Content,” published several years later. Why, then, do I interpret the positions taken in those books as suggesting strong two-dimensionalism? Because they strongly suggest ST4, and because it is difficult without ST4 to maintain other theses to which Chalmers and Jackson are committed. Detailed textual criticism aside, the main interpretive points are these: (i) Chalmers and Jackson hold that S is an instance of the necessary a posteriori iff the primary intension of S is contingent (and hence a posteriori) while the secondary intension of S is necessary, and that S is an instance of the contingent a priori iff the primary intension of S is necessary (and hence a priori) while the secondary intension of S is contingent. (These claims are themselves treated as necessary.) (ii) Whenever S is an instance of the necessary a posteriori, they endorse It is not knowable a priori that S. Whenever S is an instance of the contingent a priori, they endorse It is knowable a priori that S. In general, they endorse It is knowable a priori that S iff the primary intension of ‘S’ is necessary, and they take it that (for any context C) the left-hand side expresses a truth (in C) iff the primary intension of S is necessary. (iii) From this it is reasonable conclude that, in their view (at the time), it is knowable a priori that operates on the primary intension of S, or if it operates on the primary intension of S plus something else, only the primary intension matters. (iv) There is no reasonable option to assuming that know and believe operate on whatever it is knowable a priori does. In effect, this adds up to the analysis given in ST4b.15 Next, I will sketch weak two-dimensionalism. Theses WT1–WT4a differ only in minor matters of detail from ST1–ST4a, and for our purposes may be assumed to come to essentially the same thing. The key differences are in WT4b and WT5.16 Central Tenets of Weak Two-Dimensionalism WT1. Each sentence S is semantically associated with a primary intension and a secondary intension. The former is its 15

For detailed discussions of the two works see Soames (2005, chaps. 8 and 9). The differences between strong and weak two-dimensionalism are discussed in more detail in Soames (2005, chap. 7). 16

Ambitious Two-Dimensionalism • 259

WT2.

WT3. WT4a. WT4b.

WT5a.

WT5b.

Kaplan-style character. The secondary intension of S at a context C is the proposition assigned by its primary intension to C. Understanding S consists in knowing its primary intension. Although this knowledge, plus relevant knowledge of the context C, would give one knowledge of the proposition expressed by S in C, one does not always have such knowledge of C. However, this does not stop one from using S correctly in C. As before. As before. An ascription x v’s that S, taken in a context C, is true of an individual A w.r.t. a world-state w iff there is some character M such that (i) in w, A bears R to M, and (ii) M assigns the secondary intension of S in C to a related context with A as agent and w as world-state. So propositions are objects of the attitudes, and attitude verbs express two-place relations between agents and propositions. However, this two-place relation holds between A and p in virtue of a three-place relation holding between A, a character, and p. To believe p is to accept a character M that expresses p (and believe that M expresses a truth). To know p is to justifiably accept a character M that expresses p (and know that M expresses a truth). For all necessary propositions p, p is knowable only a posteriori iff (i) p is knowable by virtue of justifiably accepting some meaning M (and knowing that M expresses a truth)— where (a) M assigns p to one’s context, (b) M assigns a false proposition to some other context, and (c) one’s justification for accepting M (and believing M to express a truth), requires one to possess empirical evidence—and (ii), p is knowable only in this way. For all contingent propositions p, p is knowable a priori iff p is knowable by virtue of justifiably accepting some meaning M (and knowing that M expresses a truth)—where (a) M assigns p to one’s context, (b) M assigns a truth to every context, and (c) one may be justified in accepting M (and believing M to express a truth) without empirical evidence.

Although some of the differences between strong and weak twodimensionalism are subtle and far-reaching, others are obvious. For example, whereas the strong and the weak two-dimensionalist agree that necessary a posteriori sentences are always associated with two

260 • Essay Nine

propositions—one necessary and one contingent—the weak twodimensionalist maintains that the necessary proposition is itself knowable only a posteriori. Because of this, the weak two-dimensionalist cannot identify propositions with sets of possible world-states. Thus, although strong and weak two-dimensionalists both presuppose that whenever a primary intension is necessary, it is knowable a priori, the weak twodimensionalist does not have the ready explanation of why this should be so that the very strong two-dimensionalist has. For the latter, it is in the nature of the one necessary proposition that it should be knowable a priori. Since the weak two-dimensionalist cannot say this, a different explanation is needed—of what sort is not obvious. This is one way in which weak two-dimensionalism is initially less attractive than strong twodimensionalism. However, this initial disadvantage is not decisive, since all three versions of ambitious two-dimensionalism face crippling problems which require their rejection. I will sketch a few of these problems, before turning to the final version of two-dimensionalism to be considered.

Problems with Ambitious Two-Dimensionalism Critique of Stalnaker’s Pragmatic Two-Dimensionalism Here is one problem for Stalnaker’s pragmatic account of the necessary a posteriori. I hold up my briefcase. You look at it closely, and ask What is it made of? Is it cow leather, some other kind of leather, vinyl, or what? I answer, It is made of cow leather. Let’s assume that although you don’t know, prior to my utterance, what the briefcase is made of, we both take it for granted that, whatever it is made of, it is an essential property of this briefcase that it be made of that stuff. Since it is, in fact, made of cow leather, my remark is an example of the necessary a posteriori. How would Stalnaker represent the conversation? Well, prior to the utterance he would have different possible worlds-states in the context set that are compatible with everything we assumed or established up to then. Presumably, these would include a context i in which the one and only briefcase I am holding is made of cow’s leather, a context j in which the briefcase I am holding is made of something else, say, pigskin, and context k in which I am holding a briefcase made of something else again—vinyl. So, he would associate my remark with the propositional concept B. B i j k

i T F F

j T F F

k T F F

Ambitious Two-Dimensionalism • 261

His rules for assertion would then yield two conclusions: (i) that on hearing my utterance you had no way of knowing which proposition was expressed, because you didn’t know which context—i, j, or k—actually obtained, and (ii) that none of the propositions assigned by B to these world-states would have served a useful purpose. To have asserted a necessary truth would have been uninformative, and to have asserted a necessary falsehood would have been a nonstarter. So, if you were to regard my utterance as informative, you had to take it as asserting some proposition other than any of the candidates assigned to members of the context set by B. Since you knew that whatever the real context turned out to be, I would be committed to my remark expressing a truth, you took me to have asserted the diagonal proposition, which is true at any of i, j, or k iff the proposition B assigns to that world-state is true when evaluated at that state. Since this proposition is neither true at all the world-states, nor false at them all, asserting it does the job that assertion is intended to do. That is Stalnaker’s account. There are two things wrong with it. First, it is wrong to suppose that you had any relevant doubt about what proposition was expressed by my utterance of It is made of cow leather. The proposition I expressed is one that predicates being made of cow leather of one particular briefcase—the one I was holding. You knew it was the object you had asked about, and about which I gave my answer. Since you also knew what cow leather was, you knew precisely which property was predicated of which object by my remark. How, then, could you have been in any real doubt about which proposition my sentence expressed? The second thing wrong with the explanation is that world-states j and k in the context set must either be ones that are not really metaphysically possible (contrary to the assumptions of the model), or ones that are not compatible with all the shared assumptions prior to my utterance (also contrary to the model). What are the world-states i, j, and k? They are total possibilities regarding how the world might be in which I am holding one and only one briefcase, which is both seen by us and the subject of our discourse. The briefcase satisfying these conditions in i is made out of cow leather, whereas the briefcases satisfying them in j and k are made out of pigskin and vinyl, respectively. Which briefcases are these in j and k? If j and k really are metaphysically possible, as Stalnaker insists, then the briefcases there can’t be the briefcase I was really holding when I answered your question. Since that briefcase is made out of cow leather in every genuinely possible world-state in which it exists, it is not made out of pigskin in j, or vinyl in k. It follows that j and k must be world-states in which I am holding some other briefcase. But how can that be? Surely, one thing we both knew prior to my remark was that I was holding this very briefcase (imagine me holding it up again now), which we saw and were talking about. But if we did have this de re knowledge, or these de re

262 • Essay Nine

beliefs, then Stalnaker’s requirement that the world-states in the context set be compatible with everything assumed and established in the conversation must have eliminated all metaphysically possible world-states in which other briefcases were under discussion. But if that is right, then there is no room for the diagonalization required by his explanation, and the account fails. This example relies on a plausible, but potentially contentious, metaphysical doctrine—the essentiality of origin, or constitution. However, there is nothing special about the particular essential property chosen. Other essential properties or relations (e.g., the property of being nonidentical with Saul Kripke, or the relation of nonidentity itself) would serve equally well. The important thing is simply that there be such properties (and relations). Given that there are, we can reconstruct different versions of the same problem that don’t rely on assumptions about material constitution. For example, if, pointing at a man, David Kaplan, whom we both clearly see and know we have been talking about, you ask Is he Saul Kripke? and I reply No, he isn’t Saul Kripke, what I say is knowable only a posteriori, even though the proposition expressed by my sentence is a necessary truth. However, if we try to apply Stalnaker’s two-dimensionalist model of discourse to this example, we will run into precisely the same problems we did in the example about my briefcase. Instances of the necessary a posteriori that hinge on the essential properties of things cannot, in general, be explained by the model. Notice what would happen if we dropped one of the antecedent philosophical commitments used in constructing the model—the restriction of the epistemically possible to the metaphysically possible. The idea is to allow the context set to include world-states that are metaphysically impossible, but epistemically possible—i.e., maximally complete properties that the world couldn’t really have had, but which we cannot know a priori that it doesn’t have (on analogy with certain properties, like being made of vinyl, that my briefcase couldn’t really have had, but which one can’t know a priori that it doesn’t have). When we allow such worldstates, the propositional concept associated with the utterance turns out to be different from what we originally took it to be. On this way of looking at things, i, j, and k are different epistemic possibilities involving the very same object—the briefcase I was actually holding. In i, it is made of cow leather; in j, of pigskin; and in k, of vinyl. This gives us the propositional concept B*. B* i j k

i T F F

j T F F

k T F F

Ambitious Two-Dimensionalism • 263

Since the same proposition is expressed with respect to each possible context, and since that proposition is neither trivially true nor trivially false, diagonalization is irrelevant.17 Critique of Strong Two-Dimensionalism The most obvious problem for strong two-dimensionalism is that its account of the semantics of attitude ascriptions is incorrect. When ordinary indexicals like I, now, and he are involved, their secondary intensions in contexts are crucial for attitude ascriptions in which they occur in the complement clause. Since the thesis ST4b doesn’t make room for this, it clearly fails. However, strong two-dimensionalism can be shown to be false, even if these indexicals are put aside. Here is one argument. Let n be a name. If strong two-dimensionalism were correct, then for some F, n would be synonymous with the actual F, and hence would have the same primary intension as the F —in which case, the primary intensions of n is G and the F is G would be the same. If, in addition, attitude ascriptions reported relations to the primary intensions of their complement clauses, then (5a) and (5b) would have the same secondary (as well as primary) intensions, and so would be necessarily equivalent. This would mean that (6a) and (6b) couldn’t differ in truthvalue. (5)

a. A believes (knows) that n is G b. A believes (knows) that the F is G (6) a. If it had been the case that . . . , then it would (or would not) have been the case that . . . A believed (knew) that n was G . . . . ] b. If it had been the case that . . . , then it would (or would not) have been the case that . . . A believed (knew) that the F was G....] Of course, pairs of the form (6a) and (6b) can easily differ in truth-value, as is shown by (6aa, 6bb). (6) aa. Although John truly believes (knows) that n is F, had the world been in state w, n would not have been F and John would not have believed (known) that n was F. bb. Although John truly believes that the F is F, had the world been in state w, the actual F would not have been F and John would not have believed (known) that the F was F. 17 Stalnaker’s two-dimensionalist account of assertion, the objections to it, and the distinction between aspects of it that must be abandoned and those that can be retained are discussed at greater length in Soames (2006), which is essay 8 of this volume.

264 • Essay Nine

The lesson here is that since n is rigid, its referent o is such that, at any world-state w, n is F is true iff at w it is a fact that o has the property expressed by F; thus, it ought to be the case that, for any world-state w,  John’s belief that n is F stands for a belief about o—one that is true at w only if, at w, o has the property expressed by F. Incredibly, on strong two-dimensionalism this turns out not to be so; instead the fact that n is F and the belief that n is F are, wrongly, allowed to be about different individuals. Here is another argument that makes a similar point. Where n is a name that refers to o, for any world-state w, if John believes (knows) that n is G is true with respect to w, then in w John believes (knows) of o that it “is G”. In other words, when n designates o, (7b) must be true, if (7a) is. (7) a. In world-state w [A believes (knows) that n is G] b. In world-state w [A believes (knows) that x is G ] (relative to an assignment of the referent of n to ‘x’) Strong two-dimensionalism can’t capture this, since it wrongly allows the ascription A believes (knows) that n is G to be about different individuals with respect to different world-states. For reasons like these, it is clear that strong two-dimensionalism is false.18 Critique of Weak Two-Dimensionalism Although the issues here are somewhat more complex, in the interest of economy I will be brief.19 There are two main points to make. First, the arguments that the reference of names and natural kind terms must be semantically fixed by description are defective, and do nothing to rebut the original Kripkean arguments that the vast majority of such terms do not have their referents fixed in this way. Second, even if we had reference-fixing descriptions, there would be no way of rigidifying them that would meet the needs of the weak two-dimensionalist. Regarding the first point, about the lack of reference-fixing descriptions, one must distinguish (a) foundational facts that bring it about that terms acquire (and retain) the semantic properties they have from (b) semantic facts about them, known by competent speakers, that constrain what they refer to—in the way that being constrained to refer to a female is part of what must be known in order to understand the demonstrative she. When it is claimed by Jackson, Chalmers, and others 18 19

For additional arguments, see Soames (2005, chap. 10). For a more complete discussion see Soames (2005, chap. 10).

Ambitious Two-Dimensionalism • 265

that the reference of names and natural kind terms must be fixed by description, they are mixing up these two things. For example, when Jackson claims that there must be reference-fixing descriptions associated with names and natural kind terms, because, as he puts it in From Metaphysics to Ethics,20 “it isn’t magic” that they refer to what they do, he is pointing to the fact that these expressions—like all others—get to have the semantic properties they do in virtue of some describable empirical features of how they are used. Of course, they do; and, of course, being language users, we are not entirely ignorant of what these features are. However, it doesn’t follow that we have complete and accurate knowledge of them. Still less does it follow that such knowledge is part of what a speaker must understand in order to understand the expression. To understand a word you have to know what it means—which in certain cases may amount to knowing its reference—but you don’t have to know how the word got to have the meaning or reference it does. So a guarantee that there is some accurate description of how a word got to have its meaning and reference does not provide any guarantee that the word has its reference semantically fixed by a description grasp of which is required for linguistic competence. This point bears on the suggestion, frequently made by contemporary descriptivists, that what Kripke did in giving his historical-chain account of reference transmission was simply to offer his own version of the description theory of reference-fixing.21 The idea behind this suggestion is, in simplest terms, that the reference of a name n for a particular speaker is determined by something like the following description: the individual which the person or persons from whom I acquired n referred to when they used the name. One of the things that makes this idea seem plausible is the requirement, recognized by Kripke, that in order for a reference-determining chain to be created by passing a name from speaker to speaker, the person acquiring the name must intend his reference to be parasitic on the reference of his source(s). The descriptivist proposes simply to state this requirement in terms of a reference-fixing description. Although the attraction of this idea is evident, when one examines it more closely one finds ample reason to be skeptical. It is not clear that speakers invariably have in mind, among all the different descriptions they associate with a given name, some accurate and precise referencefixing description for it. Certainly, the description the referent of the name as it was used by the first person from whom I acquired it does not 20 21

Jackson (1998a, 82). See, for example, the passage from Lewis (1999b, n. 22) referenced in n. 10 above.

266 • Essay Nine

always pick out my present referent for it.22 Since theorists have yet to develop a complete, accurate, and fully explicit historical-chain theory from which the needed description could be extracted, it is not clear which, if any, such description would be sufficient to handle all problematic cases. Even if theorists were to come up with the desired theory, it is far from obvious that ordinary speakers must have had it at their disposal all along, whenever they used a proper name. We may assume that there is some process of reference-borrowing by which later uses of a name inherit their reference from earlier uses. We may also assume that speakers know that some such process exists, even though they don’t know precisely how it works. With this in mind, consider the parasitic, reference-borrowing description D. D.

the thing referred to by those uses—whatever they turn out to be—of the name n from which my present use of n inherits its reference.

Since it is plausible to suppose that speakers know, at some level, that the reference of their uses of names is inherited from the use of these names by other speakers—even though they don’t know the precise mechanism by which this takes place—it is reasonable to suppose that they implicitly associate D, or something very much like it, with each name n that they use. However, it is important not to be confused by this. Although D does denote the referent of the speaker’s use of n, the reference of n is not semantically fixed by D. If it were, then satisfaction of D would be the mechanism by which the speaker’s use of n acquired its reference. But this can’t be the way in which reference is determined, since D itself presupposes that the speaker’s use of n already has a reference, which has been inherited from something else. In order for D to correctly pick out the referent, the speaker’s use of n must have acquired that referent in some other way. Thus, the mechanism by which its reference is fixed can’t be via satisfaction of D. The upshot of all this is that even though the reference-borrowing facts of Kripke’s historical-chain account of reference transmission play a foundational role in determining the reference of names and natural kind terms for speakers, there is no reason to think that these facts are among 22 Suppose, for example, that the person from whom I first picked up the name Plato was talking about his neighbor, whom he believed to be very wise. Suppose further that after speaking to this person I had many other conversations in which the name was used to describe Socrates’ famous biographer, whom I later started to read about under the name Plato. All of this could be true, even if I wrongly assumed that the person from whom I first heard the name was talking about the same individual as everyone else. In this sort of case, I do not refer to the neighbor of my original source when I use the name Plato. Instead, I refer to the famous philosopher.

Ambitious Two-Dimensionalism • 267

the semantic facts that must be mastered by competent speakers. Since this is precisely what would be required if the referents of these terms were semantically fixed by description, the description theory of reference fixing remains unsupported. Similar considerations can be used to scotch a familiar twodimensionalist objection to Kripke’s semantic arguments against the description theory of reference fixing. These arguments are based on thought experiments which show that often, in different counterfactual circumstances of use, a name or natural kind term n refers to something other than what is denoted by the descriptions that speakers explicitly associate with n, and that they would offer in answer to the question  To whom, or what, do you use ‘n’ to refer? From this, Kripke concludes that n does not have its reference fixed by descriptions associated with it by speakers. On the contrary, the two-dimensionalist objects, Kripke’s thought experiments undermine his conclusion by covertly presupposing unarticulated reference-fixing descriptions that implicitly guide speakers’ judgments. Here is Frank Jackson. Our ability to answer questions about what various words refer to in various possible worlds, it should be emphasized, is common ground with critics of the description theory. The critics’ writings are full of descriptions (descriptions) of possible worlds and claims about what refers, or fails to refer, to what in these possible worlds. Indeed, their impact has derived precisely from the intuitive plausibility of many of their claims about what refers, or fails to refer, to what in various possible worlds. But if speakers can say what refers to what when various possible worlds are described to them, description theorists can identify the property associated in their minds with, for example, the word ‘water’: it is the disjunction of the properties that guide the speakers in each particular possible world when they say which stuff, if any, in each world counts as water. This disjunction is in their minds in the sense that they can deliver the answer for each possible world when it is described in sufficient detail, but it is implicit in the sense that the pattern that brings the various disjuncts together as part of the, possibly highly complex, disjunction may be one they cannot state.23 Jackson maintains that our ability to identify the reference of terms in imagined Kripkean scenarios presupposes a knowledge of complete and accurate reference-fixing descriptions that is adequate for every scenario. He takes our ability to make these Kripkean judgments to demonstrate that names and natural kind terms must have their reference semantically 23

Jackson (1998b, 212).

268 • Essay Nine

fixed by these descriptions, that competence with the terms requires mastery of the descriptions, and that the terms themselves have the semantics of (indexical) rigidified descriptions. However, the ability he points to— such as it is—is not sufficient to draw these conclusions. First, there are clear cases in which we have no trouble identifying the referent of a term t, even though it is clear that there is no referencefixing description associated with t by speakers. David Kaplan’s example of the identical twins, Castor and Pollux, discussed earlier, is a case in point. We have no trouble identifying Castor as the referent of his use of I, and Pollux as the referent of his, just as we have no trouble recognizing ourselves as referents of our own uses. This is so despite the fact that the referent of I is not semantically fixed, for any of us, by descriptions we semantically associate with it. If this is true of I, it is surely also true of now, and may be true of other expressions as well. Second, even in cases in which there may be descriptions picking out the referent of a term that are, in some sense, associated with it by speakers, it remains to be shown these descriptions play any role in its semantics. One can describe possible scenarios in which our intuitions tell us that speakers use the word and to mean disjunction, the material conditional, the property of being a necessary truth, or the property of being a philosopher. Even if one were to grant the assumption that these intuitions arise from an internalized theory T that unconsciously guides us, it would not follow that the meaning of and—its character in Kaplan’s sense—is one that yields as content in a context whatever satisfies the relevant description extractable from T. Surely not every word is a descriptive indexical in Kaplan’s sense. To miss this point is to miss the distinction noted earlier between (a) foundational facts that bring it about that terms acquire (and retain) their semantic properties and (b) semantic facts about them, known by competent speakers, that constrain what they refer to. Whereas the descriptivist needs reference-fixing descriptions arising from (b), Jackson’s argument can’t exclude the possibility that the only relevant descriptions are those arising from (a). Finally, the claim that our ability to categorize cases in certain ways presupposes the sort of underlying knowledge required by the description theory is tendentious in something like the way that Plato’s attribution of a priori knowledge of mathematics to the slave boy in the Meno is tendentious. There are other ways to explain the recognition of new facts. The failure of Jackson’s argument on this point is representative of the arguments of two-dimensionalists generally on the subject of descriptive reference fixing. Typically these arguments do not involve precise and detailed proposals about the descriptions that allegedly fix the referents of specific names or natural kind terms. Instead, they attempt to establish, on the basis of very general considerations applying to all such terms,

Ambitious Two-Dimensionalism • 269

that their reference simply must be fixed descriptively—even if we are not in a position to articulate the crucial descriptions themselves. As the discussion of Jackson has illustrated, these arguments fail for a number of reasons, most notably the failure to distinguish foundational from semantic facts. Once this is recognized, and the Jackson-style arguments are out of the way, full weight can be given both to Kripke’s refutations of particular proposals for treating specific descriptions as semantically fixing the reference of particular terms, and to the fact that complete, accurate, and reliable descriptions, known by speakers to fix the reference of typical names and natural kind terms, have not been formulated by anyone. The next point to be made is that even if reference-fixing descriptions were generally available for names and natural kind terms, the weak two-dimensionalist has no good way of rigidifying them. One of two standard ways of rigidifying a description is by adding the actuality operator. However, this is no help to the weak two-dimensionalist, since it is obvious that neither names nor natural kind terms are equivalent to descriptions rigidified using this operator. For example, Aristotle isn’t equivalent to the actual F, for any F—since it is possible for agents in other possible world-states to know or believe that Aristotle was a genius without knowing or believing anything about our actual world-state— whereas it is not possible for those agents to know or believe that the actual so and so was a genius without knowing or believing anything about our actual world-state. In the presence of the defining assumptions of weak two-dimensionalism, this means that the propositions expressed by (i.e., the secondary intensions of) Aristotle was a genius and The actual F was a genius cannot be the same, no matter what F one chooses.24 Thus, if the weak two-dimensionalist is going to take names and kind terms as rigidified descriptions, he will have to take them to be descriptions rigidified using David Kaplan’s dthat operator.25 However, this choice is also problematic. For one thing, it makes weak two-dimensionalists Millians about semantic contents—and hence subject to the very problems posed by Frege’s puzzle and Russell’s problem of negative existentials that descriptivists standardly take to refute Millianism, and to motivate descriptivism in the first place. In addition, the weak two-dimensionalist explanations of the necessary a posteriori and the contingent a priori go disastrously awry if one holds (in accord with WT4b) that understanding and 24

This argument is elaborated in Soames (2002, chap. 2). See Soames (2005, 303–6). Another reason for weak two-dimensionalists to forsake analyses involving the actuality operator is directly connected to the point made in n. 11. Given an instance, n ≠ m, of the necessary a posteriori, one cannot analyze the names n and m as the x: actually Nx and the x: actually Mx, unless the x: actually Nx ≠ the x: actually Mx is knowable only a posteriori—something called into question by the argument mentioned there. 25

270 • Essay Nine

justifiably accepting dthat [the F] is G is sufficient for knowing the (Russellian) proposition it expresses. It is easy to show that, on this view, the necessary a posteriori will shrink to the vanishing point, and the contingent a priori will bloat to include virtually all contingent propositions involving individuals or kinds.26 To block this reductio ad absurdum, one must add a condition— knowing de re of the denotation of the dthat-rigidified description that it is so denoted—to what is required in order for understanding and justifiably accepting dthat [the F] is G to count as knowing p, where p is the proposition it expresses. But then there is a further difficulty. The de re knowledge mentioned in the condition is itself knowledge of a singular proposition. How does it arise—for all names and natural kind terms for which dthat-rigidified descriptions are posited as analyses by weak twodimensionalists, i.e., for every name and kind term? Since this knowledge can’t typically be explained as the result of understanding and ( justifiably) accepting still further indexical sentences, some additional, nontwo-dimensionalist, explanation of the crucial de re knowledge and belief must be given. If we had such an explanation, however, we could, presumably, apply it directly to paradigmatic Kripkean examples in which we know de re of an object that it has an essential property, even though this knowledge can only be a posteriori. But then we have an instance of the Kripkean necessary a posteriori that can’t be forced into the weak twodimensionalist mold, and the weak two-dimensionalist account of the necessary a posteriori is subject to a falsifying counterexample.27

A Hybrid View Confronted with these objections to existing versions of ambitious twodimensionalism, one might wonder whether there is some other version that is immune to the problems we have found. I don’t think there is. However, the matter is not easily settled, since it is not clear what modifications of two-dimensionalism are possible without abandoning essential features of the program. I will say a word about this by considering a proposal put forward by David Chalmers, in his 2002 paper, “The Components of Content,” where he sets his sights on solving Frege’s puzzle by treating names and natural kind terms as rigidified descriptions, and making attitude verbs sensitive to both primary and secondary intensions. The new approach implicitly rejects the strong twodimensionalist analysis of attitude ascriptions given by ST4b and the 26 27

Soames (2005, 307–10). Soames (2005, 310–13).

Ambitious Two-Dimensionalism • 271

weak two-dimensionalist analysis given by WT4b, in favor of something along the lines of H. H. Both the primary and secondary propositions (intensions) associated with S are responsible for necessary conditions on the truth of x (knows) believes that S, without providing sufficient conditions. Such an ascription, as used in C, is true of an agent A in a circumstance of evaluation w iff in w, A ( justifiably) accepts some (true) sentence or mental representation M which is such that (i) the secondary intension of M in A’s context in w is identical with the secondary intension of S in C, and (ii) the primary intension of M is “appropriately related” to the primary intension of S in C. There are various reasons to doubt the effectiveness of this proposal in solving Frege’s puzzle. For one thing, it presupposes that all names and natural kind terms have their referents semantically fixed by description, and so is vulnerable to the first objection brought against weak twodimensionalism. For another thing, the most plausible candidates for reference-fixing descriptions that might withstand Kripke’s semantic arguments involving error, misdescription, and failures of uniqueness are metalinguistic descriptions involving the historical chain of reference transmission in which the user of the term stands. However, these descriptions are very poor candidates for specifying the contents of the beliefs and assertions one ascribes to others. When I say that the ancient Babylonians believed that Hesperus was a star, I am surely not saying that their mental contents included anything having to do with the chain of reference transmission the brought the name Hesperus to me. For these, and other, reasons Chalmers’s proposal for solving Frege’s problem doesn’t seem very promising.28 But even if we put Frege’s puzzle aside, adopting H is a more effective strategy for challenging ambitious two-dimensionalism than for saving it. Such a strategy threatens the conjunction of two principles, A priori 1 and A priori 2, which have been central to the approach. A priori 1 A sentence S is an instance of the a priori iff the character of S assigns every context a proposition that is true in that context. A priori 2  It is knowable a priori that S is true iff S is an instance of the a priori. 28

Soames (2005, 322–24).

272 • Essay Nine

Since the objection against taking names to be descriptions rigidified using the actuality operator within the framework of weak two-dimensionalism remains in force against the hybrid version of the view, the hybridist must analyze names and natural kind terms as dthat-rigidified descriptions. But then, by A priori 1, If there is a unique F, then n is F is analyzed as If there is a unique F, the dthat [the F] is F, and so is characterized as a priori—when the referent of n is fixed by the F. The right-to-left direction of A priori 2 will then characterize the ascriptions that result from prefixing it is knowable a priori that to them as true. However, this can be correct only if It is knowable a priori that if there is a unique F, then x is F is true relative to an assignment of the referent of n to ‘x’.29 Since such de re knowledge of individuals can’t, in general, be a priori, the right-to-left direction of A priori 2 is in jeopardy. Nor would denying this be of much help, since if such knowledge of individuals can be a priori, then If there is a unique F, then x is F (taken relative to an assignment of the referent of n to ‘x’) will, according to A priori 2, count as an instance of the a priori, even though, typically, its character (relative to such an assignment) will not assign a truth to every context—thereby falsifying the conjunction of the left-to-right directions of A priori 1 and A priori 2. An independent argument against this conjunction can be constructed as follows. Step 1. If H is correct, then there will be cases in which an utterance, by a1 in context C1, of A2 knows that n is F is true, where (i) the referent a2 of A2 justifiably accepts a true sentence m is F the secondary intension of which in a2’s context C2 is the same as the secondary intension of n is F in C1, and (ii) the primary intensions of these two sentences are at least somewhat different because the primary intension of m (as used in C2) differs slightly from that of n (as used in C1). Here, n is a name, and m is either a name or an indexical—m may even be the same name as n, provided that one recognizes (as Chalmers knows he must) that different speakers may use it with somewhat different primary intensions in their respective contexts. 29 It might seem that the defender of H could deny the move from the truth of It is knowable a priori that if there is a unique F, then n (i.e., dthat [the F]) is F to the truth of  It is knowable a priori that if there is a unique F, then x is F relative to an assignment of the referent of n to ‘x’, on the grounds that the primary intension of the complement clause of the latter is not “appropriately related” to that of the complement clause of the former. However, this would fly in the face of the fact that if n is a name that refers to o, the truth of α knows/believes (a priori) that . . . n . . .  guarantees the truth of α knows/believes (a priori) that . . . x . . .  , relative to an assignment of o to ‘x’. See Soames (2005, 261–62, 316–18).

Ambitious Two-Dimensionalism • 273

Step 2. If there are cases of the sort indicated in step 1, then we may suppose that some will be symmetrical in that what a1 uses A2 knows that n is F to truly report about a2, the latter can use  A1 knows that m is F to truly report about a1. These reports can be jointly true because the secondary intensions of the two complement clauses are the same in their respective contexts, while the primary intensions of these clauses, though different, are close enough to satisfy constraints imposed by H. Step 3. We may assume, with the two-dimensionalist, that the primary intensions of n and m are given by a pair of descriptions Dn and Dm which pick out the same objects in their respective contexts, but which denote different objects with respect to some pairs of contexts. One way for this to happen is for Dn to be like Dm except for containing a conjunct that predicates G of its denotation whereas the corresponding conjunct in Dm predicates an unrelated predicate F. Step 4. Now consider a1’s utterance in C1 of the ascription A2 knows that if n exists, then n is F. For the reasons given in step 2, this ascription should be true, since a2 justifiably accepts the true sentence If m exists, then m is F, the secondary intension of which in C2 is the same as the secondary intension of  If n exists, then n is F in C1. As before, although the primary intensions of the two sentences (as used in their respective contexts) differ slightly, they should be close enough to satisfy the constraints imposed by H. Step 5. Since F is included in the reference-fixing conditions for m, the primary intension of If m exists, then m is F (as used in C2) should be necessary, and so the knowledge correctly attributed to a2 by a1’s utterance of A2 knows that if n exists, then n is F should (according to ambitious twodimensionalism) be counted as a priori. Step 6. So A2 knows a priori that if n exists, then n is F should be true, as used in context C1, as should It is knowable a priori that if n exists, then n is F—even though the primary intension of If n exists, then n is F (as used in C1) is not necessary (thereby falsifying the conjunction of the left-to-right directions of A priori 1 and A priori 2). What this shows is that Chalmers’s explicit proposal H for the attitudes threatens what previously had been a central and defining tenet of ambitious two-dimensionalism—the conjunction of A priori 1 and A priori 2. Thus, his proposal may be more of a threat to that view than a way of implementing it.

274 • Essay Nine

Be that as it may, there is a final difficulty that the hybrid view shares with other versions of ambitious two-dimensionalism that analyze proper names (and natural kind terms) as dthat-rigidified descriptions. The difficulty involves the following acquaintance constraints on names and their potential analyses (as well as similar constraints on natural kind terms). The Acquaintance Constraint on Names If there is an object o, such that n designates o, and a speaker s understands n, then x knows that ‘n’ designates n (if ‘n’ designates anything at all) is true of s, as is There is an object o such that x knows that ‘n’ designates o (if ‘n’ designates anything at all). The Acquaintance Constraint on Descriptive Analyses of Names If dthat [the D] is the analysis of n, as used by s in C, and dthat [the D] designates o (relative to C), and s understands the D, then  There is an object o such that x knows that ‘the D’ designates o (if ‘the D’ designates anything at all) is true of s (relative to C). The problem is that descriptions satisfying the second constraint are not generally available. What is needed are descriptions such that understanding them and correctly believing them to denote puts one in a position to have de re beliefs about objects they denote. There are such descriptions. For example, it is arguable that (i) understanding the descriptions the ancient philosopher who is the referent of the name ‘Plato’ that I have encountered in the language of my community and the ancient philosopher reference to whom is inherited by my use of ‘Plato’ from standard uses of ‘Plato’ by speakers of the language of my community and (ii) correctly believing them to denote, does put me in a position to have de re beliefs about Plato expressible using the name. However, this is only because the referent of the name, as used in my linguistic community, has already been established independently of the satisfaction of these descriptions. Since advocates of descriptivist theories such as ambitious two-dimensionalism require reference to be semantically fixed by virtue of satisfying the descriptive content associated with use of a name, descriptions that presuppose that reference has already been established independently can’t play that role. It is doubtful that any other descriptions available to ordinary speakers can do the job.

Conclusion Having reached this point, we may summarize our results as follows: The versions of ambitious two-dimensionalism we have been able to state

Ambitious Two-Dimensionalism • 275

precisely—pragmatic two-dimensionalism, plus strong and weak semantic two-dimensionalism—are clearly incorrect.30 The hybrid version of the view—which is problematic as a solution to Frege’s puzzle—both suffers from difficulties that it shares with other descriptivist analyses of names, and threatens principles central to ambitious two-dimensionalist accounts of the necessary a posteriori and the contingent a priori. At present, there is no further version of the view left to consider. If the twodimensionalist approach is to survive at all, the onus is on those who favor it to formulate a new version that avoids the objections raised here. For that, we will simply have to wait.

References Bealer, George. 2002. “Modal Epistemology and the Rational Renaissance.” In Conceivability and Possibility, ed. Tamar Gendler and John Hawthorne, 71–125. Oxford: Oxford University Press. Chalmers, David. 1996. The Conscious Mind. New York: Oxford University Press. ———. 2002. “Components of Content.” In The Philosophy of Mind: Classical and Contemporary Readings, ed. Chalmers, 608–33. New York: Oxford University Press. Jackson, Frank. 1998a. From Metaphysics to Ethics. Oxford: Clarendon Press. ———. 1998b. “Reference and Description Revisited.” In Language, Mind, and Ontology, ed. James E. Tomberlin, 201–18. Philosophical Perspectives 12. Oxford: Blackwell. Kaplan, David. 1989. “Demonstratives: An Essay on the Semantics, Logic, Metaphysics, and Epistemology of Demonstratives and Other Indexicals.” In Themes from Kaplan, ed. Joseph Almog, John Perry, and Howard Wettstein with the assistance of Ingrid Deiwiks and Edward N. Zalta, 481–563. New York: Oxford University Press. Kripke, Saul A. 1980. Naming and Necessity. Cambridge: Harvard University Press. Originally published in Semantics of Natural Language, ed. Donald Davidson and Gilbert Harman, 253–355 (Boston: Reidel, 1972). Lewis, David. 1999a. “Elusive Knowledge.” In Papers in Metaphysics and Epistemology, 418–45. Cambridge: Cambridge University Press. Originally published in Australasian Journal of Philosophy 74 (1996): 549–67. ———. 1999b. “Naming the Colors.” In Papers in Metaphysics and Epistemology, 332–58. Cambridge: Cambridge University Press. Originally published in Australasian Journal of Philosophy 75 (1997): 325–42. Salmon, Nathan. 1986. Frege’s Puzzle. Cambridge: MIT Press. Soames, Scott. 2002. Beyond Rigidity: The Unfinished Semantic Agenda of “Naming and Necessity.” New York: Oxford University Press. 30 For a different, but in some respects related, argument for a similar conclusion about the prospects for two-dimensionalism, see George Bealer’s discussion (2002, sec. 2).

276 • Essay Nine ———. 2003. Philosophical Analysis in the Twentieth Century. 2 vols. Princeton: Princeton University Press. ———. 2005. Reference and Description: The Case against Two-Dimensionalism. Princeton: Princeton University Press. ———. 2006. “Understanding Assertion.” In Content and Modality: Themes from the Philosophy of Robert Stalnaker, ed. Judith Thomson and Alex Byrne, 222–50. Oxford: Oxford University Press. Stalnaker, Robert. 1984. Inquiry. Cambridge: MIT Press. ———. 1999. “Assertion.” In Context and Content: Essays on Intentionality in Speech and Thought, 78–95. New York: Oxford University Press. Originally published in Syntax and Semantics, vol. 9, Pragmatics, ed. Peter Cole, 315–22 (New York: Academic Press, 1978).

ESSAY TEN

Actually

My topic is the metaphysics and epistemology of actuality and possibility, plus the semantics and pragmatics of the language we use to talk about it. By ‘actuality’ I mean the actual world-state. By ‘possibility’ I mean all possible world-states, both the metaphysically and the epistemically possible. The actual world-state is the way the world is. Metaphysically possible states are ways the world could have been. Epistemically possible states are ways the world can coherently be conceived to be. In what follows I will sketch a conception of what these world-states are, and explore how we know about them. To that end I will examine two characterizations of epistemic possibility. EP1. A world-state w is epistemically possible iff w is a way the world can coherently be conceived to be, which it cannot be known a priori not to be. EP2. A world-state w is epistemically possible iff w is a way the world can coherently be conceived to be, and one cannot know a priori that w is not a way the world could be (or have been). Since knowing that w is a way that the world could not be (or have been) involves knowing that w is a way that the world is not, but not vice versa, EP1 is more restrictive than EP2. I will, therefore, take it to be the default definition.1 One of my tasks will be to determine whether or not it needs to be liberalized. In addition to limning the nature of world-states, and distinguishing different types of possibility, I will also explain the semantics and pragmatics of our talk about actuality and possibility. All of this, I will argue, leads to the resolution of certain puzzling problems about the necessary a posteriori and the contingent a priori. Two Uses of Actually I begin with the standard philosophical semantics of the actuality operator. This semantics presupposes that sentences are evaluated for truth or 1

This is the definition I adopted in Soames (2005b, 2006a; and n.d.).

278 • Essay Ten

falsity at pairs of world-states—one, the designated “actual state,” which provides the interpretation of ‘actually’, and the other, which may be any possible state, which is needed to evaluate sentences containing ‘necessarily’ and ‘possibly’. As David Kaplan taught us, the first of these states may be thought of—along with a designated time, place, and agent—as a context of utterance, which, together with the meaning of the sentence uttered, determines the proposition the sentence expresses. The second world-state is a circumstance of evaluation at which that proposition is assessed for truth or falsity.2 Syntactically, the actuality operator combines with a formula to form a more complex formula. Semantically, it is an indexical, the content of which it varies from context to context. For example, the sentence ‘Actually Kaplan wrote “Demonstratives,”’ used by anyone at the actual worldstate, @, expresses the proposition that Kaplan wrote “Demonstratives” at @, while the same sentence used by a speaker at a world-state w expresses the proposition that Kaplan wrote “Demonstratives” at w. In this way, ‘actually’ stands for the world-state Cw of the context in a manner analogous to the way in which ‘now’ stands for the time, and ‘I’ stands for the agent, of the context. When p is the proposition expressed by S in C, Actually S expresses the proposition that p is true at CW (which predicates, of Cw, the property of being a world-state at which p is true). Since this proposition doesn’t change truth-value from state to state, Actually S is true at an arbitrary pair C,w iff S is true at C, Cw.3 Thus, when S is true at Cw, Actually S is a necessary truth, and when the x: Fx denotes a unique individual o at CW, the x: actually Fx denotes o at C,w, for all worldstates w at which o exists, and never denotes anything else at C,w*, for any w*. Hence, ‘actually’ is a rigidifier. However, the x: actually Fx is not directly referential, since its content is not an object but a descriptive condition, (expressed by) the unique object which is F at Cw.4 So understood, the actuality operator is a useful logical tool. However, does it capture the ordinary meaning of the English word ‘actually’? Initially, there appears to be evidence on both sides. On the positive side, 2 Kaplan (1989). Whereas Kaplan takes circumstances of evaluation to time/world-state pairs (because he takes the truth-values of propositions to be temporally changeable), I take circumstances to be world-states (because I agree with Nathan Salmon (1989b) that propositions have their truth-values eternally). 3 To say that S is true at C,w is to say that the proposition expressed by S at C is true when evaluated at w. 4 If, when evaluating the x: actually Fx at w, ‘x’ ranges over all possible individuals, then its denotation at C,w need not exist at w. If the range of ‘x’ at w is restricted to things existing at w, this is not so, leading to complications noted in Soames (2005b, 29–30). For simplicity, I let ‘x’ be unrestricted, unless otherwise indicated.

Actually • 279

(1) provides evidence that both the adverb ‘actually’, and its adjectival cousin ‘actual’, are rigidifiers. (1) It could have been the case, had just a few things gone differently, that the actual winning general (the general who actually won) at Chancellorsville lost that battle. What I say in assertively uttering (1) is true iff the general who won the battle at the actual world-state @—Robert E. Lee—lost the battle at a world-state w differing in only a few respects from @. This is just what we should expect, if ‘actual’ and ‘actually’ are rigidifiers. On the negative side, (2a) and (2b) seem to provide evidence that S and Actually S differ only rhetorically. (2) a. Actually, I live in California. b. I live in California In uttering (2a), I assert the information carried by (2b), while signaling that it may be unexpected. In general, Actually S is used rhetorically to indicate that the information expressed by S, which one is asserting, may contrast with possibilities that one’s hearers find salient. Thus, we are faced with a dilemma. The all-too-ubiquitous rhetorical use of actually seems to suggest that, in ordinary language, it is a purely rhetorical device that is logically and semantically inert, while its apparently rigidifying use points in the opposite direction. In what follows, I will show that this dilemma is merely apparent: the rhetorical use is fully explainable on the hypothesis that the ordinary word ‘actually’ is simply the actuality operator of philosophical semantics. However, there is more to be done before we reach that result.

Actual and Actually First a word about the relationship between the adverb ‘actually’—which can be prefixed to a sentence, Actually S—the adjective ‘actual’—which modifies a noun, actual N—and the predicate, ‘is actual’. To the extent that these forms are interdefinable, their grammatical differences are philosophically unimportant. Taking the indexical semantics of ‘actually’ as basic, one naturally understands the actual N as equivalent to the x: actually x is N, thereby explaining the apparent rigidity of the former. However, the predicate ‘is actual’, used by philosophers to express the property of being a world-state that obtains (or is instantiated), is another matter. Although it applies only to the way the world is, it could have applied to any way the world could have been, thereby making (3a, b) true.

280 • Essay Ten

(3) a. Every metaphysically possible world-state is one that could have been actual. b. No world-state can be known a priori to be actual. Though equivalent to (4a), this predicate is not equivalent to (4b), which—when used at @—expresses the property (4c), which is equivalent to the property being @. (4) a. is a world-state that obtains (is instantiated) b. is a world-state that actually obtains (is instantiated) c. being a world-state that obtains (is instantiated) at @—i.e., being a world-state that would obtain (be instantiated), if @ obtained (were instantiated) Since this property would make (3a, b) trivially false, ‘is actual’—as used by philosophers—is not definable in terms of the actuality operator. Given this use of ‘is actual’, one can, of course, define a corresponding use of the actual N in which it is synonymous with the thing that is N at whatever world-state is actual. However, on this use, ‘actual’ is not a rigidifier, since the actual N is trivially equivalent to the N. Since this interpretation doesn’t explain the apparently rigidifying effect of adding ‘actual’ or ‘actually’ to descriptions in English, I will assume that these descriptions contain the indexically defined operator. I will further assume that this is the operator in Actually S, since, as I will show, the rhetorical effects of assertively uttering that sentence pose no problem for this hypothesis. I use the phrase ‘the actual world-state’ to name the way, @, that the world is. A different, but referentially equivalent, understanding assimilates it to ‘the world-state w such that actually w is instantiated’— which, when used at @, has the same content as ‘the world-state w such that if @ were instantiated, then w would be instantiated’. On this understanding, it is knowable a priori that @ is the actual world-state—even though it is not knowable a priori that @ obtains (or is instantiated).5

Actuality, Necessity, and Apriority With these linguistic matters in place, I turn to the metaphysics and epistemology of actuality and possibility. Since adding the actuality operator to a contingent truth produces a necessary truth, and since it is widely assumed that adding it to a truth that is knowable only a posteriori preserves aposteriority, the actuality operator is often seen as a rich source of the 5 If ‘the actual world-state’ were understood as ‘the world-state that is actual (i.e., obtains)’, then it would be nonrigid, referring, at each w, to w itself. On that construal, one can’t know a priori that @ is the actual world-state.

Actually • 281

necessary a posteriori. A related point is made about the contingent a priori. When S is contingent, the proposition expressed by S iff actually S is also contingent, even though it is knowable a priori. Although these results appear obvious, it is wise to withhold judgment on them until we have a clearer picture of what the actual world-state really is. Since the proposition expressed by Actually S is a singular proposition that attributes the property of being a world-state at which p is true to @, knowing it, and thereby satisfying x knows that actually S, involves having de re knowledge of @. Surely, the nature of @ is relevant to whether we have such knowledge, and, if so, how we come by it. Thus, in order to assess the role of the actuality operator in generating instances of the necessary a posteriori and the contingent a priori, we need to clarify what world-states are.

The Nature of World-States My account is based on three leading ideas. From Robert Stalnaker, I take the idea that world-states are not Lewisian alternate concrete worlds (universes), spatially and temporally disconnected from ours. Rather, they are properties specifying ways the world could be, or be coherently conceived to be.6 From Saul Kripke, I take the idea that world-states may be specified in terms of the objects and properties we find around us, and so need not be given purely qualitatively. As Kripke puts it: Don’t ask: how can I identify this table in another possible world, except by its properties? I have the table in my hands, I can point to it, and when I ask whether it might have been in another room, I am talking, by definition, about it. I don’t have to identify it after seeing it through a telescope.7 From Nathan Salmon, I take the idea that the space of world-states includes not only the actual and genuinely possible, but also some that are metaphysically impossible.8 The actual world-state is the maximal, world-constituting property that the world really instantiates. Metaphysically possible world-states are maximal, world-constituting properties that could have been instantiated. Epistemically possible world-states are maximal, world-constituting properties that we can coherently conceive to be instantiated, and (assuming E1) that we cannot know a priori not to be instantiated. 6 Stalnaker (1976). Whereas Stalnaker identifies ways the world could be with ways the world can be conceived to be, I distinguish the two. 7 Kripke (1980, 52–53). 8 Salmon (1989a).

282 • Essay Ten

For insight into these properties, I turn to Rudolf Carnap’s classic notion of a state description, used in giving the semantics of an elementary first-order language L.9 Details aside, a Carnapian state-description is a complete, consistent set of atomic sentences of L, or their negations (resulting in a complete assignment of truth-values to atomic sentences). Truth-values of complex sentences relative to a state description are determined using familiar recursive clauses for quantifiers, truth functions, and modal operators. In updating this picture, I replace Carnap’s atomic sentences with structured, Russellian propositions expressed by atomic formulas, relative to assignments of objects to variables. Complete, consistent sets of such propositions, and their negations, are used to define world-states, at which complex sentences, and the propositions they express, are evaluated. Let D be the domain of objects talked about, and B the set of properties expressed by simple predicates of L, including an existence predicate. A world-description SW is a set each member of which is either an atomic proposition, consisting of an n-place property from B plus an n-tuple of objects from D, or the negation of such. SW is complete iff for every atomic proposition, either it or its negation is a member of SW. It is consistent iff its members cannot be known a priori not to be jointly true. The world-state w corresponding to SW is the property of making the propositions in SW true. To conceive of w as instantiated is to conceive of every member of SW being true, while taking the objects in the universe to include only those the existence of which is required by SW. The propositions in SW are, of course, not the only ones true at w. Others include those expressed by nonatomic, nonmodal sentences the truth of which is determined from Sw by recursive clauses for quantifiers and truthfunctional operators. All states in the structure are epistemically possible. The one that is instantiated is actual. The ones that could have been instantiated are metaphysically possible. The rest are metaphysically impossible. Possibly S is true at w iff S is true at some world-state that is metaphysically possible from w—similarly for Necessarily S. World-states that are metaphysically possible from one state may differ from those that are possible from another. For example, suppose that P1 and P2 are mutually exclusive, essential properties of anything that has them (so that having one precludes having the other). Suppose further that one can determine whether an object has these properties only by empirical investigation. It follows that if it is true at w1 that o has P1, and true at w2 that o has P2, then the world-states metaphysically possible from the two states are different. 9

Carnap (1956).

Actually • 283

Since each world-state is epistemically possible, it can coherently be conceived to be instantiated. For each such state w1, there is a set of states w2 that would be metaphysically possible, if w1 were instantiated. These are properties the universe could have had, if it had had w1. For each such w2 there is a set of states w3 that would be metaphysically possible, if w2 were instantiated. These are properties that it could have been the case that the universe could have had.10 This process is repeatable. The truth-values, at w, of the propositions expressed by sentences containing modal operators are determined not by w itself, but by its position in the overall space of world-states. Finally, we introduce the actuality operator, allowing each metaphysically possible world-state to be the designated state of a possible context.11 Taking @ to be designated in our present context, and S to express p, our use of Actually S expresses the proposition that p is true at @. The truthvalue of this proposition at an arbitrary state w is determined, not by the content of the world-describing set SW, nor by the world-states metaphysically possible from w, but simply by the truth-value of p at @. Actually S is true at any world-state iff the proposition expressed by S is true at @. To sum up, a world-state w is a property that gives a complete story of what the universe would be like if w were instantiated. Since it is no part of that story to specify what the universe would be like if other worldstates were instantiated, the propositions in terms of which w is defined don’t contain explicit information about other world-states. This is compatible with the fact that, for any world-state w* and proposition p, we can always evaluate the truth-value of the proposition that p is true at w* at any world-state whatsoever. We need only remember that a proposition can be true at a world-state without being one of the propositions that define it. There are, of course, limitations to this framework. Like Carnap, I have introduced a space of states to evaluate sentences of a simple firstorder, modal language L (and the propositions they express). As a result, some features of this space are tied to features of L. The properties in terms of which world-states are defined are those expressed by simple predicates of L, and the objects mentioned in the definition are those that L is used to talk about. If richer languages had been chosen, the worldstates would have been richer, and the rules for determining truth at a state would have been more complex. This raises two questions. 10 This is not a stutter. As Nathan Salmon (1989a) argues (for ship of Theseus type examples) it should not be assumed that if w is (metaphysically) possibly possible, then w is (metaphysically) possible. 11 There are some niceties excluding certain world-states from playing this role. However, these won’t matter to us.

284 • Essay Ten

Q1. Would the choice of a more complex language invalidate essential features of this framework? Q2. Should the relativization of the space of world-states to particular languages, and contextually varying inquiries involving them, be discarded in favor of an absolute space of worldstates, equally relevant for all inquiries in all languages? Stalnaker addresses Q2. One might ask, are there such things as possibilities, or possible worlds, in this sense [maximal properties that the world might have had]? I doubt that it is plausible to believe that there is, independent of context, a well-defined domain of absolutely maximally specific possible states of the world, but I do not think the proposed conception . . . requires such a domain. The alternative possibilities . . . must be exclusive alternatives made in the context at hand. But one can make sense of this requirement even if there is no ultimate set of possibilities relative to which any possible distinctions might be made. One might think of possible worlds as something like the elements of a partition of a space, rather than as the points of the space. The space might be partitioned differently in different contexts, and there might be no maximally fine partition.12 There is, I think, something right about this. World-states are properties attributed to the universe. When attributed, they are taken to capture everything relevant to the inquiry at hand. However, it is reasonable to suppose that for every inquiry that might be undertaken, there is another requiring a finer level of detail and specificity, for which a more fine-grained, and fully-articulated set of world-states would be needed. If so, then there may be no absolute sense in which a worldstate is a maximally informative story about the universe that answers every conceivable question, and evaluates every conceivable proposition. Rather, world-states are properties treated as maximal for particular purposes. The stories they tell are maximally informative in the sense of answering every question relevant to a given inquiry. This doesn’t mean that world-states are made up to suit our interests. The properties are there independently. It is the use to which we put them that is relative to us. Thus, it is no defect in the framework I have sketched that it doesn’t provide an absolute conception of maximality for world-states. It would be a defect, however, if the framework couldn’t be liberalized to overcome limitations of the simple Carnapian conception of world-states, and generalized to accommodate languages richer than L. Later, I will 12

Stalnaker (1999, 136).

Actually • 285

discuss ways of doing so. However, we already have enough to resolve some puzzles, and record some results.

The Necessary A Posteriori Revisited Kripkean instances of the necessary a posteriori are propositions that predicate essential properties or relations of objects that can be known to possess them only empirically.13 The function of empirical evidence needed for knowledge of these propositions is to rule out epistemically possible, but metaphysically impossible, world-states at which they are false. The same cannot be said for what are widely taken to be examples of the necessary a posteriori involving the actuality operator. Whenever S expresses a contingent truth p, Actually S expresses the necessary truth that p is true at @. However, since Actually S is trivially inferable from S, and since the proposition it expresses often doesn’t seem to be knowable in any other way, it has seemed to be knowable only a posteriori, whenever p is. This is problematic. If p is true at @, then the proposition that p is true at @ is true, not just at every metaphysically possible world-state, but at every epistemically possible state as well. What, then, is the role of empirical evidence needed for a posteriori knowledge? If there are no possible world-states at which this proposition is false, why is empirical evidence required to know it? Strictly speaking, it isn’t. World-states are properties we can conceive the universe as having—properties of making certain world-describing sets of propositions true. Imagine, then, a tiny universe consisting of two blocks side by side, with a third on top. This world-state, Tiny, is (in effect) the property of containing blocks 1 and 2 side by side, with block 3 on top. Since we have no trouble comprehending the content of this property, we can know, just by thinking about it, that if it were instantiated, then block 3 would be sitting on blocks 1 and 2. So, when p is the proposition that block 3 is sitting on those blocks, it is knowable a priori that p is true at Tiny. The point generalizes. If, as often seems plausible, the world-states relevant to an inquiry are finitely specifiable, then, for every such state w, and every proposition p the truth of which is calculable from those defining w, the proposition that p is true at w is knowable a priori. Since this result applies to the actual world-state (relative to an inquiry), as much as to any other, the proposition expressed by uses of Actually S is often knowable a priori. Knowable, though not, necessarily, known a priori. Since the actual world-state, relative to many inquiries, will be much more complex than Tiny, we may not be able grasp it in the nondemonstrative 13

See Soames (2006b; essay 6 in this volume) and Soames (n.d.).

286 • Essay Ten

way we grasp Tiny—in which case we won’t have any way of calculating the truth-values of propositions from a complete specification of @. In such cases, our only practical way of coming to know that p is true at @ is by inferring it from p. Thus, when p is a posteriori, our knowledge of the proposition expressed by Actually S may be a posteriori, even though what we know can, abstracting away from our cognitive limitations, also be known in another way.14 How, for example, do I come to know the necessary truth that actually over 600,000 soldiers died in the American Civil War? I derive it indexically. To say that actually over 600,000 soldiers died is just to say that at this very world-state—the one that is instantiated—over 600,000 died. Hence, if I know from historical research how many soldiers died in the Civil War, I know how many died at this very world-state, and I can express this knowledge by saying: Actually, over 600,000 soldiers died in the Civil War. In so doing, I demonstrate @, and say of it that a certain proposition is true at it, in something like the way in which, in another context, I demonstrate Southern California, and say of it that it is warm there, by uttering: ‘It is warm here’. Just as in the latter case, I demonstrate a certain large territory, and say something about it, on the basis of my limited acquaintance with it—even though my ignorance of the territory greatly exceeds my knowledge of it—so, in the former case, my limited familiarity with the way the world is allows me to refer to it indexically, and say something about it, despite being ignorant of much of its content. In this way, I come to know, a posteriori, the necessary truth that at @, over 600,000 soldiers died in the Civil War, by deriving it from a contingent truth that I know a posteriori. The inference also runs in the other direction. To know that actually over 600,000 soldiers died in the Civil War is to know that over 600,000 soldiers died in the Civil War, at this very world-state—from which one may trivially conclude that over 600,000 soldiers died then. Why, then, isn’t it knowable a priori that over 600,000 soldiers died in the Civil War? After all, I have argued, (i) that the necessary truth that at @, over 600,000 soldiers died in the Civil War is, in principle, knowable a priori, and (ii) that there is a certain way of knowing this truth, when @ is presented indexically, such that when one knows it this way, one can derive the contingent truth that over 600,000 soldiers died in the Civil War from it. Thus, the contingent truth is an a priori consequence of an a priori truth. How, then, can it fail to be knowable a priori? It fails to be knowable a priori because the route to it from the necessary proposition that at @, over 600,000 soldiers died in the Civil 14 This position is suggested in Soames (2005b, 304–5 n. 16)and Soames (2006c, 231–32; essay 8 in this volume).

Actually • 287

War is different from, and at odds with, the route to the apriority of the latter. In order to derive p from the proposition that p is true at @, @ must be given as this very word-state (the one that is instantiated). However, in order for one to know a priori that p is true at @, @ can’t be given in this way, but must be known by grasping the propositions that define it. The proposition that p is true at @ is entertainable in two radically different ways. One way—which, as a practical matter, may exceed our cognitive abilities—involves grasping the propositional content of @. One who entertains the proposition in this way can know it a priori—by deriving p from the propositions that define @. But when @ is presented in this way, there is no way of knowing that it is instantiated. Hence, when one entertains the proposition that p is true at @ in a way that allows one to know it a priori, there is nothing in one’s knowledge that allows one to infer p from it. The second, indexical, way of entertaining the proposition that p is true at @—which is how it is presented using the actuality operator—does not involve grasping the full propositional content of @. When presented with the proposition in this way, we cannot determine it to be true a priori, though we can move a priori from it to p, and vice versa. Since on neither way of knowing that p is true at @ is there a way of establishing p a priori, p is knowable only a posteriori. Is it strange that the proposition that p is true at @ should, in principle, be knowable a priori, even though, in practice, our knowledge of it is often a posteriori? Not when one realizes the kind of proposition it is— namely, one that relates one propositional content (a particular proposition) to another propositional content (a particular world-state that is itself propositionally defined). As the example about the miniature worldstate Tiny shows, propositions of this kind are, in general, knowable a priori. The fact that the complexity of world-states often exceeds our psychological limitations is balanced by the fact that many other a priori truths do, too. Some arithmetical truths are too complex for us to effectively evaluate, even though they are expressed by theorems of correct arithmetical theories. We shouldn’t deny that these propositions are knowable a priori—even if our psychological limitations afford us no way of knowing them, short of using a posteriori methods, like running a reliable computer program for a long time. The point about propositions expressed by uses of Actually S is similar.

A Puzzle about the Contingent A Priori In explaining Kripkean instances of the necessary a posteriori, I argued that the function of empirical evidence required to know them is to rule

288 • Essay Ten

out metaphysically impossible, but epistemically possible, world-states in which they are false. This may seem to suggest (5). (5) If p is false at some epistemically possible world-state, then p isn’t a priori. So, if p is a priori, then p isn’t false at any epistemically possible world-state, and so (we may assume) p is true at every such state. But then, if p is contingent a priori, it will follow that p is true at all epistemically possible world-states, while being false at some metaphysically possible state. Thus, if (5) is correct, some metaphysically possible worldstates are epistemically impossible. How can that be? A metaphysically possible world-state is a property the universe could have had. But surely, one is inclined to think, if the universe could have had w, then there can’t be anything incoherent, or a priori inconsistent, in supposing that it does have w. And, if there is no such inconsistency, then w won’t be knowable a priori not to be instantiated. In short, any metaphysically possible world-state should be epistemically possible. Since this contradicts our earlier result, one of the suppositions leading to the contradiction must be abandoned—either (i) that there are instances of the contingent a priori, (ii) that (5) is correct, or (iii) that it is never a priori inconsistent to suppose, of any metaphysically possible world-state, that it is instantiated. That there are instances of the contingent a priori is shown by the fact that, when S is contingent, anyone who, at @, knows the a priori (6a) is in position to derive the contingent (6e) by steps that can be known a priori to be truth-preserving.15 (6) a. S iff S b. So, it is true at this very world-state (said demonstrating @) that S iff S c. So, it is true at this very world-state (said demonstrating @) that S iff it is true at this very world-state (said demonstrating @) that S d. So, S iff it is true at this very world-state (said demonstrating @) that S e. So, S iff actually S The step from (a) to (b) is based on the principle that for any proposition p, if at world-state w, an agent A knows p, then A needs no further justifying evidence to come to know, of w, that it is a world-state at which p is true. Thus, our a priori knowledge of proposition (a) is sufficient to allow us to come to know (b)—i.e., to know, a priori of @, that it is a 15

A version of this argument is given in Soames (2005b, 120–22).

Actually • 289

world-state at which proposition (a) is true. But if we know, of this very state @, that it makes the proposition expressed by S iff S true, then we need no further information to come to know the same thing about the proposition expressed by S iff it is true at this very state that S. Thus, proposition (c) is knowable a priori. The next step, to (d), is based on the principle that if a use, at w, of It is true at this very world-state that R (said demonstrating w) expresses knowledge based on evidence E (where E may be null), then a corresponding use of R does too. Given the apriority of (c), we conclude that the proposition expressed by (d) and (e) is also knowable a priori—despite being contingent. Hence, there is an instance of the contingent a priori for each contingent truth. Niceties aside, we may, therefore, take it that for every metaphysically possible world-state w ≠ @, there is a proposition (expressed in @ by (6e)) which is false at w, despite being knowable a priori. So, if (5) is true, all metaphysically possible world-states other than @ are epistemically impossible. However, (5) isn’t true. The temptation to think otherwise is linked to the temptation to accept (7)—which (in the presence of EP1) is interderivable with (5). (7) If p is true at w, and it can be known a priori that p is false, then w can be known a priori not to be instantiated (in which case w is epistemically impossible). The initial plausibility of (7), as well as its ultimate falsity, is illustrated by (8). (8) a. Saul philosophizes iff actually (i.e., it is true at @ that) Saul philosophizes b. ~ Saul philosophizes & actually (i.e., it is true at @ that) Saul philosophizes Since it is contingently true that Saul philosophizes, there is a metaphysically possible world-state w at which (the proposition expressed at @ by) (8a) is false, and (8b) is true—despite the fact that (8b) is knowable a priori to be false. Thus, the antecedent of (7) is true for . If (7) were true, it would follow that w was knowable a priori not to be instantiated. But (7) isn’t true. World-states, it will be remembered, are properties defined by sets of basic, world-describing propositions—where the propositions true at w exceed, not only those that define w, but also those the truth of which is calculable from the ones that do. Crucially, the truth-values, at w, of propositions that ascribe truth-values to propositions at other worldstates are determined, not by the complete story about the universe told by w, but by the space of world-states of which w is a part. The following table applies this idea to example (8).

290 • Essay Ten

A Simplified Space of World-States P1 = the proposition that Saul Kripke philosophizes P2 = the proposition that Scott Soames philosophizes @ P1 P2

w ~P1 P2

w* P1 ~P2

w# ~P1 ~P2

Here we pretend that complete world-stories can be told in terms of P1 and P2. Of course, in any realistic example, the number of basic propositions would be greater, and the world-state-defining sets larger, and more numerous. However, this doesn’t affect our result, as long we retain the assumption that world-states are not defined in terms of the truth-values of propositions at other world-states. Given this, we can reduce the question ‘Can the world-state w in which (8b) is true be known a priori not to be instantiated?’ to the question of ‘Can it be known a priori that the propositions defining w, ~P1, and P2, aren’t jointly true?’ Since this can’t be known a priori, it can’t be known a priori that w isn’t instantiated. Thus, the consequent of (7), (7) itself, and (5) are all false—as is (7AP). (7AP) If it can be known a priori both (i) that p is false, and (ii) that p is true at w (i.e., that if w were instantiated, then p would be true), then it can be known a priori that (iii) w is not instantiated (in which case w is epistemically impossible in the sense of EP1). When p is (8b), and w is a world-state in which p is true, we have already shown that (i) is knowable a priori, and that the consequent of (7AP) is false. Thus, (7AP) is false iff (ii)—which amounts to (9)—is knowable a priori. (9)

If it were true that (~P1 & P2), then it would be true that (~P1 & it is true at @ that P1).

This is knowable a priori iff it is knowable a priori that it is true at @ that P1. But, as I argued using the world-state Tiny, when the truth of q is calculable from the propositions defining w, it is always knowable a priori that q is true at w. Since P1 is a defining proposition for @, (9) is knowable a priori, and (7AP) is false. The falsity of (7AP) is another example of an earlier result: a priori consequences of propositions that are knowable a priori are sometimes themselves not knowable a priori. In this case, one proposition—(i) of (7AP)—can be known a priori only when @ is presented indexically—as in ‘It is false that ( ~ Saul philosophizes & actually Saul philosophizes)’— while another proposition—(ii) of (7AP)—can be known a priori only when it is known a priori that Saul philosophizes at @—which requires

Actually • 291

@ to be presented nonindexically, in terms of its propositional content. Since there is no way of merging the a priori routes to (i) and (ii) into a single route to (iii), an agent can’t derive (iii) from a priori knowledge of (i) and (ii). To recap, propositions like those expressed by S iff actually S are genuine examples of the contingent a priori. Initial appearances to the contrary, the metaphysically possible world-states at which they are false are also epistemically possible in the sense of EP1. Why, then, doesn’t knowledge of these propositions require empirical evidence to rule out the epistemically possible world-states in which they are false? The answer is illustrated by (6), which shows how certain contingent propositions expressed using the actuality operator (or an analogous indexical) can be derived, by uniformly a priori steps, from corresponding truths the necessity and apriority of which are uncontentious. Because of this, these contingent propositions can be known a priori at the world-state, @, they are about. What rules out world-states at which they are false is not empirical evidence, but the transparent, indexical reference to the very worldstate at which the knower evaluates them. This contrasts with standard instances of the Kripkean necessary a posteriori, which don’t involve reference to the actual world-state.

A Unified Treatment of the Two Uses of Actually I now return to the two uses of ‘actually’ with which I began. The first, rhetorical, use is one in which an utterance of Actually S signals that the information expressed by S, which one is asserting, may be unexpected, or may contrast with possibilities one’s hearers find salient. In such cases, assertive utterances of S and Actually S say the same things, while differing in what they implicate or convey. The second, rigidifying, use is one in which the addition of ‘actually’ affects the (modal) truth-conditions of what is asserted. Despite their differences, these uses can be seen as two sides of the same coin, sharing a single indexical semantics, in which ‘actually’ directly refers to the world-state of the context. This reference is responsible for its rigidifying affect on definite descriptions, and for the fact that Actually S is necessary when S is contingent. It follows from the nature of world-states that the former is a priori when the latter expresses an a posteriori truth, calculable from the propositions defining the referent of ‘actually’. As we have seen, the apriority of the proposition expressed by Actually S is consistent with the fact that, in practice, our knowledge of it is often a posteriori. The indexicality of ‘actually’ allows speakers routinely to pass back and forth between S and  Actually S, even though the propositions they semantically express differ

292 • Essay Ten

dramatically. Because of this effortless inferential interchange, an assertive utterance of either sentence standardly results in the assertion of the propositions semantically expressed by both. This assertive equivalence is responsible for the rhetorical effect of uttering Actually S. Since adding ‘actually’ doesn’t change what is asserted, one who adds it is presumed to have some nonassertive reason for referring to @, and explicitly saying that the proposition p (expressed by S) is true in it, rather than simply asserting p by using S on its own. Standardly the reason is to contrast @ with other states that one’s hearers find salient, or expect to be instantiated. For example, when I said ‘Actually I live in California’, I did so in recognition that some people may have thought that I still lived in Princeton. By calling attention to the actual world-state, and explicitly saying that in it, I live in California, I implicitly contrasted it with possible states in which I live elsewhere. Thus, the indexical semantics that gives ‘actually’ its logical and philosophical punch also explains its rhetorical use in ordinary conversation.16 The same rhetorical effect could, of course, also be achieved by referring to @ nonindexically—as the world-state that is actual (in the philosopher’s sense of being instantiated). However, since the indexical ‘actually’ is needed independently, the rhetorical effect provides no reason to posit a second, nonrigidifying sense. If there is such a sense in ordinary language, it must be motivated in some other way.

Broader Issues Having illustrated a framework for thinking about actuality and possibility, I will briefly consider some challenges to it. The first concerns indexical reference to the actual world-state—which, I have argued, is the property of making a certain set of basic propositions true. Although the property that plays this role varies from inquiry to inquiry, in many cases the one that does will be very complex, often outstripping our cognitive capacities. In these cases, any knowledge that p is true at @ which we possess will be knowledge in which @ is presented to us by the actuality operator, or some indexical variant. It is, therefore, crucial that our acquaintance with, and ability to directly refer to, complex properties should enable us to know singular propositions about them in roughly the way in which our acquaintance with, and ability to directly refer to, 16 This argument parallels Grice’s (1989, 56–57) argument that the performative effects of uttering it is true that S, rather than S, noted by Strawson, are conversational implicatures arising from the semantics of ‘true’, rather than additions to that semantics.

Actually • 293

complex physical objects enables us to know singular propositions about them, despite being ignorant of many of their features. That it does is suggested by an example involving a box on my shelf. Three sides of it, which you can see, are square. Two others are square, and the back is a pyramid—though you can’t see them. We agree that a certain property, which we name ‘S’, is the shape of the box. By looking at the box, you know, of S, that its instantiation involves sides 1, 2, and 3 being square—knowledge you express by saying: ‘S, or this (threedimensional) shape, is one that involves sides 1, 2, and 3 being square’. In so doing, you succeed in directly referring to, and expressing your knowledge of, a complex property, despite the fact that you are only partially familiar with its content. Adding complexity—more sides, or shapes—doesn’t seem to change the situation. Knowledge expressed using the actuality operator is a more complex version of the same thing. It is, of course, true that direct reference plus propositional attitudes sometimes produces strange results. You may, after hearing me referred to by name, learn something that you express to me by saying ‘You are Scott Soames’. However, what you learned is (arguably) not the proposition semantically expressed by your sentence—which is also expressed by the uninformative ‘You are you’. Similar remarks apply to my box. When, after examining all sides, you say ‘This shape, S, is one that involves five square sides plus a pyramid’, you may express new (empirical) knowledge, despite the fact that the proposition semantically expressed by your sentence may (arguably) be one that you have known (a priori) all along. Similarly peculiar cases can be constructed with ‘actually’. Though potentially puzzling, these peculiarities arise for all directly referential expressions, and so do not count specially against the analysis given here. At most, they locate questions about it within a larger debate.17 Once indexically based knowledge of properties like @ is accepted, the next question is whether the complexity of these properties prevents us from having the nonindexical knowledge-by-content of them needed to know a priori that certain propositions are true at them. For many highly complex world-states, there is certainly a sense—analogous to the sense in which I can’t dunk the basketball, because I can’t jump high enough—in which I can’t derive the truth-values of propositions from sets of propositions defining them—and so cannot know a priori which propositions are true at them—because I am psychologically incapable of entertaining those sets. In what sense, then, is it knowable a priori that p is true at @ (or at w generally), for many propositions p? It is knowable a priori in the sense that it is possible to know such propositions a priori—where a use 17

Solutions to the puzzles of direct reference are proposed in Soames (2002, 2005a).

294 • Essay Ten

of It is possible to know a priori that S in a context C expresses a proposition that is true (at w1) iff there is a world-state w2 that is metaphysically possible (from w1) in which we, or beings relevantly like us, know the proposition expressed by S (in C) a priori. For example, it is knowable a priori that, at @, hundreds of thousands died in the Civil War because there is a metaphysically possible world-state in which we—or similar agents without our limitations on intelligence, memory, and attention span—can, and do, derive the proposition that hundreds of thousands died in the Civil War from the basic propositions defining @. In this way, the possibility of knowing propositions to be true at @ a priori is explained in terms of cognitively enhanced versions of ourselves whose a priori knowledge of these propositions parallels our own unproblematic a priori knowledge of propositions true at the world-state, Tiny.18 If this is right, then, in many cases, when S expresses a true proposition p, it is possible to know that p is true at @ either indexically, corresponding to (10a), or by content, corresponding to (10b). (10) a. It is true that S at this very world-state. b. It is true that S at the world-state at which it is true that Saul is a philosopher & Alfred is a logician & . . . (one conjunct for each of the basic propositions defining @). Does this conception falsely assume that these sentences semantically express the same proposition (in their respective contexts)? I don’t think so. First, it is not evident that such an assumption would be false. If worldstates are the complex properties I take them to be, then it is natural to regard the semantic contents of the italicized phrases in (10,a b) as bearing the same relation to one another as those in (11a, b) do. (11) a. This very proposition (said demonstrating the proposition that Saul is a philosopher & Alfred is a logician & . . . ) is true. b. The proposition that Saul is a philosopher & Alfred is a logician & . . . is true. Since the semantic contents of these phrases is, arguably, the same, the idea that (10a) and (10b) semantically express different propositions may be another of the familiar illusions connected with direct reference. Second, and more importantly, my argument doesn’t assume that (10a) and 18 This modal perspective also sheds light on certain puzzling examples. Suppose that the proposition expressed by S & it is never known that S is calculable from the propositions defining @. Then this proposition is true, even though this is never known—at @, or any other metaphysically possible state. However, this is no barrier to the apriority of the proposition expressed by Actually (S & it is never known that S).

Actually • 295

(10b) express the same proposition. What it assumes is that understanding and justifiably accepting (10b), and thereby knowing the proposition it semantically expresses, is sufficient for knowing (or at least being in a position to know) de re, of @, that p is true at it, and hence for knowing the singular proposition that p is true at @—whether or not this proposition is semantically expressed. Since knowing the proposition in this way doesn’t require empirical evidence, while knowing it when @ is presented indexically does, we get the result that it is possible to know that p is true at @ in two different ways, without having to decide what (10b) semantically expresses. For what it is worth, I do take (10a) to semantically express the proposition that p is true at @, and I find it plausible to think that (10b) does too. But the latter is an optional part of the package. The conclusion that some possible agents know a priori that p is true at @ raises another worry. In identifying world-states with properties incorporating complete stories of what the universe would be like if they were instantiated, I argued that it is no part of the story told by any world-state to specify what the universe would be like if a different worldstate were instantiated. This was one reason for excluding world-stateindexed ascriptions of truth-value from the basic propositions defining world-states. This exclusion may seem to be threatened by our recognition that agents at some world-states know certain of these ascriptions about other states. Let’s see whether it is. We may assume that when agents at a world-state w have beliefs, the set Sw of propositions defining w will include propositions ascribing beliefs to them. Since the truth, at w, of (12) requires only the truth of p at w*, the truth, at w, of (13) doesn’t require (12) to be an a priori consequence of Sw. (12) p is true at w*. (13) A believes truly that p is true at w*. What about (14)—which we may take to be true at w? (14) A knows that p is true at w*. If (14) is a priori calculable from (13), plus members of Sw specifying the causal sources of A’s belief, the reliability of A’s cognitive processes, etc., then Sw needn’t include any propositions from which (12) can be derived. Thus, the truth of (14) at w requires no modification of the story I have told, provided that (15) is an a priori consequence of the basic propositions about belief, reliability, etc. used in defining w. (15) If S, then A knows that S However, if (15) isn’t an a priori consequence of those propositions, then Sw will have to include (14), thereby having (12) as an a priori consequence.

296 • Essay Ten

This presents a problem. For if the propositions defining certain worldstates have world-state-indexed truth-value ascriptions about other states as a priori consequences, then instances of the contingent a priori will create trouble for the definition of epistemic possibility given by EP1. For example, let @ be the actual world-state, and w be a metaphysically possible world-state at which the contingent a priori proposition (expressed at @ by) (8a) is false, because the proposition (expressed at @ by) (8b) is true. (8) a. Saul philosophizes iff actually (i.e., it is true at @ that) Saul philosophizes b. ~ Saul philosophizes & actually (i.e., it is true at @ that) Saul philosophizes Suppose further that proposition (8b) is known, at w, by some agent A, and that the proposition, (8c), ascribing this knowledge to A, is one of the propositions defining w. (8c) A knows that: ~Saul philosophizes & it is true at @ that Saul philosophizes Under these assumptions, it is knowable a priori that w isn’t instantiated. First, it is knowable a priori that w is instantiated only if (8c) is true. So, it is knowable a priori that w is instantiated only if (8b) is true. Hence, (8d) is knowable a priori. (8d) If (8b) isn’t true, then w isn’t instantiated. Second, since (8a) is knowable a priori, it is knowable a priori that (8b) isn’t true. Thus, the claim that w isn’t instantiated is an a priori consequence of (8a) and (8d), both of which are knowable a priori. Moreover, it is possible for us (here and now) to know both of these propositions a priori when @ is presented indexically. (It doesn’t matter that A’s knowledge, in w, of @ is nonindexical.) Thus, it is knowable a priori that w isn’t instantiated. If one assumes, as I do, that metaphysically possible world-states, like w, are always epistemically possible, this result contradicts EP1, which defines an epistemically possible world-state as one that is not knowable a priori not to be instantiated. In sum, if true knowledge ascriptions aren’t a priori consequences of more basic claims about truth, belief, and other factors, then the fact that agents at some world-states know the truth-values of certain propositions at other states requires rejecting EP1, in favor of EP2. Since EP2 stipulates that w is epistemically possible iff it can’t be known a priori that w couldn’t be (or have been) instantiated, examples of the contingent a priori can’t pose problems for it. Given that the metaphysically possible world-states in which such examples are false are, by hypothesis, states that could have been instantiated, knowing that they couldn’t

Actually • 297

have been instantiated is impossible. Nor, as far as I can tell, would the replacement of EP1 by EP2 undermine other aspects of the overall picture. However, since it hasn’t been shown that knowledge claims aren’t a priori consequences of more basic claims, such replacement isn’t mandated. I am skeptical that it can be, since doing so requires much more than showing simply that knowledge isn’t definable in more basic terms. Therefore, I continue to favor EP1, while recognizing that the issue remains open. What about other ways of liberalizing the simple Carnapian framework? In setting up the structure of world-states, I used a simple, firstorder language. The basic propositions used to define world-states were expressed by atomic formulas, and their negations, relative to assignments of values to variables. The truth-values of the remaining nonmodal propositions at a world-state w were determined by the basic propositions of w, plus recursive rules for quantifiers and truth-functional operators. Truth-values of modal propositions at w followed from the truth-values of propositions at other world-states, related to w, in the total space of states. On this conception, both the richness of individual world-states and the scope of the total space of states, varies with the richness of the underlying language, the properties expressed by its predicates, the domain of objects talked about, and the uses to which the language is put. It is not important that there be one absolute space of world-states with respect to which all inquiries are conducted. What is important is that every space of states we need can be understood within the broad outlines of this framework. What might an extension of this simple system look like? No matter how syntactically and semantically complex the underlying language, nothing essential to the framework is threatened by allowing many propositions expressed by nonatomic, nonmodal formulas (relative to assignments) to count as basic. How, in this setting, should we distinguish those sets of basic propositions that define world-states from those that don’t? Putting aside syntactic criteria for completeness and consistency, we might stipulate that to be complete a world-state-defining set must determine (a priori) the truth-value of every nonmodal proposition, and to be consistent it must be epistemically possible in the sense of EP1 (or of EP2, if EP1 turns out to be unsustainable). What about propositions which, though not overtly modal, have consequences that are? Are they world-state-defining? They can be, and often are. Suppose, for example, that truths about what causes what are constitutive of w, without being a priori consequences of other, more basic, truths of w. On this supposition, causal propositions will play a role in defining w, even if they constrain which world-states are metaphysically possible from w. This is not unusual. It is routine for the constitutive truths

298 • Essay Ten

of a world-state w—e.g., that Saul Kripke is a human being and that I am the father of Greg and Brian Soames—to have modal consequences by partially determining which world-states are metaphysically possible from w. The sensitivity of metaphysical possibility to the contents of world-states is no threat to the framework. There is, however, a related question that could be raised. In excluding modal propositions from those that define world-states, the framework presupposes that world-states that agree on all nonmodal facts are identical. Thus, those that support different modal truths must also differ nonmodally. Although I find this plausible, those who don’t may avoid this result by allowing definitions of world-states sometimes to include modal propositions. A deeper question involves the ontology of the framework. Do worldstates other than @—i.e., properties that the universe doesn’t have, but which it can coherently be conceived to have—(actually) exist? Of course they do, just as other complex, but uninstantiated properties do. However, it is also true there could have been world-states different from those that actually exist. Since world-states are properties the constituents of which are objects and properties, (actually) existing world-states are those the constituents of which (actually) exist. Since there could have been many objects that don’t actually exist, and since they could have been constituents of world-states, there could have been world-states that don’t actually exist. Accordingly, what the truth of It could have been that case (i.e., is metaphysically possible) that S really requires is not that there exists a metaphysically possible world-state at which S is true, but that there could have been such a state. This, of course, precludes giving a reductive analysis of modal notions in terms of possible world-states. However, that should have been obvious anyway—since, on my account, the notion of a possible world-state is itself defined with the help of modal notions.19

References Carnap, Rudolf. 1956. Meaning and Necessity. 2nd ed. Chicago: University of Chicago Press. Grice, Paul. 1989. “Further Notes on Logic and Conversation.” In Studies in the Way of Words, 41–57. Cambridge: Harvard University Press. Kaplan, David. 1989. “Demonstratives: An Essay on the Semantics, Logic, Metaphysics, and Epistemology of Demonstratives and Other Indexicals.” In Themes from Kaplan, ed. Joseph Almog, John Perry, and Howard Wettstein with the assistance of Ingrid Deiwiks and Edward N. Zalta, 481–563. New York: Oxford University Press. 19 Thanks to Nathan Gadd, Ali Kazmi, and David Manley for useful comments on an earlier draft.

Actually • 299 Kripke, Saul A. 1980. Naming and Necessity. Cambridge: Harvard University Press. Originally published in Semantics of Natural Language, ed. Donald Davidson and Gilbert Harman, 253–355 (Boston: Reidel, 1972). Salmon, Nathan. 1989a. “On the Logic of What Might Have Been.” Philosophical Review 98:3–34. ———. 1989b. “Tense and Singular Propositions.” In Themes from Kaplan, ed. Joseph Almog, John Perry, and Howard Wettstein with the assistance of Ingrid Deiwiks and Edward N. Zalta, 331–92. New York: Oxford University Press. Soames, Scott. 2002. Beyond Rigidity: The Unfinished Semantic Agenda of “Naming and Necessity.” New York: Oxford University Press. ———. 2005a. “Naming and Asserting.” In Semantics vs. Pragmatics, ed. Zoltán Szabó, 356–82. Oxford: Clarendon Press; New York: Oxford University Press. ———. 2005b. Reference and Description: The Case against TwoDimensionalism. Princeton: Princeton University Press. ———. 2006a. “Kripke, the Necessary Aposteriori, and the Two-Dimensionalist Heresy.” In The Two-Dimensional Semantics, ed. Manuel García-Carpintero and Josep Macia, 272–92. Oxford: Clarendon Press; New York: Oxford University Press. ———. 2006b. “The Philosophical Significance of the Kripkean Necessary Aposteriori.” Philosophical Topics 16:288–309. ———. 2006c. “Understanding Assertion.” In Content and Modality: Themes from the Philosophy of Robert Stalnaker, ed. Judith Thomson and Alex Byrne, 222–50. Oxford: Oxford University Press. ———. n.d. “Kripke on Epistemic and Metaphysical Possibility: Two Routes to the Necessary Aposteriori.” In Saul Kripke, ed. Alan Berger. Cambridge: Cambridge University Press, forthcoming. Stalnaker, Robert. 1976. “Possible Worlds.” Noûs 10:65–75. ———. 1999. “Indexical Belief.” In Context and Content: Essays on Intentionality in Speech and Thought, 130–49. New York: Oxford University Press. Originally published in Synthese 49 (1981): 129–51.

This page intentionally left blank

PA RT T H R E E

Truth and Vagueness

This page intentionally left blank

ESSAY ELEVEN

What Is a Theory of Truth?

Alfred tarski’s theory of truth and its successors enjoy a perplexing double status. On the one hand, they are mathematical theories characterized by a rich class of mathematical results. On the other hand, they are commonly believed by philosophers to provide analyses of the nature of truth and, hence, to be philosophically significant. With this broader significance comes a kind of controversy not normally associated with mathematical theorems. No one disputes the correctness of Tarski’s formal results. In that sense, there is no doubt that his theory is true. However, there is considerable doubt about whether, or in what sense, it is a theory of truth. One main reason for this uncertainty is the difficulty of determining what a theory of truth ought to be. Generally, theories of truth have tried to do one or the other of three main things: (i) to give the meaning of natural-language truth predicates; (ii) to replace such predicates with substitutes, often formally defined, designed to further some reductionist program; or (iii) to use some antecedently understood notion of truth for broader philosophical purposes, such as explicating the notion of meaning or defending one or another metaphysical view. In order to do the first of these things, a theory must analyze the content of paradigmatic examples in which what is said to be true is a proposition, rather than a sentence or utterance. (1) a. The proposition that the earth moves is true. b. Church’s theorem is true. c. Everything he said is true. There are theories that try, in my opinion unsuccessfully, to do just this.1 Tarski’s theory, which restricts itself to cases in which truth is predicated This essay was presented in the spring of 1983 at Yale and Dartmouth, where I profited from useful discussion. Thanks are also due to Paul Benacerraf, Axel Buhler, John Burgess, Gilbert Harman, David Lewis, and Walter Sinnott-Armstrong for reading and commenting on an earlier draft. 1 Various versions of the “redundancy theory” fall into this category. Although these versions deal with “propositional” contexts like those in (1), they deny that ‘true’ is predicated of propositions, or anything else. Arguments against these approaches are given in Soames (1999).

304 • Essay Eleven

of sentences of certain formal languages, is not one of them. Thus, Tarski cannot be seen as even attempting to give the meaning of naturallanguage truth predicates. Nor can he be seen as attempting to use the notion of truth for broad philosophical purposes. In order to do that, one must regard truth as essentially unproblematic and philosophically productive. For Tarski, truth itself is what has to be legitimated. Once it is, it turns out to be useful for certain primarily technical purposes, but useless for ambitious philosophical programs. For example, Tarski recognized that his notion of truth could not be used to give the meanings of logical constants (or, I suspect, anything else).2 He also thought of it as epistemologically and metaphysically neutral. Thus, in “The Semantic Conception of Truth” he says: . . . the semantic definition of truth implies nothing regarding the conditions under which a sentence like . . . snow is white can be asserted. It implies only that, whenever we assert or reject this sentence, we must be ready to assert or reject the correlated sentence . . . The sentence “snow is white” is true. Thus, we may accept the semantical conception of truth without giving up any epistemological attitude we may ever have had; we may remain naive realists, critical realists or idealists, empiricists or metaphysicians—whatever we were before. The semantic conception is completely neutral toward all these issues. (1952, 33–34) It is helpful, in understanding this remark, to focus on something that the truth predicate is good for—-namely, what W. V. Quine has called “semantic ascent.”3 The simplest example of this is provided by Tarski: (2) a. Snow is white. b. The sentence ‘Snow is white’ is true. Any speaker of English knows that these sentences are at least materially equivalent. Because of this, they can often be used to convey essentially the same information. To choose (2b) is to use a semantic statement to convey information that could have been conveyed nonmetalinguistically. To do this is to engage in semantic ascent. The importance of semantic ascent is illustrated by cases like (3), in which we want to generalize.

2 3

Tarski (1952, sec. 15). Quine (1970, 10–13).

What Is a Theory of Truth? • 305

…

(3) a. Snow is white → (Grass is blue → Snow is white) b. The earth moves → (The sun is cold → The earth moves)

Each of these examples is something one could feel safe in asserting. However, if one wanted to get the effect of asserting all of them, one would have to quantify, replacing sentences with variables. In English such quantification is most naturally, though not inevitably, construed as first-order and objectual. Thus, if the variables are taken to range over sentences, we need a metalinguistic truth predicate. Semantic ascent gives us (4) For all sentences p, q (p is true → (q is true → p is true))4 That which is conveyed by (4) is closely related to that which is conveyed by (3). However, here the truth predicate is especially handy, since we don’t have the alternative of asserting each member of (3). Truth predicates can be used in the same way in more obviously philosophical cases. For example, consider: (5) There is a duplicate of our sun in some remote region of space, but we will never find (sufficient) evidence that there is. Someone who asserted this would, by contemporary standards, be counted a metaphysical realist—i.e, as being someone who thinks that what there is doesn’t depend in any way on what we may rationally believe. Of course, one can be a realist without believing (5). One may think that what there is doesn’t depend on us, while believing that there is no duplicate of our sun, or being uncertain whether there is, or while holding that evidence will someday be found to settle the matter. What, then, distinguishes realism from antirealism? One is tempted to answer that it is the belief that

…

(6) Either there is a duplicate of our sun in some remote region of space, but we will never find (sufficient) evidence that there is; or there is no duplicate of our sun in any remote region of space, but we will never find (sufficient) evidence that there isn’t; or there is intelligent life elsewhere in the universe, but we will never find (sufficient) evidence that there is; or

But this is awkward. We ought to be able to state the realist’s position without having to gesture toward an infinite list. Semantic ascent provides 4

Or, equivalently, (4'):

(4') For all sentences p, q ( p → ( q → p )) is true.

306 • Essay Eleven

a convenient way of doing this. With the help of a truth predicate and quantification over sentences, we can characterize the realist as believing, and the antirealist as denying: (7) There is at least one sentence s such that s is true (in English), but we will never find (sufficient) evidence supporting s. The relationship between (6) and (7) is like that between (2a) and (2b) and between (3) and (4). In each case, the semantic sentence may not say exactly the same thing as its nonsemantic counterpart; but if knowledge of English is assumed, the two can be used to convey essentially the same information.5 The utility of the truth predicate in stating this dispute has led some to believe that the dispute is about truth and, hence, that truth is a deeply metaphysical notion. However, there is no reason to suppose this. The realist and antirealist may agree about truth; they may even accept something like Tarski’s definition. Where they differ is in their conceptions of reality. Since statements about truth mirror direct statements about nonlinguistic reality, semantic ascent makes the truth predicate a convenient vehicle for expressing competing metaphysical views. But a convenient vehicle is all it is. As Tarski puts it, the notion of truth is completely neutral toward all these issues. The upshot of this is that Tarski’s definition of truth is neither an attempt to analyze the meaning of natural-language truth predicates nor an attempt to use the notion of truth for broad philosophical purposes. Rather, Tarski’s goal was to replace natural-language truth predicates with certain restricted, but formally defined substitutes. He thought such replacements were needed both to remove the doubts of certain scientifically minded truth skeptics and to eliminate what he took to be the incoherence in our ordinary notion brought out by the liar paradox. For Tarski, these two motivations were connected, since the paradoxes constituted one source of skepticism about truth.6 However, the truth skeptics of his day also had other, more broadly philosophical grounds for their doubts. These included the frequent use of truth in metaphysical discussions, the tendency to confuse truth with epistemological notions like certainty and confirmation, and the inability to see how acceptance 5 There are, I presume, many versions of realism, of which (6) and (7) represent only one. A different version might hold that some sentence of English is such that it is metaphysically possible for it to be true (keeping the semantics fixed) in cases in which the proposition it expresses cannot (ever) be known or rationally believed. This thesis is no more linguistic than (6) and (7) are. 6 And also of skepticism about related notions like definability. Tarski cites the paradoxes as a source of skepticism in Tarski (1983a) and, as John Burgess has pointed out to me, in Tarski (1983b).

What Is a Theory of Truth? • 307

of a truth predicate could be squared with the doctrine of physicalism and the unity of science.7 Although Tarski’s work was historically effective in alleviating each of these worries, the only one discussed by Tarski was the final one, involving physicalism.8 Tarski’s version of physicalism was a moderate one, allowing both physical and mathematical elements, without requiring the latter to be reduced to the former. Roughly, this “moderate physicalism” asserts that (i) all facts are physical or mathematical facts; (ii) all scientific (or descriptive) claims are reducible to claims about the physical and mathematical characteristics of things; and (iii) all scientific (or descriptive) concepts are definable in terms of physical and mathematical concepts.9 Tarski took this doctrine to require that truth be eliminable via an explicit, physicalistic definition. Anything else—for example, taking truth to be a primitive whose extension is fixed by a set of axioms—was deemed to be undesirable. It is worth pointing out that this emphasis on definition is primarily philosophical rather than technical. What is at issue is not the technical results achievable, but the philosophical significance of those results.10 It is possible to view a Tarski truth characterization for a language L as simply specifying the extension of ‘true’ for L, explaining how the truthvalue of a sentence depends on the semantic properties of its parts, and providing the basis for accounts of logical truth and logical consequence. Even if the truth characterization is put in the form of what is technically an explicit definition, it doesn’t have to be viewed as an explication of truth in any interesting philosophical sense. If one’s philosophical views differ from Tarski’s, one can accept his formal results while taking truth to be primitive. 7 See Hempel (1935); Carnap (1963, 63; 1949); Reichenbach (1938, sec. 22); Neurath (1959a, 1959b); Popper, (1965, 274); and Field (1972). 8 Tarski (1983a, 406). 9 Tarski’s physicalism countenanced both physical science and “logic,” where the latter was construed as including set theory and everything obtainable from it. In what follows, I will use the term ‘physicalism’ in the moderate sense of (i)–(iii) above. In particular, physicalism, in my sense, does not require the reduction of set-theoretical facts, mathematical facts, or syntactic facts about expression types. 10 There are, of course, technical issues as well. When the metalanguage contains quantifiers ranging over arbitrary subsets of the domain of the object language, an explicit definition of object-language truth is possible in the metalanguage. On the other hand, if a classical object language containing set theory has quantifiers ranging over all sets, then an explicit metalanguage definition of truth is impossible. Tarski’s emphasis on explicit definition is philosophical in the sense that he saw significant philosophical advantages in explicit definitions of truth, where they are possible.

308 • Essay Eleven

I will not comment directly on this way of viewing Tarski, but will instead concentrate on his own view of his work. I do this not out of any commitment to physicalism, but rather out of a sense that his deflationist attitude toward truth is interesting and worth defending. Tarski’s attitude needs defense because his definition of truth fails to satisfy certain initially plausible demands one might place on an explication of truth. His attitude is defensible because these demands turn out to be dubious or illegitimate. The importance of this defense extends beyond Tarski to the general question of what ought, and what ought not, to be expected from a theory of truth.

2. Tarski’s basic idea is that for certain languages L, adequate for natural science, one can define a truth predicate using only notions already expressible in L, plus certain syntactic and set-theoretic apparatus. Thus, if L is physicalistically pure and if syntax and set theory are unproblematic, then defining a metalanguage truth predicate can’t introduce any difficulties. Following Hartry Field, we can think of such a definition as divided into two parts. The first part is concerned with what Field calls “primitive denotation”; here one defines what it is for a name to refer to an object and for a predicate to apply to one or more objects. The second part of the definition defines truth in terms of primitive denotation. The end result is a metalanguage sentence: (8) For all sentences s of L, s is true iff T (s) in which T(s) is a formula with only ‘s’ free, made up entirely of logical, set-theoretic, and syntactic apparatus, plus translations of the primitives of L. If these translations are (extensionally) correct, then T(s) will be coextensive with ‘true’ over L. Tarski’s technique can be illustrated using a particularly simple example. Let L be a language whose only logical constants are ‘∨’ and ‘−’, whose nonlogical constants consist of finitely many names and one-place predicates. (R) and (A) define reference and application for L; (T) uses these notions to define truth: R. For all names n of L and objects o, n refers in L to o iff n = ‘a’ and o = Arizona, or n = ‘b’ and o = Boston, . . . (and so on for each of the names of L). A. For all one-place predicates P of L and objects o, P applies in L to o iff P = ‘C’ and o is a city, or P = ‘S’ and o is a state, . . . (and so on for each one-place predicate of L).

What Is a Theory of Truth? • 309

T. For all sentences S of L, S is true in L iff S ε the set K such that for all x, x ε K iff (i) x = Pn for some predicate P and name n of L, and there is an object o such that n refers in L to o and P applies in L to o; or (ii) x = (A ∨ B) for some formulas A and B of L, and A ε K or B ε K; or (iii) x = − A for some formula A of L and A ∉ K. Let (T′) be just like (T) except for containing the right-hand sides of (R) and (A) where (T) contains n refers in L to o and P applies in L to o, respectively. (T′) is then an explicit Tarskian truth definition for L, with ‘T(s)’ in (8) representing the right-hand side of (T′).11 Although truth definitions for richer languages are technically more interesting, their philosophical status as putative physicalistic reductions of truth is essentially the same as that of the simple definition just given. On the basis of such definitions, Tarski concluded that he had shown truth, reference, and application to be physicalistically acceptable terms. In a well-known critique of Tarski, Hartry Field (1972) argues that this conclusion is unjustified. The problem, according to Field, is that the proposed replacements for the notions of primitive denotation are not physicalistically acceptable reductions of our pretheoretic notions of reference and application. Because Field takes Tarski to have reduced truth to primitive denotation (350), he concludes that Tarski has not legitimated the notion of truth for physicalists. Field does not, of course, dispute the fact that Tarski’s definitions are extensionally correct. He maintains, however, that extensional correctness is not enough. In addition, any genuine reduction must show semantic facts about expressions to be supervenient on physical facts about their users and the environments in which they are used. Tarski’s definitions don’t do this. This can be seen by considering a simple example. Suppose that ‘Cb’ is a sentence of L and that the relevant semantic facts about it are given in (9): (9) a. ‘b’ refers (in L) to Boston. b. ‘C’ applies (in L) to cities (and only cities). c. ‘Cb’ is true (in L) iff Boston is a city. 11

Note, since there are only finitely many atomic formulas in L, that we could have got an equivalent result by substituting (i') for (i) in (T). (i') x = ‘Ca’ and Arizona is a city, or x = ‘Cb’ and Boston is a city, or x = ‘Sa’ and Arizona is a state, or x = ‘Sb’ and Boston is a state, . . . (and so on for each atomic formula).

310 • Essay Eleven

If Tarski’s definitions really specify the physicalistic content of semantic notions, then, in each case, we ought to be able to substitute the physicalistic definiens for the semantic definiendum without changing the physical fact thereby specified. Performing this substitution and simplifying results, we obtain (10) a. ‘b’ = ‘b’ and Boston = Boston. b. For all objects o, ‘C’ = ‘C’ and o is a city, iff o is a city. c. ‘Cb’ = ‘Cb’ and there is an object o such that o = Boston and o is a city, iff Boston is a city. But there is a problem in identifying these facts with those in (9). As Field points out, it is natural to suppose that the expressions of a language have semantic properties only in virtue of the ways they are used by speakers. Thus, he holds that the facts given in (9) wouldn’t have obtained if speakers’ linguistic behavior had been different.12 Since the facts in (10) are not speaker-dependent in this way, Field concludes that they are not semantic facts and that Tarski’s attempted reduction fails. Tarski’s truth predicate is both physicalistic and coextensive with ‘true in L’; but it is not, according to Field, a physicalistic conception of truth. On Field’s view, Tarski’s truth characterization inherits its inadequacy as a reduction from the pseudo-reductions that constitute its base clauses. Thus, Field’s strategy for solving the problem is to provide genuine reductions for the notions of primitive denotation, on something like the model of the causal theory of reference. The picture that emerges from his discussion is one in which an adequate definition of truth is a twostage affair. Stage 1 is Tarski’s reduction of truth to primitive denotation. Stage 2 is the imagined causal-theory-like reduction of the notions of a name referring to, and a predicate applying to, an object in a language.13 If the physical facts that determine denotation in one language do so in all, then these relations will hold between expressions and objects, for variable ‘L’. When logical vocabulary and syntax are kept fixed, the result is a notion of truth that is not language-specific, but is itself defined for variable ‘L’. Although the resulting picture appears rosy, there are several problems with it. One concerns reference to abstract objects, for which a causal account seems problematic. Another involves Quinean worries about ontological relativity and referential indeterminacy. These, of course, are obstacles to a physicalistic reduction of primitive denotation. However, 12 I use the phrase ‘linguistic behavior’ in a broad sense to include all facts about speakers relating to their use of language. 13 Field also includes the notion of a function sign being fulfilled by a pair of objects. In the interest of simplicity, I am ignoring this.

What Is a Theory of Truth? • 311

there are other difficulties which become clear when one notices that Field has understated his objection to Tarski. If the alleged dependence of semantic facts on facts about speakers shows that Tarski has not reduced primitive denotation to physical facts, then the very same point shows that he has not reduced truth to primitive denotation. This can be seen by considering a pair of elementary examples. Imagine two languages, L1 and L2, which are identical except that in L1 the predicate ‘R’ applies to round things, whereas in L2 it applies to red things. Owing to this difference, certain sentences will have different truth conditions in the two languages. (11) a. ‘Re’ is true in L1 iff the earth is round. b. ‘Re’ is true in L2 iff the earth is red. Under Tarski’s original definition, this difference will be traceable to the base clauses of the respective truth definitions, where the applications of predicates are simply listed. Field’s objection to this is that although Tarski’s definitions correctly report that ‘R’ applies to different things in the two languages, they don’t explain how this difference arises from the way in which speakers of the two languages use the predicate. What Field fails to point out is that exactly the same objection can be brought against Tarski’s treatment of logical vocabulary and syntax in the recursive part of his definition. This time let L1 and L2 be identical except for their treatment ‘∨’. (12) a. A formula (A ∨ B) is true in L1 (with respect to a sequence s) iff A is true in L1 (with respect to s) or B is true in L1 (with respect to s). b. A formula (A ∨ B) is true in L2 (with respect to a sequence s) iff A is true in L2 (with respect to s) and B is true in L2 (with respect to s). Owing to this difference, sentences containing ‘∨’ will have different truth conditions in the two languages. In order to satisfy Field’s requirements on reduction, it is not enough for a truth characterization to report such differences. Rather, such differences must be explained in terms of the manner in which speakers of the two languages treat ‘∨’.14 Since Tarski’s truth definitions don’t say anything about this, their recursive clauses should be just as objectionable to the physicalist as the base clauses. This means that Field’s strategy of achieving a genuine reduction of truth by supplementing Tarski with nontrivial definitions of primitive denotation cannot succeed. The reason it can’t is that, given Field’s strictures 14 Presumably, speakers of L1 differ in some way from speakers of L2 regarding their beliefs, intentions, attitudes, brain-states, or conditioned responses involving ‘∨’.

312 • Essay Eleven

on reduction, Tarski has not reduced truth (for standard first-order languages) to primitive denotation. At best he has reduced it to the class of semantic primitives listed in (13):15 (13) the notion of a name referring to an object the notion of a predicate applying to objects the notion of a formula being the application of an n-place predicate P to an n-tuple of terms t1 . . . tn the notion of a formula A being a negation of a formula B the notion of a formula A being a disjunction of formulas B and C the notion of a formula A being an existential generalization of a formula B with respect to a variable u and a domain D of objects This way of looking at things requires a restatement of every clause in Tarski’s truth definition. For example, the recursive clause for negation, which had been given by (14a), is now given by (14b). (14) a. If A = − B, then A is true in L (with respect to a sequence s) iff B is not true in L (with respect to s). b. If A is a negation of a formula B, then A is true in L (with respect to a sequence s) iff B is not true in L (with respect to s). The resulting abstraction extends the generality of the truth definition to classes of first-order languages that differ arbitrarily in syntax, plus logical and nonlogical vocabulary. Although this generality is appealing, it has a price. Whereas the original definitions simply stipulated that − A is a negation, A ∨ B is a disjunction, and ∃xAx is an existential generalization over a range D of objects, the revised definition doesn’t provide a clue about which formulas fall into these categories. Moreover, Field’s physicalist now has to provide reductions of each of these semantic notions. How might this be done? We are accustomed either to using truth to explain the logical notions or to taking them as primitive, while stipulating 15 Field (1972, nn. 5, 10) partially anticipates this point. In n. 5, he notes that in model theory quantifiers are given an “unusual” semantics in which they range over the members of some specified set, rather than over all (actually existing) things. In such a case, Field claims, Tarski has reduced truth to primitive denotation, plus the notion of the range of the quantifiers. (For Tarski this constituted the usual case, since it is only when the range of quantifiers is restricted that explicit truth definitions can be given—for languages with a certain minimal richness.) In n. 10 Field notes, without specifying, the existence of problems that must be faced when the definition of truth is generalized so as not to contain a particular logical vocabulary.

What Is a Theory of Truth? • 313

that certain symbols are to count as instances of them. Neither of these policies is open to Field. He cannot characterize negation as a symbol that attaches to a formula to form a new formula that is true (with respect to a sequence) iff the original is false (with respect to the sequence); for that would make the reduction of truth to the notions in (13) circular. Nor can he take negation to be primitive and stipulate that − S is to be the negation of S; for that would fail to give the facts about speakers that explain the semantic properties of − S. Although there are alternative approaches, none that I know of is clearly successful.16 For example, in The Roots of Reference17 Quine attempts to characterize truth-functional operators in terms of community-wide dispositions to assent and dissent. He ends up concluding that indeterminacy between classical and intuitionist construals of the connectives is inevitable. Although I do not accept Quine’s argument for this,18 I do think that the task confronting Field’s physicalist is nontrivial. The problems involved in reducing primitive denotation to physical facts are hard enough; adding the logical notions makes the job that much harder. As I have stressed, the source of this difficulty is the demand that semantic facts be supervenient on physical facts about speakers. In effect, this demand limits adequate definitions to those which legitimate substitution for semantic notions in contexts like (15) and (16). (15) If L-speakers had behaved differently (or been differently constituted), then ‘b’ wouldn’t have referred (in L) to Boston, and ‘C’ wouldn’t have applied (in L) to cities, and ‘Cb ∨ Ca’ wouldn’t have been true (in L) iff Boston was a city or Arizona was a city. (16) The fact that L-speakers behave as they do (and are constituted as they are) explains why ‘b’ refers (in L) to Boston, etc. Field’s critique of Tarski is based on the conviction that there ought to be a way of spelling out (15) and (16) so that they come out true when physicalistic substitutes replace semantic terms and their initial clauses are construed as expressing contingent physical possibilities.19 As we have seen, Tarski’s definition doesn’t have this character. 16

The most interesting, in my opinion, is briefly sketched in Harman (1982). Quine (1974). 18 For an excellent critique of Quine, see Berger (1980). 19 In stating this requirement in terms of the replacement of a semantic term by its physicalistic definiens, I have tacitly relied on Tarski’s emphasis on explicit definition. However, I don’t think the philosophical point of the requirement depends on this. In cases in which only an axiomatic treatment is possible, Field could require that the axioms governing ‘true’, together with empirical facts about speakers and their environments, have statements of type (15) and (16) as consequences. 17

314 • Essay Eleven

3. It is helpful in understanding the issues at stake to compare this criticism of Tarski to a parallel objection. Whereas Field’s critique is based upon a view about the relationship between speakers and semantic properties like truth, the parallel objection is based on a view about the relationship between meaning and truth. It is widely held that the meaning of a sentence is closely related to its truth conditions and that knowledge of the one constrains knowledge of the other. Thus, many philosophers would accept arbitrary instances of (17) and (18): (17) If ‘S’ had meant in L that p, then ‘S’ would have been true in L iff p. (18) If x knows that it is not the case that ‘S’ is true in L iff p, then x knows (or has sufficient grounds for concluding) that ‘S’ does not mean in L that p.20 A natural demand growing out of this view is that substituting an adequate explication for ‘true in L’ in (17) and (18) should result in true sentences with contingent antecedents.21 20 (18) is considerably weaker than the claim that knowledge of truth conditions is sufficient for knowledge of meaning. (18) says only that knowledge of truth conditions is capable of providing some information about meaning. In effect, it says that even if knowledge that

(i) ‘S’ is true in L iff q. is not sufficient for knowing that (ii) ‘S’ means in L that q. it should be sufficient for knowing that (iii) ‘S’ does not mean in L that p. (where the sentences replacing ‘p’ and ‘q’ are obviously incompatible). 21 Although the contexts in question are intensional, this demand does not require that an adequate explicatum for the pretheoretic notion of truth be intensionally equivalent to an ordinary, pretheoretic truth predicate. Rather, it requires that all legitimate theoretical purposes served by the explicandum (truth), be equally well served by the explicatum. For example, if knowledge of that expressed by (i) . . . is true . . . is used to help explain the nature of some capacity (say, the capacity to understand sentences), then knowledge of that expressed by (ii) . . . T . . . (where the explicatum T is substituted for ‘is true’) should be sufficient for the same purpose. An explication that meets this requirement of theoretical productivity will allow the explicandum to be eliminated from one’s total scientific and philosophical theory without

What Is a Theory of Truth? • 315

As before, it is obvious that Tarski’s definition does not satisfy this demand. For example, let ‘Ws’ be a sentence of L meaning that snow is white. Using Tarski’s definition of truth, we can produce the following counterparts of (17) and (18):22 (17T) If ‘Ws’ had meant in L that snow is black, then it would have been the case that snow was white iff snow was black. (18T) If x knows that it is not the case that snow is white iff snow is black, then x knows (or has sufficient grounds for concluding) that ‘Ws’ does not mean in L that snow is black. These are clearly not what the defender of (17) and (18) has in mind. The reason they aren’t is that Tarski’s set-theoretic truth predicate doesn’t impose any conditions on the meanings of the sentences to which it applies. To be sure, Tarski wouldn’t count any predicate T as a truth predicate unless α is T were materially equivalent to any metalanguage paraphrase of the object-language sentence named by α. On the basis of this, one might interpret Tarski as implicitly supposing that instances of (19) are necessary or a priori. (19) If ‘T’ is a truth predicate for L, and ‘S’ means in L that p, then ‘S’ is T iff p. However, this is quite different from maintaining that if ‘T’ in (20) is replaced with a truth predicate for L, then the resulting instances of the schema will be necessary or a priori: (20) If ‘S’ means in L that p, then ‘S’ is T iff p It is this that the advocate of (17) and (18) demands and that Tarski appears not to provide.23 loss of explanatory power. Thus, substitution of explicatum for explicandum in intensional contexts contained in one’s total explanatory theory must be countenanced, even if such substitution is not always countenanced in ordinary discourse. The qualification in note 19 above regarding substitution and explicit definition also applies here. 22 (17T) and (18T) are simplifications of the sentences that would result from substituting Tarski’s explicatum (the right-hand side of (T') in section 2) for ‘true in L’ in (17) and (18). The simplifications are based on the fact that, where T is Tarski’s explicatum, ‘Snow is white’ is T and ‘Snow is white’ are necessarily equivalent (in the presence of elementary set theory). In light of this equivalence, replacing one with the other should not affect the philosophical issues at stake in (17) and (18). 23 Hilary Putnam has used a version of the argument involving (17)/(17T) against Tarski (in a lecture at Princeton, fall 1982). Michael Dummett has used a version of the argument involving (18)/(18T) against Tarski (in the preface to 1978b and in 1978a). The arguments given above are intended as stand-ins for a variety of related arguments, all designed to show that Tarski’s notion of truth has nothing to do with semantic interpretation or understanding. For example, it is probably best to understand Davidson not as attempting to analyze

316 • Essay Eleven

4. We have, then, two major objections to Tarski. Field demands that semantic properties be dependent on speakers in a way in which Tarski’s substitutes are not. A familiar sort of semantic theorist demands that meaning and truth conditions be contingent, but analytically connected, properties of a sentence in a manner incompatible with Tarski. The only way to defend Tarski’s philosophical interpretation of his work is to reject these demands. Although this might initially seem to be a desperate strategy, it is not. Think of a standard first-order language L as a triple (SL, DL, FL), where SL is a family of sets representing the various categories of well-formed expressions of L; DL is a domain of objects; and FL is a function that assigns objects in DL to the names of L, subsets of the domain to one-place predicates of L, and so on.24 Let J be a class of such languages. Truth can now be defined in nonsemantic terms for variable ‘L’ in J in a straightforward meaning in terms of truth, but rather as eliminating the notion of meaning in favor of the notion of truth. Since (18) utilizes the notion of meaning, a defender of the Davidson of “Truth and Meaning” might want to trade it for something like (i): (i) If x knows that which is expressed by the relevant instance of ‘S’ is true in L iff p for each sentence of L, then x is a competent speaker of L. If ‘true in L’ is understood as short for the definiens provided by Tarski, (i) is as absurd as (18T). Just this sort of absurdity is present in familiar and often repeated remarks like the following (which would allow Tarski’s definiens to be the central notion in a theory of meaning): (T) s is T if and only if p What we require of a theory of meaning for a language L is that without appeal to any (further) semantical notions it place enough restrictions on the predicate ‘is T’ to entail all sentences got from schema T when ‘s’ is replaced by a structural description of a sentence of L and ‘p’ by that sentence. . . . It is worth emphasizing that the concept of truth played no ostensible role in stating our original problem. That problem, upon refinement, led to the view that an adequate theory of meaning must characterize a predicate meeting certain conditions. It was in the nature of a discovery that such a predicate would apply exactly to the true sentences. I hope that what I am doing may be described in part as defending the philosophical importance of Tarski’s semantical concept of truth. (Davidson 1971, 455–56) Earlier statements of essentially the same absurdity can be found in Carnap (1947, 5–6) and in Carnap (1943, sec. 7 ). 24 This sort of construction is familiar from model theory. However, its use here is different from model-theoretic treatments. Here we are not defining truth in L relative to a model, but rather truth in L (simpliciter) for an enriched conception of a language. This way of looking at things was suggested to me from two sources: Lewis (1975); and one of Saul Kripke’s seminars on truth, Princeton, 1982.

What Is a Theory of Truth? • 317

Tarskian fashion. The only significant change from before is that the notions of primitive denotation are no longer given language-specific list definitions, but rather are defined for variable ‘L’ using the “interpretation” functions built into the languages. In particular, a name n refers to an object o in a language L iff FL(n) = o.25 The resulting truth predicate is just what is needed for metatheoretical studies of the nature, structure, and scope of a wide variety of theories. What the truth definition does not do is tell us anything about the speakers of the languages to which it applies. On this conception, languages are abstract objects, which can be thought of as bearing their semantic properties essentially. There is no possibility that expressions of a language might have denoted something other than what they do denote; or that the sentences of a language might have had different truth conditions. Any variation in semantic properties (across worlds) is a variation in languages. Thus, semantic properties aren’t contingent on anything, let alone speaker behavior. What is contingent on speaker behavior is which language a person or population speaks and which expression a given utterance is an utterance of. Let L1 and L2 be two languages in J which are identical except for the interpretations of certain nonlogical vocabulary—perhaps the color words in L1 are shape words in L2. We can easily imagine a situation in which it is correct to characterize L1, rather than L2, as the language of a given population. To ask what such a characterization amounts to, and what would justify it, is to ask not a semantic question about the languages, but a pragmatic question about their relation to speakers. Although Tarski had nothing to say about this relation, other philosophers have. David Lewis, using a different, but equally abstract, conception of language has proposed (1975) an analysis in terms of a convention of truthfulness and trust. Discussions of what Donald Davidson calls “radical interpretation” can also be reconstructed as dealing with this issue. For physicalists, the interesting question is whether any purely physical explication can be given. If so, then the physicalist can accept both semantic notions that apply to sentences and those which apply to utterances. If not, then either the latter or physicalism itself must be abandoned. It is interesting to note that much of Hartry Field’s concern is with the semantic properties of utterances rather than sentences. In describing the physicalist’s position he says: People utter the sounds ‘Electrons have rest mass but photons don’t’ . . . , and we apply the word ‘true’ to their utterances. We don’t 25 Note, FL is a purely mathematical object—a set of pairs, if you like. Thus, it does not incorporate any undefined semantic notions. This was one of the points noted by Kripke in the seminar mentioned in note 24.

318 • Essay Eleven

want to say that it is a primitive and inexplicable fact about those utterances that they are true, a fact that cannot be explicated in nonsemantic terms; this is as unattractive to a physicalist as supposing that it is a primitive and inexplicable fact about an organism at a certain time that it is in pain. (1972, 359) In effect, Field criticizes Tarski for not providing a physicalistically acceptable truth predicate of utterances. But Tarski wasn’t concerned with utterances. Thus, confronted with the question (i) In virtue of what are certain sounds utterances which are true in L? Tarski’s response ought to be to break it up into two subsidiary questions: (ii)

In virtue of what are certain sounds utterances in L of its sentences? (iii) In virtue of what are sentences of L true (in L)? Whereas Tarski answered the second question, the first was no part of his task. It is hard to see how Field himself could avoid this division of labor. At one point he suggests that in order to handle ambiguous and indexical expressions, truth definitions should be formulated in terms of tokens rather than types (351–53). The idea is that utterances are contextually disambiguated and that semantic notions should apply to unambiguous entities. This means that all clauses in a truth definition must be formulated as applying to tokens. To this end, Field reformulates the clause for negation as (21): (21)

A token of − e is true (with respect to a sequence) iff the token of e that it contains is not true (with respect to the sequence). (352)

However, this won’t do. As I indicated earlier, Field can’t accept any truth definition in which a certain syntactic form is simply stipulated to be a negation; for to do this would be to fail to explicate the facts about speakers in virtue of which negative constructions have the semantic properties they do. Instead, (21) must be replaced with something along the lines of (22). (22) A token of a formula A, which is a negation of a formula B, is true (with respect to a sequence) iff some designated token of B is not true (with respect to the sequence). But now there is a problem. Even if the notion of a formula A being a negation of a formula B can be given a physicalistic definition in terms

What Is a Theory of Truth? • 319

of the behavior of speakers, there is no clear way of specifying the relevant token of B needed in (22); indeed, there is no way of ensuring that it will exist. If we could count on utterances of negative sentences always containing, as proper parts, utterances of the sentences they are negations of, then the problem would not arise. Although this is a feature of certain artificial languages, it is not a characteristic of natural languages actually spoken by people. In order to avoid arbitrarily restricting truth definitions to (utterances involving) this subset of artificial languages, we need some way of eliminating undue dependence on empirically unreliable tokens. The most straightforward way of doing this is to define truth for types, thereby acknowledging the theoretical division of labor I have attributed to Tarski.26 Once this is done, the physicalist is free to accept Tarski-like truth definitions applying to sentences, while leaving it open whether the pragmatic relations between languages, expressions, speakers, and utterances are purely physicalistic.27 It should be emphasized that although the linguistic threat to physicalism has been moved from semantics to pragmatics, it is still a serious one. It is by no means evident that physicalistic reductions of the crucial relations can be given. One physicalist who seems to think they cannot be given is Quine. Although he doesn’t conceptualize matters in just the way that I have, it is illuminating to interpret him as accepting Tarski’s semantic definitions while rejecting any physicalistic reduction of the pragmatic notions. On this interpretation, there is no indeterminacy about the claim that ‘rabbit’ refers to rabbits in a certain specified language, call it “English”, or about the claim that ‘gavagai’ refers to rabbits in another language, call it “Junglese”. What is indeterminate is whether I speak English, as opposed to some related rabbit-stage language, and whether the native speaks Junglese, as opposed to some similar counterpart. The upshot of this is that it is all right for a Quinean physicalist to use a Tarskian language to describe the world, and even to attribute Tarskian semantic properties to expressions in that language. What he cannot do is identify the language he is using. When it comes to describing linguistic behavior—even one’s own—identifiable Tarskian languages are excluded 26 Ambiguity can then be treated as a case of homonomy. For example, instead of thinking that English contains a single (ambiguous) word type ‘bank’, one can take English to contain two different words, ‘bank1’ and ‘bank2’, whose tokens are phonologically identical. The contextual factors that Field relies on to disambiguate tokens can then be thought of as determining whether particular utterances are tokens of the type ‘bank1’ or the type ‘bank2’. 27 Acknowledging the need to formulate truth definitions in terms of types does not force one to think of the semantic properties of sentences as invariant from world to world and not dependent on the properties of tokens. However, it does make this a natural position.

320 • Essay Eleven

in favor of dispositions to verbal behavior. The strain in this position is a measure of the challenge that language use presents to physicalism. What is not problematic is the physicalist’s acceptance of Tarski.

5. This discussion illustrates a general strategy for answering Tarski’s critics. Field’s objection was that Tarski’s semantic properties are not dependent on facts about speakers. The Tarskian reply is that nothing is lost by thinking of semantics abstractly and relegating the interpretation of speakers’ behavior to pragmatics. In so doing, one gains the advantages of a truth predicate for metatheoretical discussions, while retaining the ability to raise deep philosophical problems in other areas. As I pointed out earlier, Field’s is not the only objection to Tarski. Any theory of semantic competence that makes knowledge of truth conditions the central notion implicitly rejects Tarski’s claim to have provided a notion of truth adequate for all theoretical purposes. The defense against this objection is that such theories are flawed in any case. The problem with these theories lies in specifying what truth conditions are in such a way that knowledge of them is necessary and sufficient for understanding. If we assume that truth conditions involve the notion of truth, then it is natural to suppose that they are given by T-sentences of the form (23): (23) ‘S’ is true in L ≡ P. (Instances are formed by replacing ‘P’ with a sentence that means the same as the sentence replacing ‘S’.) However, it is easy to show that knowing the propositions expressed by T-sentences is neither necessary nor sufficient for understanding meaning (where ‘true’ is taken to be a non-Tarskian primitive and ‘≡’ represents either material or necessary equivalence). Thus it is not obvious that what one knows when one understands a language involves the notion of truth at all. If it doesn’t, it may be that nothing is lost by adopting a Tarski-like explication of truth together with an independent account of semantic competence. Although I won’t try to show it here, I think that this is the right approach for both truth and semantic competence. This does not mean that Tarski’s semantic predicates really are adequate for all theoretical purposes. Saul Kripke’s theory of truth is a genuine advance on Tarski’s treatment of the liar.28 In addition, semantic predicates for richer languages, as

28

Kripke (1975).

What Is a Theory of Truth? • 321

well as for propositions, are needed. What does seem right about Tarski’s approach is its deflationist character. Theories of truth for sentence types need not specify the facts about speakers in virtue of which their utterances have content; nor should such theories be seen as issuing in theorems knowledge of which is necessary and sufficient for semantic competence. Instead, theories, or definitions, of truth should provide accounts of the content of familiar truth predications, while resolving the semantic paradoxes (and their propositional variants). Beyond this, and the attendant dissolution of confusions, it is best not to expect too much. Truth is a useful notion, but it is not the key to what there is, or to how we represent the world to ourselves through language.

References Berger, Alan. 1980. “Quine on ‘Alternative Logics’ and Verdict Tables.” Journal of Philosophy 77:259–77. Carnap, Rudolf. 1943. Introduction to Semantics. Cambridge: Harvard University Press. ———. 1947. Meaning and Necessity. Chicago: University of Chicago Press. ———. 1949. “Truth and Confirmation.” In Readings in Philosophical Analysis, ed. Herbert Feigl and Wilfred Sellars, 119–27. New York: Appleton-CenturyCrofts. ———. 1963. “Intellectual Autobiography.” In The Philosophy of Rudolf Carnap, ed. Paul A. Schilpp. 3–84. LaSalle, Ill.: Open Court. Davidson, Donald. 1971. “Truth and Meaning.” In Readings in the Philosophy of Language, ed. Jay F. Rosenberg and Charles Travis, 455–80. Englewood Cliffs, N.J.: Prentice-Hall. Dummett, Michael. 1978a. “Truth.” In Truth and Other Enigmas, 1–24. Cambridge: Harvard University Press. Originally published in Proceedings of the Aristotelian Society 59 (1959): 141–62. ———. 1978b. Truth and Other Enigmas. Cambridge: Harvard University Press. Field, Hartry. 1972. “Tarski’s Theory of Truth.” Journal of Philosophy 69:347–75. Harman, Gilbert. 1982. “Beliefs and Concepts: Comments on Brian Loar, ‘Must Beliefs Be Sentences?’ ” Proceedings of the Biennial Meeting of the Philosophy of Science Association 2:654–61. Hempel, Carl G. 1935. “On the Logical Positivists’ Theory of Truth.” Analysis 2, no. 4: 49–59. Kripke, Saul. 1975. “Outline of a Theory of Truth.” Journal of Philosophy 72:690–716. Lewis, David. 1975. “Languages and Language.” In Language, Mind, and Knowledge, ed. Keith Gunderson, 3–35. Minneapolis: University of Minnesota Press. Neurath, Otto. 1959a. “Protocol Sentences.” In Logical Positivism, ed. A. J. Ayer, 199–208. Glencoe, Ill.: Free Press.

322 • Essay Eleven ———. 1959b. “Sociology and Physicalism.” In Logical Positivism, ed. A. J. Ayer, 282–317. Glencoe, Ill.: Free Press. Popper, Karl. 1965. The Logic of Scientific Discovery. New York: Harper Torchbook. Quine, W. V. 1970. The Philosophy of Logic. Englewood Cliffs, N.J.: PrenticeHall. ———. 1974. The Roots of Reference. LaSalle, Ill.: Open Court. Reichenbach, Hans. 1938. Experience and Prediction: An Analysis of the Foundations and the Structure of Knowledge. Chicago: University of Chicago Press. Soames, Scott. 1999. Understanding Truth. New York: Oxford. Tarski, Alfred. 1952. “The Semantic Conception of Truth.” In Semantics and the Philosophy of Language, ed. Leonard Linsky, 13–47. Urbana: University of Illinois Press. ———. 1983a. “The Establishment of Scientific Semantics.” In Logic, Semantics, Metamathematics: Papers from 1923 to 1938, ed. John Corcoran, trans. J. H. Woodger, 401–8. Indianapolis: Hackett. ———. 1983b. “On Definable Sets of Real Numbers.” In Logic, Semantics, Metamathematics: Papers from 1923 to 1938, ed. John Corcoran, trans. J. H. Woodger, 110–42. Indianapolis: Hackett.

ESSAY TWELVE

Understanding Deflationism

A Deflationary Conception of Deflationism My aim here will be to say what is right about deflationism about truth—including how deflationism is best understood, and why, in the end, truth is deflationary. I will do this without presenting any one deflationary theory that tells us what truth is, what the predicate ‘true’ means, or what it is to understand this predicate. I have three reasons for avoiding this level of specificity. First, any precise and specific theory would require a full-fledged analysis of the liar and related paradoxes. Although much progress has been made toward this end, and although this is where the real philosophical action in theories of truth now lies, in my opinion too many questions remain unresolved for us to be fully confident about any specific and precise analysis of truth. When all is said and done, we will, of course, want our understanding of ordinary ascriptions of truth to nonparadoxical bearers to harmonize with the more complex story that emerges from an account of the paradoxes. To this end, we should try to make sure that insights from these different investigations inform and constrain one another—as I will try, briefly, to illustrate at the very end of my discussion. However, I am not sure that we now know enough to finish the job. Certainly I don’t. The second reason for avoiding a specific deflationary theory is that every such theory that I know of is either clearly false, or at least dubious.1 Apart from the unacceptable, or dubious, details of particular proposals, one large and daunting difficulty stands in the way of developing any deflationary analysis of what ‘true’ means, or of what it is to understand the truth predicate. In order to avoid circularity, any such analysis must avoid notions that themselves presuppose the notion of truth being analyzed. For example, if one appeals to meaning, propositions, equivalence relations, or the like, one undertakes an obligation to explain how these can be understood independent of truth itself. This is a difficult problem, to say the least.

1

See Soames (1999, chap. 8) for criticisms of some well-known deflationist theories.

324 • Essay Twelve

My final reason for not offering a specific deflationary theory of truth is my belief that we can clarify the notion of truth in a way that gives us most of what we need for ordinary philosophical purposes, without resolving the contentious issues raised by any complete and precise theory. Instead of proposing any such theory, I will offer a conception of what it is for a truth predicate, or a theory of such a predicate, to be deflationary. I will then argue that our ordinary truth predicate of propositions is deflationary, but that deflationary theories of truth for sentences don’t tell the whole story about sentential truth. I begin with what often is taken to be a paradigm of a deflationary truth predicate—a Tarskian truth predicate for a restricted fragment E of present-day English, encompassing the constructions of the predicate calculus, but not containing any ordinary semantic predicates. Imagine that we have an explicit definition of a Tarskian truth predicate for this language, formulated in a slightly extended metalanguage. The definition makes use of elementary facts about the ontology and syntax of E, plus a little bit of set theory, but it does not employ any undefined semantic terms. Instead, it relies on familiar Tarski-style list-definitions of reference and denotation for the nonlogical vocabulary, plus the usual recursive clauses for the logical vocabulary. The definition is understood as introducing, and giving the content of, the predicate ‘is a Tarski-true sentence of E’; and the definiens provided by the definition is substitutable for this predicate, without loss of meaning, in every sentence of the metalanguage in which it occurs. For each sentence S of E, the claim that S is a Tarskitrue sentence of E is materially equivalent to the claim made by S itself. That isn’t all; the claim made by S is both a necessary and an a priori consequence of the claim that S is a Tarski-true sentence of E; in addition, if we take the ontology and syntax of E as given, then the necessary and a priori consequence relations run in the other direction as well—in other words, the claim that S is a Tarski-true sentence of E is a necessary and a priori consequence of the claim made by S together with a claim specifying the ontology and syntax of E. But for this relatively minor complication, S and the claim that S is Tarski-true in E are necessarily and a priori equivalent. Given elementary facts about set theory, anyone who understands the claims made by the two sentences is in a position to infer one from the other without further ado. In addition, since these inferences are relatively obvious and straightforward, it is plausible to suppose that the reason that a particular sentence of E, for example ‘The moon affects the tides’, is Tarski-true is simply the reason that the moon affects the tides. These features are hallmarks of a deflationary truth predicate. It is instructive to compare this deflationary predicate with the oneplace predicate ‘is a true sentence of E’, which is formed by filling in the second argument place of the relational predicate ‘is a true sentence of’

Understanding Deflationism • 325

that occurs in ordinary language. Since one can know that S is a true sentence of E without understanding S, and without believing or being in a position to come to believe the claim it expresses, neither the claim made by S nor the claim that S is a Tarski-true sentence of E is an a priori consequence of the claim made by ‘S’ is a true sentence of E, using our ordinary truth predicate. If, as seems plausible, sentences of a natural language mean what they do contingently, and sometimes even change their meanings, then neither the claim made by S, nor the claim that S is a Tarski-true sentence of E, is a necessary consequence of the claim that S is a true sentence of E, either. Analogous points run in the opposite direction. Since one can believe the claim expressed by S without knowing what S means, and without being in a position to determine that S is a true sentence of E, the claim that S is a true sentence of E is not an a priori consequence of the claim made by S, or of the claim that S is a Tarskitrue sentence of E. And again, if, as seems plausible, the sentences of natural language have their truth conditions contingently, then the claim that S is a true sentence of E is not a necessary consequence of the claim made by S, or of the claim that S is a Tarski-true sentence of E. Therefore, although the claim expressed by each sentence S of E can be regarded as both necessarily equivalent and a priori equivalent to claim expressed by the corresponding metalanguage sentence ‘S’ is a Tarski-true sentence of E, by no stretch of the imagination can the claim expressed by S be taken to be either necessarily or a priori equivalent to the claim expressed by the sentence ‘S’ is a true sentence of E, which contains our ordinary truth predicate of sentences. This is connected with another difference. The reason that the sentence ‘The moon affects the tides’ is a true sentence of E is not the same as the reason that it is a Tarski-true sentence of E. As we have seen, it is plausible to take the reason that the sentence is Tarski-true to be nothing more than the reason that the moon affects the tides. This reason may be part—but only part—of the reason that the sentence is true in the ordinary sense. The sentence ‘The moon affects the tides’ is a true sentence of E because of two things: (i) that it means that the moon affects the tides, and (ii) that the moon does affect the tides. Since the first of these reasons is a substantive matter, the predicate ‘is a true sentence of E’ is not deflationary. So is truth for E deflationary or not? That depends on what one means in asking the question. If one means ‘Is there a deflationary truth predicate for E—a predicate that applies to all and only the truths of E which is such that the claim that it applies to a sentence is equivalent, in an appropriately strong sense, to the claim made by the sentence itself?’—then the answer is ‘Yes, truth for E is deflationary, in this sense’. However, if one means, ‘Is the ordinary truth predicate ‘is a true sentence of E’ deflationary?’—then the answer is ‘No, it isn’t deflationary’. A more interesting question than either

326 • Essay Twelve

of these is whether one needs a truth predicate for E beyond the deflationary Tarskian one. This depends on what one needs a truth predicate for in the first place. It is clear that for some purposes—for example securing the advantages of semantic ascent, and providing the basis for a model-theoretic treatment of logical consequence in E—one doesn’t need anything more than Tarski-style predicates. However, it is far from obvious that we never need a nondeflationary truth predicate of sentences. Surely we need a substantive theory of meaning. If some notion of sentential truth plays an essential role in such a theory, then it won’t be deflationary. Even if a theory of meaning can be gotten without a notion of sentential truth, any such theory that assigns meanings, or propositions, to sentences will provide the resources for defining a nondeflationary truth predicate. Since we can define sentential truth in terms of the truth of that which a sentence means or expresses, a needed nondeflationary notion of sentential truth should be extractable from a theory of meaning, provided that we can make sense of what it is for that which is expressed by a sentence to be true. Suppose, then, that one agrees that we cannot get everything we need from deflationary truth predicates of sentences. One still might wonder whether propositional truth is deflationary, where a proposition is the sort of thing that can be believed, asserted, denied, assumed, and so on. What would it mean to say that the property of being a true proposition is deflationary? As I indicated earlier, in answering this question I will put aside any consideration of the liar, or related paradoxes. With this limitation, we may consider pairs of nonparadoxical and nonpathological propositions p and the proposition that p is true.2 My suggestion is that to say propositional truth is deflationary is to say: (i) that p and the proposition that p is true are trivial, necessary, and a priori consequences of one another,3 and (ii) that any warrant for asserting, believing, denying, 2 In restricting myself to nonparadoxical and nonpathological propositions, I am allowing cases in which propositional truth is ascribed to propositions which themselves ascribe truth (or untruth) to other propositions, provided that all these propositions are unparadoxical and grounded. Thus, although I do not restrict the propositional truth predicate in the way in which Tarski’s sentential truth predicates are restricted, I remain noncommittal about what should be said about paradoxical and pathological (ungrounded) cases. 3 The claim that a proposition p is an a priori consequence of the proposition that p is true should be understood in the following sense: there is a way of apprehending the proposition that p is true (which also involves apprehending p) such that when the proposition that p is true is apprehended in this way, the agent is able, a priori, to infer p. This corresponds to what I called a priori consequence in the weak sense in Soames (1999, chap. 8, n. 3). If we had selected a closely related truth claim—namely the one expressed by it is true that S—this complication would not have been required, and we could have appealed to what I there called a priori consequence in the strong sense. These technical differences don’t matter for present purposes.

Understanding Deflationism • 327

doubting, assuming (or taking any of a variety of related attitudes) toward one of those propositions is a warrant for asserting, believing, denying, doubting, assuming (or taking the relevant related attitude) toward the other.4 Is this really enough for a propositional truth predicate to count as deflationary? Some might object that it isn’t, on the grounds that I haven’t explained what true means, which, of course, I haven’t. But why is this an objection? It is no part of deflationism that the truth predicate must be definable. Consequently, it would be expecting too much to demand that deflationists provide a synonym for the truth predicate, or identified the proposition expressed by the sentence the proposition that S is true with the proposition expressed by some other sentence in which the word ‘true’ doesn’t occur. Some deflationists may think that truth can be defined, but not all do, and it is not an essential part of the deflationary position. Others might object that in giving my characterization of deflationism I haven’t said anything informative about truth, or the meaning of the ordinary truth predicate of propositions. But why not? To be told that the claim that p is true is necessarily and a priori equivalent to p, and that any warrant one has for bearing an attitude toward one is warrant for bearing that attitude toward the other is to be told something decidedly nontrivial. After all, many other properties of propositions don’t have these characteristics. A further potential objection to my characterization of deflationism is that I haven’t explained what it is to understand the truth predicate. In particular, it might be objected (i) that I haven’t explained the truth predicate in such a way as to guarantee that someone who did not previously understand it now could, and (ii) that I haven’t given noncircular, necessary, and sufficient conditions for understanding the predicate. And why not, one might ask? Well, consider my remark that p is necessarily and a priori equivalent to the claim that p is true. What does it mean for two propositions to be necessarily and a priori equivalent? It means that each is a necessary and an a priori consequence of the other. And what does that mean? First consider necessary consequence: According to the standard definition, which I accept, a proposition q is a necessary consequence of a proposition p if and only if any possible circumstance in which p is true is one in which q is true. Note, in defining necessary consequence I have used the notion of truth. Thus, my claim that p is necessarily equivalent 4 (ii) should be understood as subject to the following complication: a warrant for believing etc. the proposition that p is true is a warrant for believing etc. p (provided that the agent’s warrant for believing the proposition that p is true involves apprehending not only it, but also p itself).

328 • Essay Twelve

to the proposition that p is true presupposes an independent grasp of truth. Although this claim of necessary equivalence is true, it cannot be part of any noncircular explanation of the knowledge which is necessary and sufficient for understanding the truth predicate, and it wouldn’t be very informative to one who didn’t already know what truth was. If someone didn’t already understand what it was for a proposition to be true, then being told that the proposition that p is true is necessarily equivalent to p might help him eliminate some possible interpretations of the truth predicate, but not all, and not enough to ensure correct understanding of the predicate. Nor, I think, should one be tempted to take necessary consequence to be primitive, in the hope of avoiding this objection. In order for that strategy to have any hope of success, one would have to posit other primitives as well—including a priori consequence, and perhaps other notions. This seems extravagant. Moreover, there is no reason to be so desperate either for a definition of truth, or for an analysis of what it is to understand the truth predicate. As I will argue, there is much to be learned from deflationary characterizations of truth, even if they don’t provide such definitions or analyses. Similar points can be made about a priori equivalence. As I have indicated, two propositions are a priori equivalent iff each is an a priori consequence of the other. To say that q is an a priori consequence of p is, I suppose, to say that q can be deduced from p using a priori reasoning alone. But what is deductive reasoning of this sort, if not reasoning each step of which can be known a priori to be truth preserving? If this is right, then although it is true that p is a priori equivalent to the proposition that p is true, this statement presupposes an independent grasp of the notion of truth. Thus, the claim of a priori equivalence cannot be part of any noncircular specification of knowledge that is necessary and sufficient for understanding the truth predicate. Moreover, if one didn’t already know what truth was, being told of the equivalence could not be expected to ensure that one would arrive at a correct understanding. But what do these objections show? That the conditions I have specified as being necessary and sufficient for a propositional truth predicate to be deflationary are not noncircular necessary and sufficient conditions for understanding the predicate. But this is no objection to the characterization of deflationism, or to the correctness, or informativeness, of the claim that our ordinary truth predicate of propositions is deflationary in this sense. It merely shows that I haven’t tried to provide an analysis of what truth is, or what it is to understand it. There are other potentially significant questions I haven’t attempted to answer either—for example, whether the notion of a proposition (or of the meaning of a sentence) can be understood independently of truth. Answers to these questions might well be crucial if I were attempting to

Understanding Deflationism • 329

provide a full-fledged analysis. But I am not. I am merely trying to state some important truths about truth. It is natural to wonder whether we could do better. For example, could we provide noncircular necessary and sufficient conditions for understanding the propositional truth predicate by switching from statements of the a priori and necessary equivalence of the propositions expressed by S and the proposition that S is true (which presuppose the notion of truth) to talk about propositions expressed by instances of the schema, ‘The proposition that S is true iff S’? Perhaps, but it is not obvious how. Moreover, one approach clearly won’t work. Since no one is familiar with all propositions, no one is familiar with all propositions expressible by instances of the schema. Consequently, understanding the truth predicate is not believing or accepting all those propositions. Nor is there any special subset of the entire collection of (nonparadoxical) propositional instances of the schema such that one understands the notion of truth iff one believes or accepts all the members of that subset. Can we get around this difficulty by maintaining that to understand the notion of truth is to be disposed to accept each such proposition, once it is presented? No we can’t, if being disposed to accept these propositions involves accepting them in nearby counterfactual situations. There are propositions expressible by instances of the T-schema that I couldn’t be presented with, propositions too complex for me to entertain, as well as propositions which, if I did try to entertain them, would be too complex for me to respond to in any coherent way. At this point one is tempted to say something that no one knows for certain how to evaluate—namely, that one does have the relevant dispositions in these problematic cases, even if there are no straightforward counterfactual circumstances in which one would manifest them. Is there a notion of disposition that would legitimate this claim? Perhaps, but it is hard to be sure. In light of this, appeal to such dispositions strikes me as too uncertain a basis on which to rest one’s characterization of deflationism. To make matters worse, there is another problem. Shocking as it may seem, some apparently competent speakers of English, known in the trade as nonfactualists, proudly and self-consciously reject certain instances of the propositional T-schema. In my opinion, those who do this are mistaken. However, I am loath to characterize them as not understanding the word ‘true’, used as a predicate of propositions. In this, I agree with Mark Richard, who gives the following illuminating illustration of what nonfactualism about a realm of discourse amounts to. . . . consider Victoria. She is a film buff, but thinks that Buñuel’s work was in fact an elaborate hoax, with different people the directors of different [parts] of Buñuel’s oeuvre. She will tell you quite baldly that

330 • Essay Twelve

she believes that Buñuel doesn’t exist. Victoria is committed in some sense to the existence of the proposition, that Buñuel doesn’t exist; suppose she recognizes and accepts the commitment. And suppose that she will not only say she believes that Buñuel did not exist; she will say that Buñuel did not exist.5 Richard’s Victoria is a nonfactualist about existence-talk who argues for nonfactualism as follows: (1) Obviously, there is such a proposition as the proposition that Buñuel doesn’t exist. The sentence ‘Buñuel doesn’t exist’ is significant, I and others understand it, etc. (2) It is obvious to me that if there is such a proposition, I ought to believe it. Believing that Buñuel doesn’t exist maximizes the accuracy of my overall picture of the world. And believing it does a much better job of making my behavior fit the way the world is, than not believing it. (3) Assertion is the (canonical) way we convey belief. So since it is often appropriate for me to convey my belief that Buñuel does not exist, I often ought to assert that Buñuel does not exist. (4) A sentence . . . Buñuel . . . makes a claim that is true or false only if there is an object, named ‘Buñuel’, which potentially satisfies . . . x . . . There is no such object. So the claim made by ‘Buñuel does not exist’—the proposition that Buñuel does not exist—isn’t true or false. So I reject the claim that it is true that Buñuel does not exist; to accept this would, in effect, be a recognition that the proposition that Buñuel exists is truth evaluable. But I don’t recognize this.6 What is important about this example is not the realm of discourse selected, but the structure of the nonfactualist position. The nonfactualist advocates belief in, and assertion of, certain propositions, even though he or she denies that these propositions are true, on the grounds that they can’t properly be evaluated for truth at all. This well-known position has been taken about various philosophically puzzling realms of discourse, including morality and mathematics. Thus, in expanding the position of his fictitious heroine, Victoria, Richard maintains that there are, or could be, many examples of perfectly rational people who hold analogous positions. Someone who takes mathematical discourse to be without truth conditions (because of the non-existence of referents for its terms) may still find it significant. He may acknowledge 5 6

Richard (1997, 59). Richard (1997, 60).

Understanding Deflationism • 331

that many people believe, for example, that 2 + 2 = 4. He may find it impossible, and unnecessary, to eschew mathematical talk in everyday life. His frustrated exclamation to a clerk, that twice $4.75 is so $9.50, is arguably as much an assertion and as much a reflection of a motivating belief, as any “normal” one. He may still insist that he is not committed to there really being numbers; and so he is not committed to what he said really being true.7 Such is Richard’s characterization of philosophical nonfactualism. What should a deflationist say about philosophers who take this kind of nonfactualist position? One thing to say is that they are wrong about truth. However, although these nonfactualists may have false semantic or philosophical views about what the truth predicate means, many are normal speakers of English, who satisfy the usual standards of competence. Because of this, they must be counted as understanding the predicate ‘true’, even though they explicitly reject certain nonparadoxical, nonpathological instances of the propositional T-schema. These philosophical speakers are, therefore, prima facie counterexamples to the claim that in order to understand the ordinary truth predicate, one must be disposed to accept all propositions expressed by instances of the schema ‘The proposition that S is true iff S’. Is this, then, a refutation of that claim? Perhaps a proponent of the claim would insist that our nonfactualist speakers really are disposed to accept the very instances of the schema they explicitly reject—they simply allow their false philosophical views to override their robust, linguistically based dispositions to accept the instances. Here again we are in murky waters about what it is to have a disposition without manifesting it. Surely, however, it is conceivable that our nonfactualist speakers might start out having dispositions to accept certain instances of the propositional T-schema that are overridden by their philosophical views, but later lose those specific dispositions altogether, without changing any of their beliefs or linguistic behavior. It would, I think, be highly implausible to suppose that in such a case nonfactualist speakers lose their competence with the truth predicate simply by losing dispositions that previously had been systematically overridden, and rendered ineffective anyway. Thus, it seems that we do have a counterexample to the claim that understanding the notion of propositional truth is a matter of being disposed to accept all nonparadoxical, nonpathological instances of the propositional T-schema. Perhaps some weakening of the claim could be found that would prove workable. But then again, perhaps not. In any case, we are still without obviously correct deflationary criteria for understanding the truth predicate. 7

Richard (1997, 60).

332 • Essay Twelve

But why is this important? In general the criteria for understanding words in a natural language are pretty loose and quite minimal. One lesson of recent antidescriptivism in semantics is that understanding a word is not in general to be identified with accepting some privileged set of claims involving the word. This seems to be so for names and natural kind terms. It may also be true of our ordinary truth predicate of propositions. The idea that even if this predicate is primitive, we ought to be able to give informative, necessary, and sufficient conditions for understanding it may itself be unfounded. So how, in the end, should we characterize deflationism? Here, I return to my earlier suggestion. A theory of our ordinary notion of truth for propositions is deflationist just in case it holds (i) that p and the claim that p is true are necessary and a priori consequences of one another, and (ii) that any warrant for asserting, believing, denying, doubting, assuming (or taking any of a variety of related attitudes) toward one of those propositions is a warrant for asserting, believing, denying, doubting, assuming (or taking the relevant related attitude) toward the other. So characterized, deflationism—thought of as a general approach to truth— neither defines truth, nor states what it is to understand the notion of truth in terms that don’t themselves presuppose truth. Although particular deflationist theories may attempt to do one, or both of these, deflationism per se simply states certain important truths about truth. Note also that these truths are compatible with the scenarios that created problems for the claim that to understand the truth predicate is to be disposed to accept all propositions expressed by instances of the schema ‘The proposition that S is true iff S’. The first of those scenarios involved propositions expressed by instances of the schema that are too complex to entertain coherently. Obviously, such cases are no threat either to the claim that p is equivalent to the proposition that p is true, or to the claim that anything that would constitute warrant for accepting one of these propositions would constitute warrant for accepting the other. The same is true of the second scenario, involving nonfactualist speakers who, in certain special cases, accept p while rejecting the proposition that p is true (having lost any initial disposition to accept it). The point to notice about this scenario is that such speakers have warrant for accepting the proposition that p is true (assuming they have warrant for accepting p), even though they don’t recognize that they do. In this respect, their position is analogous to skepticism about the external world. The perceptual experience of a certain sort of skeptic, gazing at his outstretched hand in normal conditions, provides him with a warrant for believing that he has a hand, even though he wrongly, and unjustifiably, refuses to acknowledge such warrant because he thinks that the mere possibility of deception by an evil demon makes such warrant impossible. Similarly, the nonfactualist who

Understanding Deflationism • 333

has warrant for accepting p also has warrant for accepting the proposition that p is true, even though he wrongly, and unjustifiably, refuses to acknowledge such warrant because of his incorrect semantic and philosophical views about truth.

Assessing Deflationism With our conception of deflationism in place, we are now in a position to assess various deflationist, and inflationist, claims. (Inflationism is the denial of deflationism.) Some of these involve sentential truth predicates. For example, inflationists sometimes hold (i) that our ordinary predicate ‘is a true sentence of L’ is such that ascriptions of it to a sentence S are not equivalent in a requisitely strong sense to the claim made by S itself, and (ii) that deflationary predicates of sentences that are specially introduced so as to give rise to the requisite strong equivalences are not sufficient for all the tasks for which a truth predicate is needed. As I have already indicated, I believe both of these claims to be correct. Nevertheless, I still consider myself to be a deflationist because I take our ordinary truth predicate of sentences to be definable in terms of a sentence’s expressing a true proposition, and I am a deflationist about propositional truth. By contrast, full-blown inflationists either reject propositions altogether, or they accept propositions, but maintain that propositional truth is also nondeflationary. If I am right about the proper way to understand propositional deflationism, there is just one way for those who accept propositions to reject such deflationism; they must reject some equivalences between p and the proposition that p is true that the deflationist accepts. The central theses of deflationism assert certain kinds of equivalences between truth bearers, and the claims that those bearers are true. Propositional inflationism is not a view that accepts these theses, while adding further controversial claims. Rather, propositional inflationists deny what deflationists take to be obvious—including some instances of the propositional T-schema. In this connection, it is worth pointing out that deflationism, as I have characterized it, is compatible with various expansionist views that have traditionally masqueraded under the guise of theories of truth. For example, deflationism is compatible with the view that p is true if and only if p is knowable, as well as with the view that p is true if and only if p is useful for us to believe. It is even compatible with the view that the proposition that p is true is necessarily and a priori equivalent to the proposition that p is knowable, or to the proposition that p is useful for us to believe. What deflationism does require is that if someone holds such a view, then he is committed to the claim that for each proposition

334 • Essay Twelve

p, p itself is equivalent to the proposition that p is knowable, or to the proposition that p is useful for us to believe. Few sensible people will accept such extravagant claims. But, strictly speaking, disputes over such claims have little to do with deflationism. To paraphrase a famous remark of Tarski’s, “we may accept [deflationism about propositional truth] without giving up any epistemological [or metaphysical] attitude we may have had; we may remain naive realists, empiricists, or metaphysicians— whatever we were before. [Deflationism about propositional truth] is completely neutral toward all these issues.”8 Of course, if I am right, the deflationist will maintain that anyone who is warranted in asserting, believing, or assuming p is warranted in asserting, believing, or assuming that p is true, and vice versa. This relationship surely does not hold between the proposition that p is true and the proposition that p is knowable, between the proposition that p is true and the proposition that p useful for us to believe, or between the proposition that p is true and any other claim that traditional expansionists have wanted to link it with. So the deflationist will properly note that the relationship between p and the proposition that p is true is undeniably closer than the relationship between either p or the proposition that p is true, on the one hand, and any other expansionist candidate, on the other. This, in turn, may justify viewing expansionists’ characteristic claims—that p is knowable is equivalent to p, that p is useful to believe is equivalent to p, etc.—more as surprising tenets of unusual theories of reality than as corollaries of a theory of truth. The lesson here is that serious challenges to propositional deflationism come not from those who add controversial doctrines to the equivalences about truth maintained by deflationists, but rather from those who straightforwardly deny what deflationists assert. The real opponents of propositional deflationism are nonfactualists who believe, and are willing to assert, p while rejecting, and being unwilling to assert, that p is true. What do these inflationists have to say for themselves? Presumably not that p fails to entail that p is true. In order for that to be so it would have to be possible for p to be true, even though the claim that p is true was not true. Surely that cannot be, if entailment is defined in the standard way, in terms of truth; and nonfactualists who accept the standard definition do not claim otherwise.9 Conceivably, if nonfactualists were 8 Tarski (1944, sec. 18). In the paraphrased passage Tarski was talking about his semantic theory of truth. However, it has always struck me that the philosophical neutrality he claimed for his semantic theory was a good model of what deflationists should be aiming at. 9 I leave aside possible alternative characterizations of entailment, not involving the notion of truth. In order to help the nonfactualist an alternative characterization would have to be given according to which p does not entail the claim that p is true. I know of no plausible characterizations of this sort.

Understanding Deflationism • 335

skeptics about propositions, they might reject all talk of propositions, including talk of propositional truth. However, such skepticism can be neutralized by taking propositions in a minimal and uncontentious way. When S is a meaningful declarative sentence which may be embedded as the antecedent of a conditional, or the complement of a propositional attitude verb, a speaker who assertively utters S “says something,” and a person who sincerely accepts such a remark believes “what was said.” Nonfactualists interested in discourses including such sentences typically are willing to grant that speakers who assertively utter them do say or assert things which, if they are sincere, they also believe. Since propositions may be taken to be nothing more than what is said and believed in this sense, nonfactualists are well advised to grant the existence of propositions. Having done so, they have little alternative but to admit that the proposition that p is true is a necessary consequence of p. How, if they do admit this, can they accept and believe p while rejecting and disbelieving that p is true? The situation, as I envision it, is this: a reflective and rational speaker who understands S, and believes the proposition it expresses, assertively utters S, thereby asserting that proposition; in addition, he understands another sentence R, and recognizes the proposition expressed by R to be a necessary consequence of the proposition expressed by S; nevertheless, he does not believe the proposition expressed by R, he rejects any assertive utterance of R, and he protests that he is not committed to the proposition it expresses. How can this be? Take entailment to be the converse of necessary consequence. How can one fail to be committed to a recognized entailment of something one is fully committed to? Mark Richard provides an answer to this question on behalf of his imagined nonfactualist, Victoria. In Richard’s imagined scenario, Victoria accepts the premise Buñuel does not exist while denying the conclusion It’s true that Buñuel does not exist. Richard grants that this little argument is valid, in the sense that its conclusion is a necessary consequence of the premise. He even grants that his nonfactualist, Victoria, recognizes this. What he denies is that Victoria’s acceptance of the premise and recognition of the validity of the argument provide her with any reason to accept the conclusion. To think otherwise, he argues, is to beg the question against nonfactualism. Here is what he says: To say that the argument is valid is to say that its premise entails its conclusion. That is, necessarily, if the premise is true, then so is the

336 • Essay Twelve

conclusion. If it’s possible to coherently believe something without believing it to be true, then one can recognize the entailment, believe the premises, but not be committed to the conclusion’s truth, or to the conclusion itself. And Victoria’s position is that we ought to believe the premise of the argument, but not believe that it is true.10 Richard’s idea seems to be that since entailment is defined in terms of truth, a nonfactualist who accepts and believes p, without taking p to be true, may, in good conscience, reject q, while recognizing that p entails q, since this recognition merely ensures that q must be true, if p is true, and this tells us nothing about q in a situation in which p is not true. If this thought is correct, then nonfactualists may freely reject propositions they recognize to be entailed by propositions they continue to accept. However, the thought is not correct. All rational speakers and reasoners, even nonfactualists, are governed by normative constraints arising from recognized logical and conceptual relations involving the sentences and propositions they accept. Suppose Richard’s nonfactualist heroine, Victoria, were to say: “Buñuel doesn’t exist and Ingmar Bergman doesn’t either.” Having asserted the conjunction, she is committed to what it entails, which includes each conjunct. However, if Richard’s thought were correct, this would not be so. Using his all-purpose excuse, Victoria could argue that although she recognizes that the conjunction entails each conjunct, this commits her to accepting the conjuncts only if she recognizes the truth of the conjunction. Since, as a nonfactualist, she doesn’t recognize the truth of the conjunction, she may argue that she is free to take any attitude she likes toward the conjuncts. But, surely, such an argument, on her part, would be a joke. Argumentative commitments cannot be evaded so easily. This should have been evident from the beginning, when considering Richard’s purported explanation of how Victoria could, defensibly, accept the premise p of an argument while rejecting its conclusion q, despite recognizing that the former entails the latter. According to Richard’s explanation, Victoria’s recognition of the entailment doesn’t commit her to accepting q because she doesn’t accept the truth of p, despite accepting p itself. We have seen that this is wrong. But there is another point as well. Richard seems to suggest that if Victoria were to accept the truth of p, then she really would be committed to the truth of q, and (thereby) to q itself. How so? The obvious answer is that in accepting both the truth of p, and the entailment of q by p, Victoria would be accepting a pair of claims (i) that p is true and (ii) that if p is true, then q must be true. But why should accepting these claims commit her to the truth of q? The only 10

Richard (1997, 63).

Understanding Deflationism • 337

answer seems to be that accepting these claims would commit her to the truth of q because these claims entail that q is true.11 But now we have come full circle. Richard’s argument on behalf of his nonfactualist heroine implicitly presupposes the very thing it was intended to deny—namely that one is, after all, committed to obvious entailments of propositions one asserts or accepts. Since the proposition that Buñuel doesn’t exist entails that it is true that Buñuel doesn’t exist, the nonfactualist position that accepts the former while denying the latter is untenable. There is, I think, a larger explanatory point in the offing. We are interested, pretheoretically, in the relation of logical consequence among sentences, and of necessary consequence among propositions, because we regard these notions as connected with the argumentative commitments we take up when we accept a set of premises. The reason why a definition of logical consequence in terms of guaranteed truth-preservation serves this interest is that we take the central deflationist point for granted: namely that acceptance of S (and the proposition it expresses) carries with it a commitment to the truth of S (and the proposition it expresses), and vice versa. If we didn’t take this for granted, then recognition that an argument was guaranteed to be truth preserving would give us no guidance about what acceptance of its premises committed us to. Thus deflationism about truth is central to our practice of using logical and necessary consequence to track our argumentative commitments. And what are these commitments? Consider the simplest case. Mary asserts that John is fat and happy. Since doing this commits her to John’s being happy, for her to deny that John is happy would be for her to incoherently characterize John as being happy while also denying that he is happy. Avoiding this sort of incoherence is a basic argumentative commitment. Using the deflationist insight, we can express essentially the same point using the notion of truth. In asserting the conjunction that John is fat and happy, Mary is committed to its truth, and thereby to the 11 This is the obvious answer when Richard’s reasoning is applied to the general case of an argument in which an agent accepts both the premise of an argument, and the entailment of the conclusion by that premise, but may or may not accept the truth of the premise. However, in the case of the specific argument with premise ‘Buñuel doesn’t exist’ and conclusion ‘It is true that Buñuel doesn’t exist’, things are much simpler. Since in this special case to accept the truth of the premise is identical with accepting the conclusion, no appeal to (ii) above is needed. Nevertheless, the general case is dialectically crucial. The deflationist’s objection to nonfactualism is that since (a) one who accepts p and recognizes that p entails q is committed to q, and (b) p entails that p is true, (c) one who accepts p is committed to the claim that p is true. Since Richard denies (c) while accepting (b), he must reject (a). For this he needs the general argument. What we have seen is that the general argument against (a) fails both because it implicitly presupposes that which it is supposed to refute, and because it fails to recognize such obvious commitments as one’s commitment to the conjuncts of a conjunction one has accepted.

338 • Essay Twelve

truth of both conjuncts. Thus, if she were to deny a conjunct, she would be committed to denying the truth of a proposition she has already implicitly characterized as true. In this scenario, Mary can correctly be described in two ways: (i) as incoherently predicating a property—being fat—of John while denying that predication, and (ii) as incoherently predicating a property—truth—of the proposition that John is fat, while denying that predication. These are two sides of the same coin. Contrary to Richard’s nonfactualist, our argumentative commitments do not arise only when we ascend to talk about truth. Rather, ascent to truth is a way of generalizing and systematizing our understanding of commitments we already have. The utility of the truth predicate in studying the basic forms of argumentative commitment lies in the role it plays in allowing us to abstract away from particular predications and particular argument forms, and to bring them under a small set of general headings: logical consequence, logical inconsistency, and so on. If this is right, then the deflationist insight about truth is part and parcel of our understanding of the relationship between and logic and argument. What about truth-value gaps? As I have characterized propositional deflationism, it holds that whenever one is warranted in denying that p is true, one is warranted in denying p itself. This leaves no room for any nonstandard truth-value according to which, when p has that value, a denial that p is true may be both correct and warranted, whereas a denial of p itself may not be. However, deflationism does leave room for the kinds of gaps that arise from partially defined predicates. When F is a predicate and n is a name of an object for which F is undefined, the proposition p, expressed by n is F, is neither determinately true, nor determinately not true. The same may be said for the negation of that proposition, expressed n is not F. Since, to assert either of these propositions would be to make a mistake, each should be rejected, without accepting its negation. The deflationist insight tells us the same thing about propositions that ascribe truth or untruth to these propositions. They too are neither determinately true, nor determinately not true. They too should be rejected, without accepting their negations. We may put the point this way: if any predicates are partially defined, then a deflationist truth predicate of propositions will also be partially defined. In my view, this partiality plays a role in understanding the liar and sorites paradoxes. This is one way in which a deflationist understanding of the truth predicate constrains what we say about the paradoxes. Of course, there is much more to a proper understanding of these paradoxes than just partiality, or any other deflationist-inspired idea. But that is another story.12

12

Thanks to Gilbert Harman for useful comments on a draft of this essay.

Understanding Deflationism • 339

References Richard, Mark. 1997. “Deflating Truth.” In Truth, ed Enrique Villanueva, 57–78. Atascadero, Calif.: Ridgeview. Soames, Scott. 1999. Understanding Truth. New York: Oxford University Press. Tarski, Alfred. 1944. “The Semantic Conception of Truth and the Foundations of Semantics.” Philosophy and Phenomenological Research 4:341–75.

ESSAY THIRTEEN

Higher-Order Vagueness for Partially Defined Predicates Background In this essay, I will talk about a perplexing problem that arises for the theory of vague and partially defined predicates that I sketched in my book Understanding Truth, and which can, I think, be expected to arise for other theories that employ partially defined predicates.1 The problem is that of making sense of so-called higher-order vagueness. This problem is often regarded as the chief difficulty facing analyses which treat vague predicates as partially defined. Although I can’t claim to have solved the problem, I will argue that it is more tractable than it is often taken to be. I begin by rehearsing the basic framework that I presuppose. The central idea is that vague predicates are context sensitive and partially defined. To say that a predicate P is partially defined is to say that it is governed by linguistic rules that provide sufficient conditions for it to apply to an object, and sufficient conditions for it to fail to apply, but no conditions that are both individually sufficient and disjunctively necessary for it to apply, or fail to apply. Because the conditions are mutually exclusive, but not jointly exhaustive, there will be objects not covered by the rules for which there are no possible grounds for accepting either the claim that P applies to them, or the claim that it does not. P is said to be undefined for these objects. Its extension is the collection of things to which it applies, and its antiextension is the collection of things to which it doesn’t apply. The system is disquotational in that for any name n, we accept the statement ‘P’ applies to n just in case we accept ‘Pn’ is true, which we accept just in case we accept  Pn. When P is undefined for the referent of n, we do not accept Pn,  ‘Pn’ is true, or ‘P’ applies to n, nor do we accept the negations of these claims. We regard it as a mistake to do otherwise, since (i) none of these claims is a necessary consequence of the set of underlying nonlinguistic facts together with the rules of the language governing the expressions they contain, and (ii) given the rules governing the predicates, even one who was omniscient about all nonlinguistic facts would have no grounds for accepting them. 1

Soames (1999).

Higher-Order Vagueness • 341

A distinction is made between the extension of P and its determinateextension, the latter being the set of objects o, such that the claim that P applies to o is a necessary consequence of the rules of the language plus the set of underlying nonlinguistic facts. This distinction results from the fact that there are some objects o such that the claim that o is not in the determinate-extension of P is true, whereas the claim that o is not in the extension of P is to be rejected because the predicate is in the extension of ‘P’ is, like the predicate P, undefined for o. Similar remarks apply to the distinction between the antiextension and the determinate-antiextension of P. Corresponding to these distinctions, there is also a distinction between truth and determinate truth.2 In addition to being undefined, vague predicates are also context sensitive. Given such a predicate P, one begins with a pair of sets. One, the default determinate-extension of P, is the set of things to which the rules of the language, together with the underlying nonlinguistic facts, determine that P applies. The other, the default determinate-antiextension of P, is the set of things to which the rules of the language plus the underlying facts determine that P does not apply. For all objects o, P is undefined for o just in case o is in neither of these sets. Since these sets don’t exhaust all cases, speakers have the discretion of adjusting the extension and antiextension so as to include initially undefined cases. Often they do this by explicitly predicating P of an object o, or by explicitly denying such a predication. When a speaker does this, and other conversational participants go along, the extension (or antiextension) of the predicate in the context is adjusted so as to include o, plus all objects that bear a certain relation of similarity to o. We can illustrate these points with the help of an example. The model is clearest and most intuitive with simple observation predicates like ‘blue’. A characteristic feature of these predicates is that we learn them not by being given verbal definitions, but by being given clear and obvious examples of things to which they apply and things to which they don’t. We are told when presented with some reasonable range of objects ‘This is blue’ and ‘That is not’—or, if we are not explicitly told, we note that there are certain objects that everyone we encounter seems ready to call ‘blue’ and other objects that everyone we encounter seems ready to characterize as ‘not blue’. These learning experiences give rise to beliefs about conditions for proper application of the predicate. Think of it this way: People say of a certain object that it “is blue.” We observe the object, which is perceptually represented to us as being a certain shade of color. Call this shade B1. They say of a different object that it is not blue. We observe

2

See Soames (1999, chap. 6).

342 • Essay Thirteen

that object, which we perceive to be of a different shade—call it NB1. On the basis of experiences like these, we form the belief (which virtually everyone we encounter seems to share) that objects of the first shade— B1—are objects to which ‘blue’ applies, and objects of the second shade— NB1—are objects to which the predicate does not apply. We may idealize this situation by saying that we first entertain, and then come to accept, the hypothesis that the following pair of rules governs the application of the predicate ‘blue’ in the language. Blue 1 a. If an object is B1, then ‘blue’ applies to the object. b. If an object is NB1, then ‘blue’ does not apply to the object. In saying that the agent first entertains, and then comes to accept, the hypothesis that these rules govern the predicate in the language of his community, I don’t mean that the agent formulates these rules in words. Most likely the agent has no words, at least no nonindexical words, that stand for these specific shades. Rather, he comes to accept the propositions expressed by (a) and (b). Other learning experiences with ‘blue’ lead the agent to accept other pairs of rules, involving different shades, as governing the predicate, as well. At some point in this process, the agent is counted as having successfully learned the meaning of the word, as it is used in his linguistic community. At this point, the agent will have accepted a set of rules Blue1–Blue n, the (a) versions of which provide sufficient conditions for ‘blue’ to apply to an object and the (b) versions of which provide sufficient conditions for ‘blue’ not to apply to an object. However, although these conditions will be mutually exclusive, the requirement that they be mutually agreed upon and generally adhered to by the overwhelming majority of speakers no matter what the context, will ensure that they are not jointly exhaustive. Since there will be shades of color, and objects having those shades, about which the rules say nothing, the rules do not provide a set of conditions which are individually sufficient and disjunctively necessary for ‘blue’ to apply to an object, or for it not to apply. This illustrates the partiality of the predicate in the language of the speaker’s community. Context sensitivity results from the fact that speakers have the discretion to apply the predicate ‘blue’, or its negation ‘not blue’, to objects for which it is undefined by the rules of the language. Often they do this by asserting that some contextually salient object “is blue,” or that it isn’t. Consider the positive case. If the other conversational participants accept the characterization of the object o as “blue,” then the extension of the predicate in the context is adjusted to include o and all objects that bear a certain relation of similarity to it. In general, the relation involved in

Higher-Order Vagueness • 343

these contextual adjustments is determined by the meaning of the predicate together with the intentions of speakers and hearers in the context. Putting aside various complications, let us suppose that when an agent characterizes “as blue” an object o for which the predicate is initially undefined, he adopts a contextual standard that counts o, all objects uncontroversially regarded to be bluer than o, as well as all objects that are pairwise indescriminable from o by ordinary observation in good conditions, as being “blue” as well. Let Bc be a particular shade that applies to precisely this class of objects. We may then characterize what has happened in the context as a result of the speaker’s predicating ‘blue’ of o: as a result of doing this, the speaker has adopted a rule governing ‘blue’ in the context that contains the following condition for positive application of the predicate. If an object is Bc, then ‘blue’ applies to it. Although not a rule of the language governing the predicate, this rule is one that speakers are free to adopt at their discretion in particular contexts of utterance. We have now illustrated both parts—partiality and context sensitivity— of the semantic analysis of vague predicates that I will presuppose in what follows. In my opinion these two features of the analysis naturally go together, and are mutually reinforcing. Given an analysis that posits one, we can find substantial reasons for adopting the other as well.3

Consequences for the Sorites Paradox This brings me to the sorites paradox. Since semantic theories of vagueness are often judged by the solutions they provide to the paradox, I will say a few words about this. Although all sorites predicates are vague, not all vague predicates are natural sorites predicates, with application conditions based on the position of objects in a more or less single and unified underlying continuum. Since the semantic analysis of vagueness is intended to apply to all vague predicates, it should be motivated to a substantial degree by considerations independent of the sorites. Any light it sheds on the paradox is an extra benefit. In the case of the analysis I advocate, there are two general consequences that the model has for the paradox. First, the fact that vague predicates are partially defined means that the semantic categorization imposed on the world by such a predicate will include more than two categories. There may well be sharp and precise lines dividing the objects in different categories, but typically 3

This is argued in the final section of Soames (2002).

344 • Essay Thirteen

these lines are not properly characterized as separating objects to which the predicate applies from those to which it does not apply. Second, context sensitivity tells us that the lines are movable. When one looks closely at the mechanisms by which these lines are adjusted in conversational contexts, one finds that in many cases the mechanism makes it practically impossible to display them; any attempt to display the precise line dividing objects to which the predicate applies (or doesn’t apply) from objects for which it is undefined has the effect of moving the line elsewhere. This constant and elusive movement creates the illusion that there are no sharp lines to be drawn. For example, let us assume that the predicate ‘blue’ is partially defined, with a default determinate-extension and a default determinateantiextension, plus a range of objects for which, absent temporary conversational adjustments, it is undefined. Suppose further that the conventions governing the predicate include constraints on how its extension and antiextension may be adjusted within this range. In particular, it is accepted that, typically, one who explicitly characterizes something x as “blue” on the basis of ordinary perceptual evidence is, all other things considered equal, committed to a contextual standard that counts all objects that look bluer than x, plus objects perceptually indistinguishable in color from x (when paired with x and viewed together) as “blue.” Finally, suppose that two stimuli can be perceptually indistinguishable in this sense even though they differ slightly in the physical characteristics that cause them to look “blue.” Given this supposition, we can construct a sequence connecting x1—which definitely looks and is blue—to xn—which definitely is not, and does not look, blue—in which any two adjacent items in the sequence are (pairwise) perceptually indistinguishable in color. When an agent characterizes an object xi—for which the predicate is initially undefined—as “blue,” the (determinate) extension of the predicate is adjusted to include xi, all earlier items in the sequence, plus xi+1— which is perceptually indistinguishable in color from xi. As a result of this adjustment there is now a sharp line between xi+1 and xi+2 separating items to which the predicate determinately applies (in the context) from items for which it remains undefined. However, if one attempts to display this line, by showing the agent xi+1 and xi+2 together, and asking him to characterize them, he will, quite properly, resist the invitation to treat them differently. For if he now explicitly endorses his previously implicit commitment to counting xi+1 as “blue,” then his assertion that xi+1 is “blue” will have the immediate effect of adjusting the contextual standards so as to count xi+2 as “blue” as well. By focusing on and making judgments about what had been the line separating objects to which the predicate (determinately) applied from those for which it was undefined, the agent has

Higher-Order Vagueness • 345

imperceptibly moved the line, thereby engendering the illusion that there was no sharp and precise line in the first place. In my opinion, this analysis has illuminating implications for different versions of the sorites paradox. For example, in dynamic versions of the paradox an agent presented with a sorites sequence about which he is asked to make judgments can easily be pressured into making a series of positive claims x1 is F, x2 is F, . . . , xi is F that comes to an end when he refuses to go further, and either assents to a negation xk isn’t F or refuses to make any judgment at all. At this point, pressure can be generated in the opposite direction, with the result that the agent will dissent from, or withhold judgment on, sentences xj is F to which he previously assented. The semantic model of vague predicates just sketched indicates how and why this pressure is generated, and explains why such an agent need not be viewed as contradicting himself or going back on something he originally asserted. He need not be seen as having done these things because the different judgments he makes change the extension and antiextension of the predicate in such a way that the proposition expressed by  xj is F when he assents to it differs from, and is compatible with, the proposition it expresses when he dissents from or withholds judgment about it. The semantic analysis also points to a useful lesson: although there is something about the meanings of many vague predicates that resists drawing stable boundary lines for applying them, the semantic rules governing such predicates are coherent as they stand, and there is no compelling practical or theoretical need for stable boundaries.4 In addition, the analysis provides the basis of rejecting the major premise MP of a generalized version of the sorites paradox, while also explaining the deceptive plausibility it enjoys by virtue of its association with the more plausible and defensible premise MP* that arises directly from the rules for adjusting the extension of the vague predicate—in this case ‘blue’. (The background for the paradoxical argument includes the claim that there is a sequence S starting with something B that definitely looks and is blue, and ending with something NB that definitely is not and does not look blue. Moreover, for all members si and si+1 of S, si is perceptually indistinguishable in color from xi+1 to competent observers in good light under normal conditions.) MP. For any two colored items x and y that are perceptually indistinguishable in color to competent observers in good light under normal conditions, x and y look to be and are of the same color. Hence for each si, si+1 is blue, if si is blue. 4

See Soames (1999, chap. 7).

346 • Essay Thirteen

MP* For any two colored items x and y that are perceptually indistinguishable in color to competent observers in good light under normal conditions, a person who characterizes ‘blue’ as applying to x, (in such circumstances) is, all other things being equal, committed to a standard that counts ‘blue’ as applying to y as well. Hence for each si, si+1 is counted as “blue,” if si is explicitly characterized as “blue.” By allowing us to distinguish the roughly correct MP* from the incorrect sorites premise MP, the semantic model that I here presuppose is capable of dispelling important and widespread confusions about standard versions of the sorites paradox.5 Nevertheless, I don’t regard the semantic model as providing a complete solution to the sorites. As we will see, the model remains vulnerable to certain strengthened, revenge versions of the paradox, when we take higherorder vague predicates into account. It will be evident from the way these versions arise that there is a limit to how far one can go in defusing them by appealing to my semantic model. Even if the model is more or less correct, as I believe it to be, and even if it tells us important things about the sorites, as I believe it to do, there remains a fundamental mystery brought out by the sorites that the model does not resolve or illuminate. But that is getting ahead of ourselves. Higher-order vagueness can seem to be a perplexing problem for my semantic account from the very beginning. The central problem arises from treating vague predicates as partially defined. There is a natural line of reasoning arising from this characterization that makes it difficult to see how there could be any higher-order vagueness in the first place. What is the problem?

The Prima Facie Problem of Higher-Order Vagueness Let P be a vague predicate that is undefined for objects that are in neither its default determinate-extension nor the default determinate-antiextension. Let  is determinately P apply to an object o just in case o is in the determinateextension of P.6 This predicate applies neither to any object for which P is undefined nor to any object in its determinate-antiextension. Is it partially 5 See Soames (1999, chap. 7). For objections and a reply see Williamson (2002) and Soames (2002). 6 The just in case connective is used to form biconditionals that always have truthvalues when its arguments are undefined, or are otherwise such that we must reject assignments of truth-values to them. (This is not a definition.) See Soames (1999, chap. 6) for further discussion.

Higher-Order Vagueness • 347

defined? There is reason to think that it can’t be. In giving the analysis of P, we specified three and only three relevant categories of objects—those to which P determinately applies, those to which it determinately fails to apply, and those for which it is undefined. If these categories are jointly exhaustive, then is determinately P is totally defined, and so cannot be either vague or partial. That sounds like a problem. The reason it is a problem is not, in my opinion, that there couldn’t be vague predicates for which the relevant higher-order predicates were totally defined. It seems to me that we could, if we wanted, introduce an artificial predicate P which was both context sensitive and partially defined, for which the higher-order predicate is determinately P was totally defined. On my view, P would then be vague, even though it would not give rise to higher-order vagueness. We could introduce P with this result, provided (i) that it was fully determinate what the rules governing our new predicate P were, and (ii) that these rules did not contain other vague or partially defined concepts, and so were not themselves vague or partial. Thus, higher-order vagueness is not a sine qua non for vagueness. However, the cases in which higher-order vagueness doesn’t arise are special, and different from what we find with ordinary predicates like ‘bald’, ‘blue’, ‘poor’, and ‘young’. The problem is that higher-order predicates—is determinately P— corresponding to ordinary vague predicates like ‘bald’ and ‘blue’ also appear to be vague. Not only is there no sharp and precise line dividing the objects to which ‘blue’ or ‘bald’ apply from the objects to which they don’t, there also seems to be no sharp and precise line dividing (i) the objects to which it is determined, by the rules of the language and the underlying nonlinguistic facts, that these predicates apply from (ii) the objects for which this is not determined.7 Thus, it would seem that the predicates ‘is determinately bald’ and ‘is determinately blue’ are themselves partial. If they are also context sensitive (which I will here assume), then they too should count as vague. This means that analyses of ordinary predicates like ‘bald’ and ‘blue’ that treat them as partial and vague must explain how and why the higher-order predicates corresponding to them are also partial, and vague. How might this be done? 7 Here and in what follows, I will presuppose the default settings of ordinary vague predicates when talking about objects to which they apply, don’t apply, or are undefined— unless special contextual standards are explicitly indicated. Thus, when P is such a predicate, is determinately P will standardly be taken to apply to o just in case o is in the default determinate-extension of P. Of course, when the determinate-extension of P is contextually adjusted, the extension of is determinately P is also adjusted. But such cases will concern us only when explicitly indicated.

348 • Essay Thirteen

Proposed Explanation The Idea Think again about ‘bald’. There are some individuals to which it determinately applies, others to which it determinately does not apply, and still others for which it is indeterminate whether or not it applies, and so is undefined. Although these three categories are mutually exclusive, we should not assume that they are jointly exhaustive; after all, there may be individuals o such that we can find no possible basis for asserting that the predicate ‘x is determinately bald’ applies to o, that it doesn’t apply to o, or even that it must either apply or not apply to o. The reason for this is that there may be no possible basis to assert either (i) that the claim that ‘bald’ applies to o is a necessary consequence of all nonlinguistic facts about o plus the rules of the language governing ‘bald’, or (ii) that the claim that ‘bald’ applies to o, is not a necessary consequence of all nonlinguistic facts about o, plus the rules of the language governing the predicate ‘bald’, or (iii) that one of these claims about necessary consequence must be true. How could this be? We may think of the rules governing ‘bald’ as being of the sort indicated by the pair, Bpos and Bneg, where ‘so and so’ and ‘such and such’ in the antecedents of the two conditionals are mutually exclusive, but not jointly exhaustive.8 Rules Governing Bald Bpos For all o, if o is so and so, then ‘bald’ applies to o (and so o is bald). Bneg For all o, if o is such and such then ‘bald’ doesn’t apply to o (and so o isn’t bald). The rules governing the predicate ‘determinately bald’ are the rules governing ‘bald’ plus the rules Dpos and Dneg governing ‘determinately’.9

8

All instances of these rules are assumed to have definite truth-values. In giving the rules for ‘determinately’, I have simplified matters to focus on the most dramatic and important case—uses of a vague predicate in contexts in which it carries its default determinate-extension and antiextension. In such a context, for an object o to be “determinately bald” is for the claim that ‘bald’ applies to o to be a necessary consequence of the rules of the language governing ‘bald’ plus the underlying facts. In a context in which speakers have already exercised their discretion by adjusting the extension of ‘bald’, for o to be “determinately bald” is for the claim that ‘bald’ applies to o to be a necessary consequence of the rules already in force in the context plus the underlying facts. 9

Higher-Order Vagueness • 349

Rules Governing Determinately Dpos For all o, if o is such that the claim expressed by ‘P’ applies to x relative to an assignment of o to ‘x’ is a necessary consequence of the set of all nonlinguistic facts about o plus the rules of the language governing P, then determinately P applies to o (and the claim expressed by x is determinately P relative to an assignment of o to ‘x’ is true). Dneg For all o, if o is such that the claim expressed by ‘P’ applies to x relative to an assignment of o to ‘x’ is not a necessary consequence of the set of all nonlinguistic facts about o plus the rules of the language governing P, then determinately P does not apply to o (and the claim expressed by x is not determinately P relative to an assignment of o to ‘x’ is true). These rules are sensitive to three things—(i) the set of all nonlinguistic facts about o, (ii) the rules of the language governing P, and (iii) the relation of necessary consequence. I will take (i) and (ii) to be sets of propositions, and (iii) to be a relation holding between sets of propositions and individual propositions, which, when applied to propositions that are precise and nonvague, is itself precise and well defined. In order to simplify the discussion, I will further assume that the propositions in (i) are all fully defined, precise, and true—no vagueness allowed here. However, no such assumption will be made in the case of (ii). If there is any vagueness about what the rules of the language are, or if there is any vagueness in something which definitely is a rule of the language, then this may affect the results achievable by applying Dpos and Dneg. With this in mind, we return to our question, How can the conditions in the antecedents of Dpos and Dneg be seen as anything other than jointly exhaustive when P is the predicate ‘bald’?. The answer is that whether or not we can establish or correctly accept the claim that these conditions are jointly exhaustive depends on whether or not we can establish or correctly accept the claim that for each potential rule R of the form , it is determinate whether or not R is a rule of the language governing ‘bald’. The crucial point is that we cannot do this. Graph G1 of the baldness continuum illustrates this point. G1 | bald

| ?

| undefined | ? | bald

| not bald

|xxxxxxxxxx|xxxxxxxxxx|xxxxxxxxxx|xxxxxxxxxx|xxxxxxxx Region 1 Region 2 Region 3 Region 4 Region 5 Region 1 consists of individuals who would be judged to be clearly bald by virtually every competent speaker, provided the speaker were fully

350 • Essay Thirteen

apprised of the relevant facts about them, for example, by observing them in normal conditions. There is no serious question about these individuals; they are bald. Similarly, there is no serious question about those rule candidates Bpos that classify only members of region 1 as individuals to which ‘bald’ applies; such candidates are included in the rules of the language governing ‘bald’. Region 2 consists of individuals about whom there is moderate uncertainty or disagreement. Most competent speakers would judge these individuals to be “bald,” and few if any would confidently characterize them as “not bald,” but a significant number would be uncertain whether they qualify as “bald,” and would be somewhat reluctant to pronounce judgment on them. This region of individuals gives rise to undefinedness in the predicate ‘is a rule of the language governing the predicate ‘bald’’. Rulecandidates Bpos that classify all members of region 1, some members of region 2, and no members of any other region, as individuals to which ‘bald’ applies are rules for which the predicate ‘is a rule of the language governing ‘bald’’ is undefined. Region 3 contains paradigmatically borderline cases of baldness. There is great uncertainty and variation among speakers, and across time, regarding whether they classify individuals in this region as “bald” or “not bald”; and often they may be reluctant or unwilling to classify these individuals as either. Rule candidates Bpos that classify ‘bald’ as applying to some individuals in region 3, as well as rule candidates Bneg that classify ‘bald’ as not applying to these individuals, are not rules of the language governing ‘bald’. They may be rules that speakers have the discretion to adopt in particular conversational circumstances; however, they are not rules that are constitutive of the language itself. Regions 4–5 are mirror images of regions 1–2, with ‘not bald’ replacing ‘bald’ and Bneg replacing Bpos. On this way of looking at things, the rules of the language governing ‘bald’ include many different pairs of positive and negative conditionals, even though one pair may subsume many others—i.e., cover every case that the others do, and more. So what are these rules? The rules governing ‘bald’ include pairs of rules in which the antecedent of Bpos applies only to individuals in region 1 and the antecedent of Bneg applies only to individuals in region 5. The rules of the language governing ‘bald’ do not include any pair in which either the antecedent of Bpos applies to individuals outside regions 1 and 2, or the antecedent of Bneg applies to individuals outside of regions 4 and 5 (though speakers may choose to adopt these rules in particular contexts). Any pair in which either (i) the antecedent of Bpos applies to individuals in region 2 (but none in regions 3–5), while the antecedent of Bneg applies only to individuals in regions 4 or 5, or (ii) the antecedent of Bneg applies individuals in region 4 (but none in regions 1–3), while the antecedent of Bpos applies only to individuals in

Higher-Order Vagueness • 351

regions 1 or 2, is such that we can draw no conclusion regarding whether or not it is a rule of the language governing ‘bald’. Speakers can decide to be guided by these rules in particular conversations, but, if they do, there will be individuals o2 in region 2, or o4 in region 4, such that we can establish no correct answer to the question ‘Are speakers’ classifications of o2 as bald, or o4 as not bald, correct because they are consequences of the facts about these individuals plus the rules of the language governing ‘bald’, or are they correct because in making these classifications speakers have exercised their option of adopting extensions of the rules of the language?’ If this is right, then there is reason to resist the claim that ‘is determinately bald’ is a totally defined predicate. The basis for the resistance is that for some rules there is simply no saying whether or not they are rules of the language governing ‘bald’. Let R be the class of such rules. For certain objects o—namely those in region 2 of G1—the question of whether the claim that ‘bald’ applies to o is, or is not, a necessary consequence of the rules of the language governing the predicate can be answered only by assuming that certain members of R are rules of the language, or by assuming that they aren’t. Since neither of these assumptions can be established, there is no possible justification for accepting them; thus, we should reject both the claim that these objects are determinately bald and the claim that these objects are not determinately bald, just as we rejected both the claim that they are bald and the claim that they are not bald. So, we reject the claim that ‘determinately bald’ is totally defined. The reason for this is that the rules governing the predicate determinately P make use of the predicate is a rule of the language governing ‘P’ which cannot be seen as total, when P is an ordinary vague predicate like ‘bald’. In saying this, I recognize that the picture I have sketched is incomplete, and that there are unfinished tasks that need to be pursued. Certainly, one would like more informative descriptions of different regions in the graph, including (noncircular?) explanations of crucial concepts—like that of being a competent speaker—employed in giving those descriptions. There is also the issue of locating vagueness in the descriptions of these regions, and exploring the sources and consequences of such vagueness. Despite these unresolved matters, I am not convinced that there is any irresolvable mystery here. As far as I can tell, the predicates ‘bald’, ‘determinately bald’, and ‘is a rule of the language governing ‘bald’’ do fit the broad-brush picture I have sketched. Let us try to fill out that picture a little further. Iterating ‘Determinately’ The points we have made so far are visually represented by the graphs G1 for ‘bald’, G2 for ‘determinately bald’, and G3 for ‘determinately not

352 • Essay Thirteen

bald’. (The question marks indicate that it is so far an open question how individuals in the region should be characterized.) G1 | bald

| ?

| undefined | ? | bald

| not bald

|xxxxxxxxxx|xxxxxxxxxx|xxxxxxxxxx|xxxxxxxxxx|xxxxxxxx Region 1 Region 2 Region 3 Region 4 Region 5 G2

| det bald |

| ?/undefined | det bald

not det bald

|xxxxxxxxxx|xxxxxxxxxx|xxxxxxxxxx xxxxxxxxxx xxxxxxxx Region 1 Region 2 Regions 3, 4, and 5 G3 | |

not det not bald

?/undefined | det not bald det not bald|

|xxxxxxxxxx xxxxxxxxxx xxxxxxxxx|xxxxxxxxxx|xxxxxxxxx Regions 1, 2 and 3 Region 4 Region 5 We have rejected the claim that ‘determinately bald’ is totally defined. Should we accept the claim that it is partial (and presumably vague as well)? If so, do partiality and vagueness go even higher? Consider again the graphs and the question marks they contain. We know that the question marks in regions 2 and 4 of G1 do not indicate that ‘bald’ is undefined for individuals in the regions. But, for all we have said up to now, the question marks in region 2 of G2 and region 4 of G3 might represent individuals for which the predicates ‘is determinately bald’ and ‘is determinately not bald’ are, respectively, undefined. Suppose this is so. We can then use (i) and (ii) to establish that the predicate ‘is determinately determinately bald’ is totally defined. (i) Just as the individuals of whom it can properly be said that they are determinately bald are the same as the individuals of whom it can properly be said that they are bald, so the individuals of whom it can properly be said that they are determinately determinately bald are the same as the individuals of whom it can properly be said that they are determinately bald. Thus, the initial section of the graph for ‘determinately determinately bald’ is the same as the initial section of the graph for ‘determinately bald’. (ii) Just as the individuals of whom it can properly be said that they are not determinately bald include all and only those of whom it can properly be said either that they are not bald or that it is undefined whether or not they are bald, so, the individuals of whom it can properly be said that they are not determinately determinately bald include all and only those of

Higher-Order Vagueness • 353

whom it can properly be said either that they are not determinately bald, or that it is undefined whether or not they are determinately bald. So, if we accept the claim that the predicate ‘determinately bald’ is undefined for every individual in region 2 of G2, and hence that every individual in that region is one of which it can properly be said that it is undefined whether or not that individual is determinately bald, then we get the graph G4 for ‘determinately determinately bald’, which is the graph of a totally defined predicate. G4

| det det bald | |

not det det bald

|xxxxxxxxxxx|xxxxxxxxx xxxxxxxxxx xxxxxxxxxx xxxxxxxx Region 1 Regions 2, 3, 4, and 5 That is a surprising result. How is that when we start with ‘bald’ and add ‘determinately’ we get a predicate which cannot correctly be characterized as total, whereas when we start with that predicate, and iterate ‘determinately’, we do get a totally defined predicate? The answer is that we have made a mistake. The crucial assumption, used in (ii), is that every individual o is either determinately bald, not determinately bald, or such that the predicate ‘determinately bald’ is undefined for o. How do we know that? If at an earlier stage—in moving from ‘bald’ to ‘determinately bald’—we had started with the assumption that every individual o is either bald, not bald, or such that ‘bald’ is undefined for o, we would have reached the conclusion that ‘determinately bald’ was totally defined—which we certainly did not. But if we didn’t make that assumption in the previous case, in moving from G1 to G2, why should we make the corresponding assumption in this case, in moving from G2 to G4? The issue concerns the regions in the graphs labeled with question marks. All we know so far is that when, in G1, o is an individual in one of these regions, we reject the claim that o is bald, we reject the claim that o is not bald, and we reject the claim that the predicate ‘bald’ is undefined for o—all for the same reason, we see that it is impossible in principle to justify these claims. This being so, we need to clarify the status of the regions in the other graphs presently marked ‘? / undefined’. The individuals in region 2 are undefined for ‘determinately bald’ just in case those individuals are not determinately determinately bald, which will be so just in case ‘determinately determinately bald’ is a totally defined predicate. How do we evaluate the claim that it is such a predicate? The first thing to notice is that the rules governing ‘determinately determinately bald’ are the same as the rules governing ‘determinately bald’; they are the rules governing ‘bald’ plus the rules Dpos and Dneg governing

354 • Essay Thirteen

‘determinately’, given earlier. In the present case, we apply the rules twice, once letting P be the predicate ‘bald’, and once letting P be the predicate ‘determinately bald’. The reasoning is given in (i) and (ii), and the results are summarized in (iii). (i) Suppose we are given that the claim that ‘bald’ applies to o is a necessary consequence of the rules governing ‘bald’ plus the underlying nonlinguistic facts about o. Then, using Dpos we derive that ‘determinately bald’ applies to o (and hence that o is determinately bald). Since the rules governing ‘determinately bald’— namely, the rules governing ‘determinately’ plus the rules governing ‘bald’—include the rules used in the foregoing derivation, this means that the claim that ‘determinately bald’ applies to o is a necessary consequence of the rules governing ‘determinately bald’ plus the underlying nonlinguistic facts about o. But then, using Dpos again, we get the result that ‘determinately determinately bald’ applies to o, and hence that o is determinately determinately bald. (ii) Suppose we are given that the claim that ‘bald’ applies to o is not a necessary consequence of the rules governing ‘bald’ plus the underlying nonlinguistic facts about o. Then, using Dneg we derive that ‘determinately bald’ does not apply to o (and hence that o is not determinately bald). Since the rules governing ‘determinately bald’ include those used in the foregoing derivation, this means that the claim that ‘determinately bald’ does not apply to o is a necessary consequence of the rules governing ‘determinately bald’ plus the underlying nonlinguistic facts about o. But then, given the consistency of these rules (with the underlying nonlinguistic facts), we conclude that the negation of that claim—namely, the claim that ‘determinately bald’ applies to o is not a consequence of the rules governing ‘determinately bald’ plus the underlying nonlinguistic facts about o. Finally using Dneg again, we get the result that ‘determinately determinately bald’ does not apply to o, and hence that o is not determinately determinately bald. (iii) When we are not given either (i) that the claim that ‘bald’ applies to o is a necessary consequence of the rules governing ‘bald’ plus the underlying nonlinguistic facts about o, or (ii) that the claim that ‘bald’ applies to o is not a necessary consequence of the rules governing ‘bald’ plus the underlying nonlinguistic facts about o, we cannot use the rules governing ‘determinately’ to get any result. We conclude that the rules of the language together with the underlying nonlinguistic facts give us the same results for ‘determinately bald’ and ‘determinately determinately

Higher-Order Vagueness • 355

bald’. Since it is impossible to justify the claim that either predicate is totally defined, we reject this claim, and for any individual o, we accept the claim that o is (is not) determinately bald just in case we accept the claim that o is (is not) determinately determinately bald. The iteration of ‘determinately’ does nothing. What Not to Say We have rejected the claim that ‘determinately bald’ and ‘determinately determinately bald’ are totally defined predicates. Are they partially defined? There is reason not to say this. I have said that partially defined predicates are those that are undefined for some objects, and that totally defined predicates are those that are not undefined for any object, where by ‘undefined’ I have meant the following: Undefinedness P is undefined for o just in case the rules of the language governing P together with the underlying nonlinguistic facts about o do not determine either that P applies to o or that P does not apply to o— which in turn holds just in case neither the claim that P applies to o nor the claim that P does not apply to o is a necessary consequence of the rules governing P together with the nonlinguistic facts about o (i.e., just in case neither x is determinately P nor x is determinately not P expresses a truth relative to an assignment of o to ‘x’). Given this, one cannot correctly say that ‘determinately bald’ is undefined for o. For if one does say this, one must then admit that o is not determinately determinately bald. But that conflicts with what we have just found—namely that just as we must reject, as unjustifiable, the claim that o is determinately bald, without accepting its negation, so we must reject the claim that o is determinately determinately bald, without accepting its negation. So is ‘determinately bald’ undefined for o or not? Since neither claim can be justified, we have no option but to reject both. A similar result holds for (i) the claim that ‘determinately bald’ and ‘determinately determinately bald’ are partially defined predicates, and (ii) the claim they are not. Since our characterization of what it is to be a partially defined predicate requires the predicate to be undefined for some objects, we must reject the claim that these are partially defined, while continuing to reject the claim that they are totally defined. What can we positively assert about these predicates, and about the regions on the graphs for them that are labeled with question marks? As for the predicates, though they cannot correctly be characterized as partial in the original sense, they can be characterized as partial in a weaker and extended sense.

356 • Essay Thirteen

What We Can Say: Weak Partiality A predicate P is weakly partial just in case there are some objects o such that, no matter how much information one is given about the rules of the language and the underlying nonlinguistic facts, one cannot correctly accept either the claim that P applies to o or the claim that P does not apply to o (or the claim that either P applies to o or it doesn’t). Ordinary, partially defined predicates like ‘bald’ are weakly partial, as are the corresponding higher-order predicates formed by attaching one or more occurrences of ‘determinately’ to them. The difference between partiality and weak partiality can be illuminated by considering the contrast between regions 2 and 3 on the graph G1 for ‘bald’. We consider a pair of claims—the claim that ‘bald’ applies to o2, and the claim that ‘bald’ applies to o3—where o2 and o3 are individuals in regions 2 and 3, respectively. Neither claim can be accepted because neither can be justified. But the reasons for the lack of justification are different in the two cases. In both cases, in order to justify the claim that the predicate applies to the object, one has to establish the premise that there is a rule of the language governing ‘bald’ which characterizes the predicate as applying to the object. In the case of o3 we can refute this needed premise. In the case of o2 we can neither refute it nor establish it. What the cases have in common is that since the needed premise can’t be established, one in possession of all the facts cannot be justified in accepting the claim that the predicate applies to the object, even though in neither case can one be justified in accepting the negation of that claim either. Genuinely partial predicates always include cases like o3; predicates which are only weakly partial include cases like o2, but none like o3. As for the regions on the graphs labeled with question marks, let us take region 2 of the graph G1 for ‘bald’ as a representative example. Let o be an individual in this region. We can’t correctly say that ‘bald’ is undefined for o because there are pairs, which are candidates for being rules of the language governing ‘bald’ according to which ‘bald’ does apply to o—where candidates are rules which we cannot show not to govern the predicate in the language. Since we can’t show this, we cannot correctly say that ‘bald’ is undefined for o. Of course, we also cannot correctly say that ‘bald’ applies to o, because there is no pair of rules , which characterize ‘bald’ as applying to o that we can show to be rules of the language that do govern the predicate. It is helpful in summarizing this situation to introduce the notion of a predicate P being undefined for an object o relative to a rule R. Relative Undefinedness P is undefined for o relative to a rule R: iff neither the claim that P applies to o nor the claim that P doesn’t apply to o is a necessary

Higher-Order Vagueness • 357

consequence of R plus the set of underlying nonlinguistic facts about o. P is defined for o relative to R just in case P is not undefined for o relative to R. Absolute undefinedness is defined in terms of relative undefinedness. Absolute Undefinedness P is undefined for o iff (i) for all rules R which are such that we can, in principle, establish that R is a rule of the language governing P, P is undefined for o relative to R, and (ii) there is no rule R which is a candidate for being a rule of the language governing P, relative to which P is defined for o. (A candidate is a rule which we cannot, in principle, show not to be a rule of the language governing the predicate). In the presence of natural background assumptions—e.g., the assumption that if two rules are such that they should both be accepted as rules of the language, then they don’t give conflicting characterizations of whether a predicate applies to any object—this definition gives the same results as the characterization of undefinedness given earlier. With this in mind, we can characterize each individual o in region 2 of the graphs as follows: Individuals o in Region 2 (i) Every rule R which is such that we can establish that R is a rule of the language that governs the predicate ‘bald’ is such that ‘bald’ is undefined for o relative to R. (ii) Nevertheless, there remain candidates for being a rule governing ‘bald’ which characterize ‘bald’ as applying to o. (iii) For these reasons, we cannot establish, or correctly accept, any of the following claims: that ‘bald’ applies to o, that ‘bald’ is undefined for o, that ‘determinately bald applies to o’, that ‘determinately bald’ does not apply to o, that ‘determinately bald’ is undefined for o (ditto for ‘determinately, determinately bald’). (iv) It is the case, however, that o is not determinately not bald. (See G3.) We have now distinguished predicates which are merely weakly partial from predicates which are (also) partial in the original sense. Ordinary vague predicates like ‘blue’ and ‘bald’ are partial without qualification. Higher-order predicates built from them using the determinately operator are weakly partial (and correspondingly weakly vague). Is this the end of the story? Is there anything more to say about higher-order vagueness for partially defined predicates? I suspect there is.

358 • Essay Thirteen

Superundefinedness, Superdeterminateness, and Sharp Lines Call the individuals in regions 2 and 4 of G1 superundefined, meaning by this that they are individuals of whom we cannot, in principle, establish that ‘bald’ applies to them, that ‘bald’ doesn’t apply to them, or that ‘bald’ is undefined for them, no matter how much information we are given. Since we cannot establish any of these claims, we cannot justifiably accept them. More precisely, we cannot accept them while maintaining that in so doing we are not exercising our discretion by contextually changing the conversational standards governing the predicate ‘bald’. Call objects that have this status objects for which the predicate ‘bald’ is superundefined. More generally, when an object o has this status for an arbitrary predicate P, we say that P is superundefined for o. With this definition in place, it seems plausible to suppose that for any predicate P and object o, either (i) P applies to o, (ii) P does not apply to o, (iii) P is undefined for o, or (iv) P is superundefined for o. These categories really do seem to be jointly exhaustive. Supposing that they are, we may introduce an operator which attaches to a predicate P to form a totally defined predicate  superdeterminately P. Superdeterminately Predicates The predicate superdeterminately P applies to an object o just in case it is not the case either that (i) P does not apply to o, or that (ii) P is undefined for o, or that (iii) P is superundefined for o. Would it be a bad result if there really turned out to be such predicates? I don’t see that it would. The point of our discussion of higher-order vagueness for partially defined predicates has not been to avoid drawing sharp lines between all categories of objects to which one might think of applying a vague predicate. The point has been to accommodate what appears to be the genuine sense in which the higher-order predicate determinately P is vague (more precisely, weakly vague), when P is an ordinary vague predicate, like ‘bald’, or ‘blue’. We have done that. As for sharp lines, the important questions are ‘If they exist, what do they separate?’ and ‘How do they arise?’ The lines I have been concerned with arise from the nature of contextual theories— theories that hold that there is a range of discretion within which speakers may acceptably adjust the contextual standards of what counts as “blue,” “bald,” and the like. Since there are limits to the range of discretion that speakers have, there must be some individuals for which the rules of the language allow no discretion. For example, there must be some individuals for which any characterization conflicting with the characterization that the predicate applies to them is incorrect, no matter what the context.

Higher-Order Vagueness • 359

Let us focus on this class of individuals, and the line separating them from the next class of individuals. This is the line between regions 1 and 2 in the graph G1 for ‘bald’. Individuals in region 1 are such that it is determinate that ‘bald’ applies to them; hence, speakers have no option to characterize them in any other way. Since we know that individuals in region 2 are not determinately not bald, we know that one can correctly characterize the predicate as applying to them. However, if one does characterize ‘bald’ as applying to these individuals, we can’t say whether the rules of the language governing the predicate leave one any discretion to do otherwise. We may put this by saying that the individuals in region 2 are such that it is always correct to characterize ‘bald’ as applying to them, but we cannot say whether the reason this is correct is because the rules of the language determine this characterization, or because in characterizing the predicate as applying to these individuals one is adopting a contextual standard that makes it correct. The line between these things in region 2 (which always may correctly be said to be bald) and the things in region 1 (which may also always be correctly said to be bald) may very well be sharp. However, it is a line which, by its very nature, one would not expect speakers to notice. Hence, it is no embarrassment to the theory that they don’t. Implications for the Sorites If I am right, then semantic models of vague predicates as both partial and context sensitive do not allow one to avoid the conclusion that the meanings of these predicates impose classifications of individuals in their domains of potential application into sharply defined categories. Because of this, strengthened versions of the sorites paradox can be constructed exploiting this fact. A Strengthened Sorites Argument A man with no hair is superdeterminately bald. For all x, if x is superdeterminately bald, then a man with one more hair is too. So everyone is superdeterminately bald. Because of this one might wonder whether in using the semantic model I have defended we have made any progress in defusing the paradox. In my opinion we have, though we certainly have not fully resolved it. The puzzle that remains is how the linguistic behavior on which the semantics of our language supervenes results in such fine-grained classifications of the objects in the domains of our predicates. This is a problem for all theories of vague terms, and nothing I have said constitutes an answer to it. However, if I am right about the semantics of these terms, then, it seems to me, these fine-grained classifications turn out to be less paradoxical and

360 • Essay Thirteen

problematic than they were before. In particular, they do not pose the threat to our notion of linguistic competence that would be posed by a sharp, fine-grained bifurcation of the domain into objects to which a predicate definitely applies and those to which it definitely does not apply. The distinction between truth and falsity, or truth and untruth, is very important to speakers; and the norms of language use presuppose that we are able to closely track the truth. One lesson that has sometimes been drawn from traditional versions of the sorites is that in order to avoid absurdity, we must embrace a semantic theory that distinguishes between those objects of which a predicate is true and those of which it is not true in such a precise and fine-grained way that we can no longer view ordinary speakers who understand the predicate as competent to make the distinction, or as able to track the truth of statements made using it. That is paradoxical. How can a distinction based on meaning that is so important to language use be opaque to fully competent speakers who understand the meanings of their words? If the meaning of an ordinary predicate imposed a precise, fine-grained classification between objects to which it applied and those to which it did not, wouldn’t fully competent speakers know this, and be able to locate the boundary with a high degree of accuracy? The virtue of the semantic account I have sketched is that it does not provoke these questions.10 The distinction between truth and falsity is important enough to speakers that we expect an account of meaning (which is grasped by competent speakers) to classify statements into those categories in ways that fully competent speakers in possession of all relevant nonlinguistic facts are able to approximate. By contrast, the sharp distinction between (i) statements the truth of which are determined by the rules of one’s language together with nonlinguistic facts and (ii) statements for which there is no saying whether their truth is so determined or whether their truth results from the exercise of speaker discretion in adjusting the boundaries of contextsensitive predicates is a highly theoretical one, of which speakers need have no clear and precise pretheoretical grasp. Since their shaky grasp of this distinction in no way impugns their competence, it is not paradoxical. Although all sharp, 10 More precisely, it doesn’t provoke these questions for ordinary predicates like blue and bald. Although related questions may arise for technical predicates, like superdeterminately bald, the sharp distinctions between things to which these predicates apply and those to which they don’t are defined in terms of the theoretically less troubling distinctions corresponding to the ordinary vague predicates they arise from.

Higher-Order Vagueness • 361

fine-grained distinctions imposed by the semantics of vague predicates are theoretically puzzling, they need not be paradoxical.11

References Soames, Scott. 1999. Understanding Truth. New York: Oxford University Press. ———. 2002. “Replies.” Symposium on Understanding Truth in Philosophy and Phenomenological Research 65:429–52. Williamson, Timothy. 2002. “Soames on Vagueness.” Philosophy and Phenomenological Research 65:422–28. 11

Thanks to Alexis Burgess for comments on an earlier draft.

ESSAY FOURTEEN

The Possibility of Partial Definition

The view of vagueness I favor is one according to which vague predicates are partially defined, in the sense of being governed by rules that provide sufficient conditions for them to apply, and sufficient conditions for them not to apply, but no conditions that are both individually sufficient and disjunctively necessary for them to apply, or not to apply, to an object. Objects for which such a predicate P is undefined are those for which neither the claim that P applies to them, nor the claim that it doesn’t, is sanctioned. For any name n, which we know to refer to o, we accept the claim that P applies to o just in case we accept Pn, which we accept just in case we accept the claim that Pn is true. When P is undefined for o, these sentences and claims are also undefined. Since even complete knowledge of all linguistic and nonlinguistic facts wouldn’t justify accepting or believing them, such acceptance or belief is always mistaken. The extension of P is the collection of things to which P applies; the antiextension is the collection to which P doesn’t apply. The determinateextension of P is the set of objects o, such that claim that P applies to o is a necessary consequence of the rules of the language plus all relevant underlying nonlinguistic facts. For some objects o the claim that o is not in the determinate-extension of P is true, whereas the claim that o is not in the extension of P is undefined. Similar remarks apply to the antiextension and determinate-antiextension of P. Corresponding to these distinctions, there is also a distinction between truth and determinate truth. In addition to being partially defined, vague predicates are context sensitive. Given such a predicate P, one begins with its (default) determinateextension and (default) determinate-antiextension. P is undefined for o just in case o is in neither of these sets. Since the sets don’t exhaust all cases, speakers have the discretion of adjusting the extension and antiextension to include initially undefined cases. When one does this by predicating P of o, or by denying such a predication, and one’s hearers go along, the extension (or antiextension) of P is contextually adjusted to include o, plus all objects that bear a certain relation of similarity to it. Observation predicates like ‘is blue’—which we learn by example rather than definition—are good illustrations. When learning the word, we note that certain objects are uniformly called ‘blue’, while certain others are uniformly called ‘not blue’. People say of o—which we note to

Possibility of Partial Definition • 363

be of a certain shade BE1—“That’s blue,” while saying of o*—which we observe to be of shade BA1—“That’s not blue.” On this basis, we come to accept the rule Blue 1. Blue 1 If o is BE1, then ‘is blue’ applies to o If o is BA1, then ‘is blue’ does not apply to o Further experience leads us to accept additional rules involving different shades, until, at some point, we are counted as understanding the predicate. At this point, our rules provide a rich set of sufficient conditions for application, plus a similar set for nonapplication. However, the requirement that the rules be adhered to by the great majority of speakers ensures that these conditions won’t be jointly exhaustive. Since there are shades of color, and objects having them, about which the rules say nothing, the predicate is partially defined. Context sensitivity means that we are free to adjust the extension or antiextension of the predicate to include objects for which it is undefined by the rules of the language. Suppose I call such an object o ‘blue’, and my hearers go along. Then, the extension of ‘is blue’ is contextually expanded to include o, plus others discriminately bluer than, or perceptually indiscriminable in color from, o. Let BEc be a shade that applies to precisely this class. The rule—‘If an object is BEc, then ‘is blue’ applies to it’—is thereby implicitly adopted in the conversation. Although not a rule of the language governing the predicate, it is one that speakers may adopt in particular contexts. The basic rule of the language governing the predicate (by providing its default determinate-extension and antiextension) is Blue-English, where BE and BA are families of shades uniformly characterized as blue, and not blue, respectively (leaving a gap). Blue-English If o exemplifies one of the shades in BE, then ‘is blue’ applies to o. If o exemplifies one of the shades in BA, then ‘is blue’ does not apply to o. What are these shades, and how do they become associated with the word ‘blue’? Colors are natural kinds, and color shades are surface reflectance properties of objects. Their association with ‘blue’ is illustrated by a simplifying idealization. Imagine a small, homogenous community introducing the word into their language. They notice a set BEo of perceptually similar objects (of varying shades within BE) which are easily discriminable from another set BAo of objects (of varying shades within BA). They introduce the word ‘blue’ with a reference-fixing stipulation.

364 • Essay Fourteen

Intro.

‘Blue’ is to designate the property of object surfaces causally responsible for the fact that (nearly) all members of BEo appear similar to one another, and different from BAo. Hence, ‘is blue’ will apply (at any world-state) to all and only those objects the surfaces of which have the property which (in the actual world-state) causally explains why members of BEo look similar to us, and different from members of BAo.

This stipulation is, of course, a fantasy. The term ‘blue’ could have been introduced in this way, and it behaves pretty much as if it had been so introduced. However, no such stipulation need ever have occurred. It is enough if speakers simply started calling things ‘blue’, with the intention that the predicate was to apply, not only to certain objects they had encountered, but also to those sharing the property of surfaces that explained their appearance. Finally, it must be remembered that in discussing the idealized stipulation, as well as the more realistic process of introduction it summarizes, we are not talking about a semantic rule of the language, mastered by speakers, stating the meaning or reference of a term. Although the stipulation mentions particular objects involved in the introduction of ‘blue’, it is not a semantic rule of English that this, that, or the other object is blue. Instead, the stipulation summarizes a crucial element in the explanation of how the word ‘blue’ acquired the semantic properties it has— among them, the property of being partially defined.1

The Alleged Impossibility of Partial Definition That, in brief, is the account of vague predicates I favor. I now turn to an objection that seeks to establish, not just that the account is wrong, but that it is incoherent. According to the objection, made by Michael Glanzberg, there aren’t, and couldn’t have been, partially defined predicates in any language.2 His main argument, which is presented as an elaboration on one given earlier by Michael Dummett, is based on global claims about assertion.3 It is supposed to show that there can be no truth-value gaps—all propositions must be either true or false. Glanzberg states the argument as follows: 1 This sketch of vague predicates, including color terms, summarizes more detailed discussions in Soames (1999, chap. 7) and essays 7 and 13 in this volume. 2 Glanzberg (2003). 3 Dummett (1978).

Possibility of Partial Definition • 365

(i) Speech acts, including assertions, are moves within a practice of using language which is (partially) rule-governed. . . . As such, speech acts have intrinsic purposes [norms]. (ii) The intrinsic purpose [norm] of assertion is to convey the information that something is the case, i.e., to assert s is to convey the information that s.4 (iii) Combining (ii) with the idea that propositional contents encapsulate truth conditions implies a form of the ‘truth-assertion platitude’, for the intrinsic purpose [norm] of assertion: the intrinsic purpose [norm] of assertion is to assert that truth conditions obtain. (iv) The truth of a claim is thus fundamentally a matter of a purposive act achieving its intrinsic purpose [conforming to its intrinsic norm].5 Elaborating on this conclusion, Glanzberg says: Assessing for truth is a matter of assessing a purposive act for success. We may thus think of truth itself as having a point or purpose, in so far as it is correctly applied exactly when a purposive act achieves its purpose. The same may be said for truth values. . . . Any assignment of truth value amounts to an assessment of whether a purposive act has achieved its purpose. [My emphasis] [T]he value true corresponds to the intrinsic purpose of an assertion being achieved, and false corresponds to it failing to have been achieved. It appears evident that these are the only ways that an assertion can be assessed for whether it has achieved its intrinsic purpose. It either has or has not done so.6 [My emphasis] This points to the following conception of the intrinsic purpose, or norm, of assertion. The Glanzberg-Dummett Account of the Norm of Assertion GD1. For any proposition p, an assertion of p is correct (satisfies the intrinsic norm of assertion) just in case p is true. GD2. An assertion is incorrect (fails to satisfy the intrinsic norm governing assertion) just in case p is false. 4 Glanzberg’s (2003) use of metalinguistic variables and corner quotes requires correction. The final clause of (ii) should be understood: i.e., to assert (the proposition expressed by) s is to convey the information (proposition) denoted by æthat s”. 5 Glanzberg (2003, 159). 6 Glanzberg (2003, 159, 165–66).

366 • Essay Fourteen

The import of this conclusion for theories of truth-value gaps, and/or partial definition, is easy to see. Any theory that maintains both that some propositions are neither true nor false, and that the assertion of such a proposition is incorrect because it violates the norm of assertion, is incompatible with (GD2). Thus, establishing (GD2) would be sufficient to refute any such theory. It would also be sufficient to refute theories that embrace partial definition, in my sense. On the one hand, these theories insist that asserting an undefined proposition violates the norm of assertion, and so is incorrect. On the other hand, in calling the proposition undefined, the proponent of partial definition is committed to rejecting the claim that it is untrue—thereby violating (GD2). Thus, accepting (GD2) requires rejecting partial definition. Moreover, the friend of partial definition—who doesn’t assert the existence of propositions that are neither true nor false, doesn’t object to identifying falsity with untruth, and is happy with contraposition (in the sense of accepting If ~B, then ~A whenever he accepts If A, then B)—recognizes that (GD1) entails (GD2). Thus, (GD1) is incompatible with partial definition. But how, exactly, is (GD1) supposed to follow from the premises of Glanzberg’s argument? Premise (ii) tells us that the aim of asserting the proposition p expressed by a sentence S is to convey that which S expresses, namely p. But the claim that conveying p is the aim of asserting p doesn’t advance the argument. Nor does premise (iii), which says, in effect, that the aim of asserting p is to assert that the truth conditions of p “obtain.” Since for conditions to “obtain” is just for them to be satisfied, this amounts to the claim that the aim of asserting p is to assert that p is true— which is, at best, parasitic on the triviality that the aim of asserting p is to assert p. The problem reappears in a further remark Glanzberg makes. The intrinsic purpose of assertion is to say that the truth conditions expressed obtain. This purpose is achieved just when the proposition expressed is true.7 [My emphasis] But this is a non sequitur. If my purpose is simply to say that the truth conditions of p obtain, and hence to commit myself to the claim that p is true, I can easily achieve that purpose even if p is false, or undefined. After all, it is perfectly possible to say of any proposition that it is true. Thus, Glanzberg has no argument for (iv), which is supposed to abbreviate (GD1). However, this needn’t be fatal, since (GD1) and (GD2), which can be broken into pairs of quantified conditionals, have some impendent plausibility.8 7

Glanzberg (2003, 164). In discussing these issues I take “correct” and “incorrect” to be jointly exhaustive (when applied to assertion). Although this is an idealization, it doesn’t affect the issues at hand. 8

Possibility of Partial Definition • 367

GD1a. For any proposition p, and assertion A(p) of p, if A(p) is correct (satisfies the intrinsic norm of assertion), then p is true. GD1b. For any proposition p, and assertion A(p) of p, if p is true, then A(p) is correct (satisfies the norm). GD2a. For any proposition p, and assertion A(p) of p, if p isn’t true, then A(p) is incorrect (doesn’t satisfy the norm). GD2b. For any proposition p, and assertion A(p) of p, if A(p) is incorrect (doesn’t satisfy the norm), then p isn’t true. (GD1a) and its contraposed version, (GD2a), are unproblematic. Since the assertion of an undefined proposition p violates the norm of assertion, instances of these principles corresponding to p will be true—by falsity of antecedent in the case of (GD1a), and by truth of the consequent, in the case of (GD2a). Thus, it is only (GD1b) and (GD2b) that are potentially problematic for partial definition. However, these principles are incorrect. In the presence of the (a) principles, what the (b) principles tell us is that all there is to the intrinsic norm of assertion is the directive to assert truths. But, as Timothy Williamson has argued, this is implausible.9 Assertion isn’t the only speech act that aims at truth. That other truth-directed acts— like conjecturing or predicting—put less stringent demands on the agent than does assertion suggests that there is more to assertion than aiming at truth. This is born out by cases—e.g., those involving lotteries—in which we aren’t warranted in asserting certain truths, even though they are highly probable on our evidence. In these cases one believes, but fails to know, some true proposition p, even though the odds in favor of p are very heavy. The fact that one isn’t warranted in asserting p, despite reasonably believing p to be true, suggests that assertion requires what is missing in these cases—knowledge. As Williamson notes, this explains why the question “How do you know?” is a standard way of challenging an assertion. The question presupposes that an agent who has asserted p should know p— which is just what one would expect if knowledge, rather than truth, was the norm of assertion. These and related considerations support replacing the GlanzbergDummett truth-based norm with the Williamsonian knowledge-based norm. Williamson’s Account of the Norm of Assertion W1. For any proposition p, an assertion of p is correct (satisfies the norm of assertion) just in case the agent knows p. W2. An assertion of p is incorrect (fails to satisfy the norm) just in case the agent doesn’t know p. 9

Williamson (1996).

368 • Essay Fourteen

As before, we can divide each of these claims into a pair of claims. W1a. For any proposition p, and assertion A(p) of p, if A(p) is correct (satisfies the norm of assertion), then the agent knows p. W1b. For any proposition p, and assertion A(p) of p, if the agent knows p, then A(p) is correct (satisfies the norm). W2a. For any proposition p, and assertion A(p) of p, if the agent doesn’t know p, then A(p) is incorrect (doesn’t satisfy the norm). W2b. For any proposition p, and assertion A(p) of p, if A(p) is incorrect (satisfies the norm), then the agent doesn’t know p. Glanzberg’s principles (GD1a)/(GD2a)—which, as we have seen, are unproblematic for theories of partial definition and undefined propositions— are entailed by (W1a)/(W2a), and so have the status of derived norms of assertion. Since (GD1b)/(GD2b) conflict with (W1a)/(W2a), they must be rejected. What remains of his argument against partial definition is, therefore, reducible to the question of whether accepting partial definition is compatible with accepting (W1)/(W2). Since these two principles are interderivable, we may concentrate on (W2). Can I admit that the assertion of an undefined proposition is incorrect because it can’t be known, without attributing its unknowability to its not being true. I should think so. From the beginning, I have said that it is a mistake to assert an undefined proposition p because even complete knowledge of linguistic and nonlinguistic facts wouldn’t justify accepting p, as opposed to its negation. If this point can be extended to an explanation of why one can’t know p, then partial definition and undefined propositions will have been rendered compatible with the correct account of the norm of assertion—and the Glanzberg-Dummett argument will have been rebutted.

The Unknowability of the Undefined Why, then, can’t one know the undefined? Since the Glanzberg-Dummett argument purports to rule out very possibility of a language containing partially defined predicates, I will frame my rebuttal around a simple, artificial example, which parallels, for color words, an example I have used in other contexts.10 Imagine members of a small linguistic community living on a desert island, cut off from the outside world. Sharing no antecedent common language, they set about to create one. Color words are introduced by authoritative stipulation. One of these, ‘bluege’, is introduced by 10 I refer to the smidget example (the idea for which was originally suggested to me by Nathan Salmon) discussed in Soames (1999, chap. 6).

Possibility of Partial Definition • 369

applying it to examples. As it happens, the island contains objects of various shades of blue each of which is stipulated to be bluege, and various shades of green and other colors, which are stipulated not to be bluege.11 However, a few shades remain unclassified, because they aren’t exemplified on the island. Among them are shades intermediate between the least blue of those stipulated to be bluege and the most blue-like of the greens stipulated not to be bluege. Since there is no pressing need to decide the status of these shades, the gap goes unremarked. At this point, speakers agree that their language contains a meaningful term ‘bluege’, governed by the authoritative, meaning-giving stipulations summarized in BluegeIsland. (BIE is the family of exemplified blue shades on the island; BIA is a family of exemplified nonblue shades.) Bluege-Island If o exemplifies one of the shades in BIE, then ‘bluege’ applies to o If o exemplifies one of the shades in BIA, then ‘bluege’ does not apply to o Is ‘bluege’ partially defined? Consider a shade, INT, intermediate between blue and green. A sequence of four barely discriminable shades separates INT from the least blue shade in BIE, and a similar sequence separates it from the most blue-like shade of green in BIA. Does ‘bluege’ apply to objects exemplifying INT? The stipulations don’t tell us. Objects exemplifying INT haven’t been stipulated to be bluege, or not to be. Since they are as perceptually similar to those stipulated to be bluege as they are to those stipulated not to be bluege, the case for classifying them one way is no better than the case for classifying them the other. The issue isn’t how to extend the meaning of ‘bluege’. That’s a future matter for the Islanders to consider. Rather, the issue is whether ‘bluege’ already applies, or doesn’t apply, to objects exemplifying INT. Since there is no more support for one of these alternatives than for the other, the facts don’t determine either one. Hence ‘bluege’ is undefined for objects exemplifying INT, and complete knowledge of all relevant facts would neither justify taking ‘bluege’ to apply to them, nor justify taking it not to apply to them. This explains why someone omniscient about all the relevant facts wouldn’t know any proposition p predicating ‘bluege’ of an object for which it was undefined.12 That explanation doesn’t say that p isn’t true. 11

We may imagine that the comparative, ‘blueger than’, is similarly introduced. Here and throughout I adopt the simplifying assumption that, for all the cases under discussion, an agent who understands a sentence S that expresses a proposition p (in a context C) will accept/believe-true (be justified in accepting/believing-true) S (in C) iff the agent believes (is justified in believing) p. Although Kripkean Pierre-type cases show that this principle needs modification, the complications don’t affect the issues raised here. 12

370 • Essay Fourteen

It says that knowing p isn’t possible because the agent lacks the justification needed for knowledge. In general, an agent whose justification for a proposition is no better than his justification for its negation won’t know that proposition, even if he believes it. But if agents who know all the ‘bluege’-relevant facts lack the justification required for knowledge of p, the same will be true of ordinary, nonomniscient agents, who know less. To be sure, the acquisition of additional information bearing on a hypothesis q sometimes puts an agent with partial information about q in a worse position to know it—as when coming to know a true, but misleading, defeater undermines one’s initial knowledge of q. In all such cases, however, further knowledge, defeating the defeater, reinstates one’s original knowledge. Although selective bits of additional knowledge sometimes put one in a worse position to know something, knowing all the relevant facts never does. Thus, if the nonomniscient could know p, the relevantly omniscient could too. But since the omniscient can’t, no one else can either. Appreciating this point requires distinguishing justification, in the sense that it is needed for knowledge, from mere reasonableness of belief. Undefined propositions can be highly probable on one’s evidence, making it perfectly reasonable to believe them. However, this doesn’t provide the justification needed for knowledge. As lottery examples have taught us, even when the probability of a proposition on one’s evidence is arbitrarily high, one may fail to know it—because one’s evidence isn’t of the right sort. This point, which holds for true propositions, doesn’t cease to hold for the undefined. When justification is understood as the cognitive requirement needed for knowledge, believing the undefined is unjustified, no matter how reasonable it is in certain cases. This completes my rebuttal of the Glanzberg-Dummett argument that the intrinsic norm of assertion rules out partial definition. The first step was to replace their inadequate conception of this norm with a better one—according to which what assertion requires is knowledge of the proposition asserted. The second step was to explain why undefined propositions can’t be known in a way that doesn’t commit one to their not being true. Combining both, we have an account of why asserting an undefined proposition violates the norm of assertion, and so is a mistake.

Partial Definition and the Excluded Middle Rebutting the objection doesn’t, of course, establish that partial definition really is possible. It does, however, justify giving that possibility some weight. Absent compelling arguments to the contrary, we are, I think, prima facie justified in taking the possible term ‘bluege’ to be partially

Possibility of Partial Definition • 371

defined. What about the “law” of the excluded middle? It is often thought that acceptance of partial definition brings with it rejection of some of instances the “law”—on the grounds that a disjunction is undefined when both disjuncts are. While plausible, this point is less obvious than it first seems. When S is undefined, the rules governing S, plus the totality of facts relevant to evaluating S, don’t determine that S is true, or that it isn’t. As a result, one can’t know that S is true, and asserting that it is true is a mistake. Similar points hold for the proposition expressed by S. But when S is a disjunction Φ or Ψ, how do we show that S is undefined in this sense if Φ and Ψ are? It does not, in general, follow from the fact that asserting each of two propositions is a mistake that asserting their disjunction is. Nor does it follow that one who fails to know each of two propositions also fails to know their disjunction. What about determination of truth by the totality of linguistic and nonlinguistic facts? Does the claim that the truth of a disjunction is not a necessary consequence of those facts follow from the claim that the truth of neither disjunct is? That depends on what counts as a necessary consequence of what—which in turn depends in part on whether Φ or ~Φ is itself necessary. Since this is an instance of the very question we are trying to decide, we must be careful not to presuppose the answer we are trying to justify. How, then, might one combine partial definition with unqualified acceptance of excluded middle? On scheme for doing so is supervaluationism. One starts with an intended model M that assigns interpretations in which some sentences are true, some are false, and some are neither. A sentence S is counted as true simpliciter iff S is true in every admissible bivalent extension of M. S is false simpliciter iff S is false in all such extensions. Otherwise S is neither true nor false. Since S or ~S is true in all bivalent extensions, the “law” of the excluded middle is preserved, even when both disjuncts are neither true nor false. Nevertheless, classical supervaluationism doesn’t reconcile acceptance of excluded middle with the kind of partial definition given here. For the classical supervaluationist, the claim that S isn’t true follows from the claim that S is undefined. For me, it doesn’t; rather, the claim that S isn’t true is undefined when S is. This mismatch could, in principle, be repaired. Instead of holding that S is true iff S is true in all admissible bivalent extensions of the initial partial model M, and false iff S is false in all such extensions, one might stipulate that S is true, if S is true in all admissible bivalent extensions of M, and that S is not true, if S is false in all such extensions. Since these stipulations give sufficient conditions for being true, and sufficient conditions for not being true, while saying nothing about sentences with different truth-values in different classical extensions, such sentences will be undefined in my sense. The resulting system preserves the “law” of the excluded middle, while

372 • Essay Fourteen

allowing a form of partial definition that assimilates sentences and propositions that are not true to those that are false. However, it is still not what we want. In either the classical or the revised form, supervaluationism violates the truism that D or t is true just in case D is true or t is true, and hence that a disjunction can’t be true unless one of its disjuncts is. Since this truism is essential to our ordinary understanding of ‘or’ and ‘true’, supervaluationism doesn’t give the right account of the truth conditions of complex sentences of natural language. It is also explanatorily baroque. In order to determine whether S is true or not, supervaluationism requires one to first determine whether S is true in all admissible bivalent models, false in them all, or true in some and false in others. This presupposes a notion of truth in a model antecedent to the official supervaluationist notion of truth, plus an antecedent logic used to calculate which sentences are true in which models. The idea that there is both a hidden truth and a hidden logic—conceptually prior to the ordinary notion of truth we apply in language, and the logic we employ when using it—is implausible, as well explanatorily tendentious. Since the (classical) laws of the hidden logic are simply taken for granted, it is hard to see how supervaluationism can be used to explain or justify them. The ineffectiveness of supervaluationism as a semantics doesn’t preclude limited supervaluationist reasoning from resolving certain kinds of indeterminacy. Suppose it is determinate that an agent asserts some proposition, but indeterminate which of p1 to pn is asserted.13 Although what is asserted is indeterminate, supervaluationist reasoning might still be used to explain how it is determinate that the speaker said something true, or something untrue—provided p1 to pn are all true, or all untrue. However, this isn’t the kind of indeterminacy to which partially defined predicates give rise. Suppose a speaker says of o “That’s bluege,” when in fact o exemplifies the shade, INT, indeterminate between blue and green. There is no indeterminacy about what is asserted in this case. It is not as if there is a family of totally defined properties B1…Bn such that it is determinate that the agent asserted the proposition that o has one of these properties, but indeterminate which. Rather, the asserted proposition predicates the partially defined property being bluege of o. The reason it is indeterminate whether what is said is true, is that the truth-value of this proposition—which was determinately asserted—is indeterminate. Supervaluationism doesn’t fit this kind of case. This leaves us back where we started—trying to decide whether partial definition requires rejecting some instances of excluded middle. We have seen that supervaluationism does not show that partial definition can be combined with unqualified acceptance of the law. But, I haven’t yet argued 13

See Soames (2002a, 81–83, 337–38) for potential examples.

Possibility of Partial Definition • 373

that the two can’t be combined. Can a disjunction be determinately true, even if its disjuncts are undefined? Can the claim that D or t is true be a necessary consequence of the rules of the language, plus the underlying nonlinguistic facts, even though neither the claim that D is true, nor the claim that t is true, is? When D and t are unrelated, it’s hard to see how the rules of the language, which are silent about both, could be definitive about the disjunction. How about when one disjunct is the negation of the other? Is there something about this case that gives it a special status, rendering the attribution of truth to (1) a necessary consequence of the rules of the language, plus the underlying facts, even though the same can’t be said for the disjuncts, (2a) and (2b)? (1) N is bluege or N is not bluege. (2) a. N is bluege. b. N is not bluege. It is difficult to see what it might be. Suppose we did take (1) to be true. Surely, we would have to say the same about (3) and (4), where ‘N*’ names the same object as ‘N’, and ‘M’ names a qualitative duplicate of that object (both exemplifying INT). (3) N is bluege or N* is not bluege. (4) N is bluege or M is not bluege. Intuitively (1), (3), and (4) should be treated similarly—all true, or all undefined. However, we have no explanation of how to get the result that the truth the latter is a necessary consequence of the rules of the language plus the underlying facts. Appealing to (5) won’t help. (5) It is a necessary consequence of the rules of the language, plus the underlying facts, that N is bluege, N* is bluege, and M is bluege are all true, all untrue, or all undefined. To guarantee that (3) and (4) are (determinately) true, if (1) is, we need something like (6). (6) It is a necessary consequence of the rules of the language, plus the underlying facts, that substitution of M is bluege or N* is bluege for N is bluege in any true disjunction always preserves truth. Though (6) is quite reasonable, the rationale for it makes it difficult to assign truth to (1). The reason we find (6) plausible is, I think, that we find two ideas compelling: R1. The status of a disjunction is entirely dependent on the status of its disjuncts.

374 • Essay Fourteen

R2. We have as much reason for taking N* is bluege and M is bluege to be true as we have for taking N is bluege to be true. But if R2 is compelling, so is R3—which, in the presence of R1, leads to (7). R3 We have as much reason for taking N is not bluege to be true as we have for taking N is bluege to be true. (7) It is a necessary consequence of the rules of the language, plus the underlying facts, that substitution of N is not bluege for N is bluege (or vice versa) in any true disjunction always preserves truth. Although (7) is, I think, as well motivated as (6), it clearly precludes taking (1) to be true. The lesson to take from this is that accepting partial definition and undefined propositions, while trying to retain a completely unrestricted version of excluded middle, is a dubious business. Although it is formally possible to combine the two, the resulting systems seem ill motivated. I therefore conclude that our prima facie justification for taking partial definition to be possible provides prima facie justification for rejecting the unrestricted excluded middle. We aren’t, of course, justified in accepting the negations of any of its instances. That would be incoherent. Rather, we have reason to believe that some of those instances are undefined—in the sense that asserting them would be a mistake, that knowing them is impossible, and that their truth is not determined by the totality of linguistic and nonlinguistic facts. I haven’t argued that partial definition or undefined sentences exist in English. However, I have tried to make such analyses more plausible by addressing the familiar objection that they impose “the high price of giving up classical logic.” Of course, this talk of price is metaphorical. The issue is descriptive, not volitional. We aren’t deciding how to reason, and looking for the most economical way of doing so. The issue is whether unrestricted versions of all classical “laws” are true. The idea that a theory pays a high price for refusing to agree that they are is just the idea that the “laws” seem, initially, to be so. Thus, a theory that doesn’t embrace them in full generality has some explaining to do. I have tried to provide the beginning of such an explanation. Later, I will say a word about what the explanation has to say about seemingly more obvious “laws” like ~(Φ and ~Φ).

Partial Definition, Context Sensitivity, and Ignorance The view I favor is one in which vague predicates in natural language are both partially defined and context sensitive. In order to more closely

Possibility of Partial Definition • 375

approximate the natural language case, let’s add context-sensitivity to the rules governing ‘bluege’. The imagined situation is as before, with ‘bluege’ being introduced by authoritative stipulation, accepted by everyone in the community. The result is a partially defined predicate the meaning of which is stable, due in part to the fact that there are few, if any, objects for which it is undefined among those standardly talked about. At some point, the introduction of such objects changes the situation. Some speakers notice the new shades, and when speaking about them use expressions like ‘bluegish’, ‘kind of bluege’, and ‘more bluege than greenge’. Both the shades previously called “bluege,” and those called “not bluege,” continue to be uniformly so characterized. But, there is contextual variation in how shades in the intermediate range are described, and sometimes speakers find themselves at loss for words. Responding to this need, speakers start allowing themselves the freedom to apply ‘bluege’ to objects for which it had initially been undefined. There is, unsurprisingly, variation in how this is done. In some contexts— depending on the audience, subject, and time—speakers are more expansive in what they are willing to call ‘bluege’ than they are in others. However, no one worries about this. No one thinks that there is just one right way to apply the word to objects within its initially undefined range. Instead, it is recognized to be a matter of decision—with different reasons yielding different results in different cases. Nor is there any sense that the standards adopted in a given context must settle, for each possible shade of color, whether objects of that shade are to be in the extension of ‘bluege’ or not. It is enough if the adopted standard allows determinate evaluation of all conversationally relevant propositions. If the predicate does remain undefined for some objects, agents presented with such an object, and pressed to classify it, will often be indifferent—and say things like “Its sort of bluege and sort of not,” or “You really can’t say that its one or the other,” or “It doesn’t matter, call it what you like.” At this stage, ‘bluege’ is both partially defined and context sensitive. The determinate-extension and antiextension of the original term have become the default determinate-extension and default determinate-antiextension of the new context-sensitive term. Its meaning is a function from contexts of utterance to members of a restricted family of properties (many of which are partially defined). One member of the family—the default semantic content of ‘bluege’—is a partially defined property that applies to all members of the default determinate-extension, fails to apply to all members of the default determinate-antiextension, and is undefined for everything else. This property plays two roles. First, it is the semantic content of the term unless something about the context selects a different property. Second, it fixes the boundaries of allowable contextual variation in the use of ‘bluege’. Each object it (determinately) applies to is one that

376 • Essay Fourteen

every contextually possible semantic content of the term (determinately) applies to, and each object it (determinately) doesn’t apply to is one to which no such content (determinately) applies. Properties in the family of possible semantic contents of ‘bluege’ differ only in how they divide up the objects for which the default content of ‘bluege’ is undefined. Of course, not every way of dividing up those objects corresponds to a property in the family. For every such property P and object o, if P (determinately) applies to o, then P (determinately) applies to all objects as blue as, or bluer than, o. When an object onew for which the term had previously been undefined is called ‘bluege’ in a context Cnew, and conversational standards are thereby adjusted, the property selected as semantic content in Cnew (determinately) applies to objects perceptually indistinguishable in color from onew, plus those discriminately bluer. The resulting line between items to which ‘bluege’ applies and those for which it remains undefined will be as sharp as the notion of objects being perceptually indiscriminable in color.14 Although it is easy to be deceived about precisely where this line is, its location can, in principle, be identified.15 Let oI be perceptually indistinguishable from onew, and o2 be perceptually indiscriminable from o1— each located just one indiscriminable step beyond its predecessor in a sorites sequence. Looking at the pair and correctly judging them to be indiscriminable, one can come to know of o1 that it falls within the extension of ‘bluege’ in Cnew. A similar point holds for o2—which, though indiscriminable from o1, is discriminately less blue than onew. Correctly judging this to be so by looking at onew and o2 together, one can come to know that ‘bluege’ is undefined for o2 in Cnew. So far, we haven’t found any facts about ‘bluege’ that are in principle unknowable. The best place to look for such facts is the line between the default determinate-extension of ‘bluege’ and the objects for which it is initially undefined. Up to now, I have treated this line as if it were sharp. However, that’s an artifact of the example. In my fable, ‘bluege’

14 This statement of the rule for adjusting the extension of the predicate is a rough approximation. A more accurate statement stipulates that ‘bluege’ comes to apply to all objects with the surface spectral reflectance property causally responsible for the perceptual indiscriminability from onew of all, or nearly all, of those objects the color of which is indiscriminable from onew. This opens up the possibility of isolated instances of “fools bluege,” in which something may be indiscriminable from onew for nonstandard reasons, without having the surface spectral reflectance property required to count as bluege, as well as possibility that some objects with the property might—for extraneous reasons—not be perceptually indiscriminable from onew. (Similar remarks apply to clauses concerning objects discriminately bluer than onew.) Although these standard natural-kind complications are mostly irrelevant to our concerns, they do introduce another source of potential vagueness (“casually responsible . . . for all or nearly all . . .”). 15 For discussion, see Soames (1999, 212–17).

Possibility of Partial Definition • 377

is originally introduced as a partially defined term, in which the boundary between the objects to which it applies and those for which it is undefined is determined by a universally agreed upon, authoritative stipulation. When context sensitivity was introduced, I imagined this boundary simply remaining in place, with the initial semantic content becoming the default content of the new term. The resulting semantic model—though realizable in principle—isn’t realistic in practice. In more realistic cases, the default semantic content of a context-sensitive term will depend on an overall pattern of agreement among uses by different speakers. In such cases, the boundary between what’s in its default extension and what isn’t will be blurred. Since this boundary is indeterminate, no single partially defined property can, strictly speaking, be said to be the default semantic content of the term.16 Instead, there will be a family of barely differing properties, each of which is a candidate for being that content. Since it is indeterminate which of these properties is the default semantic content of the term, there will be a limited range of cases in which it is indeterminate whether the claim that ‘bluege’ applies to o is true because the rules of the language, plus facts about o, determine that its true, or because the speaker has exercised minimal discretion in adjusting the extension of ‘bluege’ to include o. What precisely is the range of this indeterminacy? Although speakers typically won’t care, the question is both troubling and theoretically puzzling. I don’t see any good alternative to the Williamsonian view that—somehow—the question has a definite answer, even though we can’t find it. If that’s right, then context sensitivity and partial definition don’t tell the whole story about vague language.

A Modest Case for Partial Definition and Context Sensitivity I conclude with a pair of modest arguments—first, that partial definition is part of the story, if context sensitivity is, and second, that context sensitivity is part of the story, if partial definition is. In other words, accounts of vague language that include both partiality and context sensitivity have advantages over those that employ only one. The first argument rests on an assumption that at least some vague terms are context sensitive, in the sense of having different semantic contents in different contexts of utterance. The meaning of such a term is a function from contexts to members of a family of related properties. Our 16 For a detailed explanation of the reasons for, and consequences of, this see essay 13 in this volume.

378 • Essay Fourteen

question is whether all of these properties have to be totally defined. The case that they don’t is based on the idea that the context sensitivity of a vague term is limited. For example, there are limits on what can truly and literally be called ‘green’, and limits on what can truly and literally be called ‘not green’. If ‘green’ is context sensitive, these limits will be reflected in the family of contextually possible semantic contents of ‘green’. No member of the family will apply to scarlet red roses, or fail to apply to healthy, well-watered patches of grass. If, in addition, ‘green’ is partially defined, then the set of things to which every member of the family (determinately) applies will be its default determinate-extension, while the set of things to which no member (determinately) applies will be its default determinate-antiextension. If, on the other hand, the term is required to be totally defined, there will be a corresponding pair of sets. Call them the minimal nonnegotiable extension and antiextension of the term. Either way, both theoretical models must specify the range of linguistically possible contextual variation in literal uses of the term. In both cases, this range will be subject to seeming indeterminacy and/or impenetrable Williamsonian ignorance. The two models differ in that, in the totally defined model, this ignorance comes on top of a different, and prior, form of impenetrable ignorance. For each contextually legitimate use of the totally defined term, there must be a precise, but unknowable boundary separating things to which it applies, from those to which it doesn’t. How this boundary is determined is, according to the model, shrouded in mystery. There is no corresponding mystery with the partially defined model. Once the range of contextual variation is determined, the extension and antiextension of the term, when used with its default semantic content, is completely given. Since subsequent contextual adjustments are essentially matters of stipulation, plus a similarity relation associated with the term, any further indeterminacy or ignorance is limited to vagueness or indeterminacy about the stipulation, or the similarity relation—unclarity about whether the speaker has really referred to this object o and called it ‘green’, or about whether, if he has, some other object o’ is discriminably greener than, or perceptually indiscriminable in color from, o.17 Although there may be cases in which there are no knowable, or even determinate, answers to 17 The similarity relation in question is the one that determines which objects are included in the contextually adjusted extension of the predicate as a result of the decision to include the new object o. On the partially defined model this is a relation born to o by all objects discriminately greener than o, plus those perceptually indiscriminable from o. On the totally defined picture, the relation will be much less restricted—applying to all objects discriminately greener than o, plus (perhaps) those objects that are more similar in color to o than to the least green object in the antecedently determined antiextension. Typically, the latter will allow much greater room for indeterminacy or mysterious ignorance.

Possibility of Partial Definition • 379

these questions, this area of indeterminacy or mysterious ignorance is sharply limited, and highly circumscribed. One loses this, if one insists on bivalence and total definition in every case. Since it is better to have less extensive theoretical mystery than more, there is something to be said for including partial definition in one’s account of vague terms—if one decides that they are also context sensitive.18 The second argument—for the desirability of context sensitivity given partial definition—involves the explanation of our differential reactions to apparent violations of different logical “laws.” I have already argued that the case for partial definition provides prima facie justification for rejecting certain instances of the “law” of the excluded middle. When F is undefined for o in the sense I have explained, and n names o, it is natural and correct to reject both the claim that Fn is true and the claim that ~Fn is true. But then, the truism that a disjunction can be true only if one of its disjuncts is, leads one to reject both the claim that Fn or ~Fn is true and Fn or ~Fn itself. Whatever one’s ultimate verdict on this result, the force of the reasoning is clear, and the attractiveness of the conclusion is easily recognizable. The same cannot be said for other consequences of the position. Corresponding to the truism about disjunction is a similar truism about conjunction—namely that a conjunction can be untrue only if one of its conjuncts is.19 Given both this truism and the undefinedness of Fn and ~Fn, one has no choice but to reject the claim that Fn and ~Fn is untrue—in which case one must reject both the claim that its negation, ~(Fn and ~Fn), is true, and ~(Fn and ~Fn) itself. But rejecting the law of noncontradiction seems far less defensible than rejecting excluded middle. Proponents of partial definition need some explanation of this. In my opinion, there is no plausible way to retain the undefinedness of  Fn or ~Fn, and its negation, without recognizing the similar undefinedness of (Fn and ~Fn), and its negation—when F is undefined for o. Thus, rejection of some instances of excluded middle should bring rejection of corresponding instances of noncontradiction in its wake. The reason that the rejected instances of noncontradiction seem correct, in a way that the corresponding instances of excluded middle don’t, is that the former are prey to a certain metalinguistic confusion that the latter aren’t. When, in addition to being partially defined, F is also context sensitive, it is easy to confuse the obvious truth (8) No matter what contextual standards one adopts governing F, it will not be the case that those standards (plus the nonlinguistic 18 19

An earlier statement of this argument can be found in Soames (2002b, 447–49). Thanks to Nathan Salmon for reminding me of the parallel.

380 • Essay Fourteen

facts) dictate that Fn and ~Fn are both true—and hence that  Fn and ~Fn is too with the undefined (and hence properly rejected) claims (9a–c). (9) a. (Fn and ~Fn) is never true. b. ~(Fn and ~Fn) is always true. c. ~ (Fn and ~Fn) It is because of this confusion that we are reluctant to reject instances of the “law” of noncontradiction, even when they are undefined. A similar confusion of the metalinguistic triviality (10a) with the undefined claim (10b) explains our (misguided) reluctance to reject undefined instances of the “law” (10c). (10) a. If a contextual standard dictates that Fn is true, than it will dictate that Fn is true. b. If Fn, then Fn is always true. c. If Fn, then Fn The proper explanation of these cases parallels the proper explanation of familiar so-called penumbral truths involving vague predicates. For example, although (11a) is undefined, it is easily confused with a metalinguistic truth (11b) that says something important about the partially defined, context sensitive predicate ‘is bald’.20 (11) a. If a man is bald, then he would be bald if he had one less hair. b. No matter what standards governing ‘is bald’ we adopt in a context, if according to those standards ‘he is bald’ applies to a man, then according to those same standards it would apply to him if he had one less hair. By contrast, there is no metalinguistic truth corresponding to the “law” of the excluded middle that makes us reluctant to reject undefined instances of them. The reason they seem easier to reject than some other classical “laws” is that rejecting them isn’t subject to the same pragmatic interference we encounter with the other laws. Logically, the different laws are in the same boat. Pragmatically, they differ in what they suggest about the context-sensitive semantic effects of context change. In this way, the proponent of partial definition can explain our differential reaction to violations of different classical “laws”—provided that the predicates in question are context sensitive, as well as partial.21 20

See Soames (2002b, 440–41). Thanks to Jeff King, John MacFarlane, Sebastiano Moruzzi, and Nathan Salmon for their helpful comments. 21

Possibility of Partial Definition • 381

References Dummett, Michael. 1978. “Truth.” In Truth and Other Enigmas, 1–24. Cambridge: Harvard University Press. Originally published in Proceedings of the Aristotelian Society 59 (1959): 141–62. Glanzberg, Michael. 2003. “Against Truth-Value Gaps.” In Liars and Heaps: New Essays on Paradox, ed. JC Beall, 151–94. Oxford: Clarendon Press; New York: Oxford University Press. Soames, Scott. 1999. Understanding Truth. New York: Oxford University Press. ———. 2002a. Beyond Rigidity: The Unfinished Semantic Agenda of “Naming and Necessity.” New York: Oxford University Press. ———. 2002b. “Replies.” Symposium on Understanding Truth in Philosophy and Phenomenological Research 65:429–52. Williamson, Timothy. 1996. “Knowing and Asserting.” Philosophical Review 105:489–523.

This page intentionally left blank

PA RT F O U R

Kripke, Wittgenstein, and Following a Rule

This page intentionally left blank

ESSAY FIFTEEN

Skepticism about Meaning: Indeterminacy, Normativity, and the Rule-Following Paradox

Quine and Kripke’s Wittgenstein both present “skeptical” arguments for the conclusion that there are no facts about meaning.1 In each case the argument for the conclusion is that (i) if there are facts about meaning (and propositional attitudes), then they must be determined by some more fundamental facts, but (ii) facts about meaning (and propositional attitudes) are not determined by any such facts. Consequently there are no facts about meanings (or propositional attitudes). Within this overall framework, Quine and Kripke’s Wittgenstein differ substantially—both in their reasons for thinking that facts about meaning (and propositional attitudes) are not determined by other facts, and in their responses to the alleged elimination of these facts. Despite this, I believe that their arguments fail for essentially the same reason; each equivocates about what it means for one set of facts to determine another. Once the equivocation is eliminated, the arguments lose their plausibility.

Kripke’s Skeptical Argument I begin with Kripke’s Wittgenstein. His argument is based on an undeniable truism: What we mean by a word is not exhausted by the cases in which we, or those who have taught us the word, have actually used it. Rather, what we mean somehow determines the correct application of the word to an indefinite range of new, so-far unencountered, cases. Kripke’s skeptical argument challenges us to explain how this comes about. What is it about us that determines that the word, as we now use it, already applies in a definite way to cases we have not yet considered? 1 In discussing ‘Kripke’s Wittgenstein’ I have in mind the philosophical position presented in Kripke (1982, chaps. 2 and 3). Kripke presents this position both as an interpretation of the leading ideas of Wittgenstein’s Philosophical Investigations and as a philosophical point of view of independent interest. My discussion of the position will be concerned only with its content and merits, not its origin. The question of whether the position outlined by Kripke is an accurate interpretation of Wittgenstein is not directly relevant to my discussion.

386 • Essay Fifteen

If we mean anything at all by the word, then something must determine this, for otherwise we would be free to apply the word in new cases in any way we liked, without changing its previous meaning (or saying anything false). But surely we are not free to do this. So, if it is a fact that we mean so and so by a given word w, then some fact about us must determine in advance how w properly applies in new cases. This much seems undeniable. The surprise comes when we examine potential candidates for such a determining fact and find that none fills the bill. Because of this, the skeptic concludes, we have no choice but to admit that it is not a fact that we mean anything by w after all. Kripke uses ‘+’ to illustrate the point. This symbol, as we normally understand it, stands for the addition function, which assigns a unique natural number to each of infinitely many pairs of natural numbers. Thus, its range of application far exceeds the (relatively) small number of cases in which we have actually used it in computing sums. Consequently if we really do mean the addition function by ‘+’, then something about us must determine that when ‘+’ is applied to a pair of numbers we have not previously considered it always yields the sum of those numbers, rather than nothing, or the result of applying some other arithmetical operation. But, the skeptic argues, nothing about us does determine this. Wait a moment, one might object: Isn’t the skeptic’s position selfdefeating? In setting up the case, he uses the words sum and the addition function to pick out a certain function, which he himself seems to characterize as applying in a definite way to infinitely many pairs of arguments. The problem, of course, is that if he is right, this is precisely the sort of thing that cannot be done. Thus, doesn’t his own argument presuppose the falsity of its conclusion? The short answer is that, yes, in my opinion, there is an important sense in which the skeptic’s argument is self-undermining; if one succeeds in stating the argument, or in using it to justify one’s belief in the conclusion, then that very act falsifies the conclusion.2 The real challenge is to find where the argument goes wrong; and to do this it is helpful to have a formulation which is not (immediately) self-defeating. Kripke provides one. In his formulation we assume that there is no doubt about what we now use our words to mean. In particular, we all agree that at present we use ‘+’ to stand for the addition function, which applies in a definite and prescribed way to infinitely many cases. All arithmetical facts about this function, including all results of applying it to particular numbers, are 2 Stating the skeptic’s argument is not just a matter of uttering the words, but rather involves taking up genuine propositional attitudes toward the contents expressed by those words.

Skepticism about Meaning • 387

taken as given, and are not in dispute. What is in dispute is what we meant by ‘+’ in the past.3 Here we face the familiar problem. Although I performed lots of calculations in the past using ‘+’, there are infinitely many pairs of arguments that I never applied it to. Pick one such pair—say, 68 and 57. Doing the calculation now, I will tell you ‘68 + 57 = 125’. Everyone grants that this is arithmetically correct. However, the skeptic raises the question whether it is correct in a certain metalinguistic sense.4 Is it the case that in the past I used the ‘+’ sign to denote a function that gives 125 as value to the pair of arguments 68 and 57? More generally, is it the case that in the past I used ‘+’ to denote the same function I now do, the addition function?5 If so, then some fact about my past understanding of the ‘+’ sign must have determined that ‘125’ was the right answer to the question ‘What is 68 + 57?’ even though I had never considered that case. The skeptic challenges us to find such a fact. Certain natural responses are immediately disqualified. If, outside the context of the skeptical challenge, one were asked what past facts determined what we meant by a particular word, it might be natural to answer that the determining facts were facts about the beliefs and intentions connected with our use of sentences containing the word. For example, one might claim that my having meant addition by ‘+’ is a consequence of the fact that in the past when I used sentences containing ‘+’ I did so with the intention of communicating my belief concerning the result of adding certain numbers, rather than the result of applying some other function. Because of this, it was addition that I meant by ‘+’. Whether or not such a response is true, it is of no help in dealing with Kripke’s skeptical challenge. The skeptic raises a question about the content of the words we use. He does so knowing all the nonintentional facts there are about our past uses of the word—our linguistic performances, our behavior, and so on. Any attempt to answer his question by citing our past beliefs and intentions will be met with a mere reformulation of the original challenge. Granted, in the past we used the words 9 + 16 = 25 to express a belief. But what belief? A belief about addition, or a belief 3 The advantage of this way of formulating the problem is that it allows us to take the meanings of the words we use in stating the argument for granted while the argument is being given and evaluated. Hence we do not have to rely on the meanings of certain words in stating the argument while at the same time questioning what those words mean. 4 Kripke (1982) introduces this sense of metalinguistic correctness on p. 8. 5 The skeptic’s challenge is made graphic by his suggestion that perhaps in the past I used ‘+’ to denote not the addition function but the quaddition function, where the latter differs from the former in assigning the value 5 to all pairs of arguments greater than any arguments I had explicitly used in calculations. It is assumed, for the sake of discussion, that 68 and 57 are such a pair.

388 • Essay Fifteen

about some other function? To settle this question we must find some fact that determined the contents of our past beliefs and intentions. But this is just our original problem all over again. Having started by challenging us to find facts that determined the content of our words, the skeptic continues by challenging us to find facts that determined the contents of our mental states. Thus, if we are ever going to be able to answer the skeptic, we will have to find facts that determine content in general— linguistic and otherwise. The same sort of difficulty plagues other natural responses to the skeptic. For example, in the case of ‘+’ it is probably true that for most of us coming to understand the term was associated with learning a rule or algorithm for applying it. When confronted with a new calculation, we simply apply the algorithm and obtain the desired result. Since, in principle, the algorithm applies to infinitely many cases, it might seem ideally suited to answering the skeptic. One might say, for example, that the fact about me that determined that in the past I used ‘+’ to stand for a function that assigns the value 125 to the arguments 68 and 57, even before I performed the actual calculation, was simply the fact that the algorithm I associated with ‘+’ gives the result 125 in that case. The same point obviously can be made for all other cases. Thus, one might say, the fact that I meant addition by ‘+’ was determined by the fact that the algorithm I associated with it in the past always yields the sum of the arguments to which it is applied. But what is an algorithm anyway? Is it a collection of symbols (the symbols we use to express it), or is it the content of those symbols (what those symbols express)? If we think of the algorithm as just a collection of symbols, without any particular interpretation, then it won’t determine the value of any function at any argument, and so will provide no answer to the skeptic. If, on the other hand, we think of the algorithm as the content of the symbols used in stating it, then the skeptic will question whether we did associate that content with ‘+’ in the past. He does not doubt that in the past we associated the same symbols with ‘+’ as we do now—symbols which, under their present interpretation, express an algorithmic content that determines the addition function. However, he does doubt that we interpreted those symbols in the past in the same way we do now. In particular he doubts that, as we interpreted the symbols of the algorithm in the past, they provided a procedure that always yields the sum of a pair of numerical arguments. Our problem, then, is this. We think that our understanding of ‘+’ determines how it applies to an indefinitely large range of cases, even though at any given time we have considered only a small number of these. Thus, the fact that we mean one thing by ‘+’ rather than another cannot consist simply in the uses we have already made of it in particular calculations. What then does it consist in? A natural thought is that it consists in associating

Skepticism about Meaning • 389

‘+’ with a linguistic rule that determines its correct use. However this only postpones the problem, since we can now ask what our understanding of the symbols used in formulating the rule consists in. At some point this process of using symbols to interpret other symbols must stop. And when it does, we have to face the skeptic’s challenge directly. In general, if we ever mean anything by any expression, then there must be some expressions whose meanings are not determined either by the particular occasions we have used them in the past, or by any linguistic rules we have associated with them, or by any introspectable, content-bearing, mental images or psychological representations that may accompany their use.6 This realization leads to a different attempt to answer the skeptic—one that Kripke calls the dispositional view. Applied to ‘+’, it says that for us to use ‘+’ to denote one function rather than another is for us to have certain mental and behavioral dispositions. Specifically, in the past I used ‘+’ to denote the addition function in virtue of the fact that I was disposed when asked any question of the form ‘What is m + n?’ to answer by giving the sum of m and n. Thus, even though in the past I never was asked the question ‘What is 68 + 57?’, if I had been asked, I would have answered 125—just as I do now. Although the particular calculations I have performed have changed over time, my dispositions to answer arithmetical questions in specific ways has not. Consequently, the dispositionalist concludes, I meant the same thing by ‘+’ in the past as I do now. Paraphrasing Kripke, we may take (D) to be a preliminary statement of the dispositional analysis of ‘+’. D. One uses ‘+’ to denote a two-place numerical function f iff for all natural numbers m, n, and z, and numerals m, n, and z denoting them, one is disposed to answer z when asked the question What is n + m? iff f(m,n) = z.7 It is worth noting that on this account, the fact about me that determined what I meant by ‘+’ was an intentional fact that is itself subject

6 Kripke’s skeptic raises a challenge about content-bearing (i.e., intentional) facts. The challenge is to locate non-content-bearing (i.e., nonintentional) facts that determine the content-bearing ones. In attempting to meet this challenge, the nonintentional facts that one is allowed to appeal to are not restricted to behavioral facts, publicly observable facts, or even physical facts. Mental images, representations, feelings, and sensations can all be appealed to, provided that their ‘interpretations,’ or contents, (if any) are not taken for granted, but are themselves given a thoroughly nonintentional explanation. In certain cases, it is quite plausible to suppose that factors such as linguistic rules or representations really do play a role in determining the meanings of some expressions. A case in point is ‘+’ itself. However, in this, as in other cases, the role is dependent upon a prior explanation of the contents of the relevant rules, algorithms, images, or representations. 7 Kripke (1982, 26).

390 • Essay Fifteen

to skeptical worries of two different sorts. First, the disposition to answer a question in a certain way is a disposition to assert a certain proposition in answer to the question. But what, the skeptic will demand, justifies our characterizing an event that consists of my producing a numeral z in response to the utterance of an interrogative sentence, What is n + m?, as my answering the question—by asserting the proposition that n + m = z—as opposed to emitting a mere verbal reflex, or giving an arbitrary response without meaning anything? A thoroughgoing skeptic about meaning and propositional attitudes would not take this for granted. Consequently our statement of the dispositional view already presupposes more than the skeptic would allow. Second, in order to use the analysis to determine the function meant from my verbal performances one would have to be given the denotations of the numerals and other symbols used in questions and answers. These are just as subject to skeptical doubt as the interpretation of ‘+’. For these reasons, the dispositional analysis (D) stated here is itself too intentional to succeed as a response to the skeptic. However, I will put this objection aside temporarily in order to review two other objections discussed by Kripke—the error and finitude objections. After looking at these, our worries about the intentionality built into (D) will resurface in the discussion of Kripke’s normativity objection. The error objection is that the analysis does not make room for dispositions to misapply the expression. Since, according to it, any answer I am disposed to give to a particular query, What is n + m?, is definitive of what I mean by ‘+’, all such answers must be correct. But this is wrong. One can mean addition by ‘+’ even if one is disposed to make a few systematic mistakes in particular cases, and hence is disposed to give incorrect answers to some questions of the form What is n + m?. The finitude objection is that the analysis cannot account for the fact that our dispositions are finite, and will give out before the application of the term does. Certainly, one can mean addition by ‘+’ despite the fact that at some point the numbers to be added get too large for us to deal with. When this happens we have no dispositions to answer the relevant questions What is n + m?, even though the addition function denoted by ‘+’ continues to be defined.8 8 Both the error and the finitude objections can be generalized to cover dispositional analyses of other terms. In this connection, it is important to note that the finitude objection to the dispositional analysis of ‘+’ is just one instance of a more general objection. The objection is that a term may apply to certain objects even in cases in which those objects are, for one reason or another, epistemically inaccessible to us, and hence are not objects about which we have any normal dispositions. In the case of ‘+’ the infinity of natural numbers ensures that at some point they will get too big for us to consider. In the case of other words the reasons for epistemic inaccessibility may be quite different.

Skepticism about Meaning • 391

For these reasons the dispositional analysis, in the form we have stated it, is surely inadequate.9 Nevertheless, it is difficult to resist the idea that our dispositions to apply a term in particular cases play some role— perhaps along with other factors—in determining what the term means. With this in mind, we turn to Kripke’s normativity objection. Applied to the dispositional analysis, it is designed to show how far from providing an answer to the skeptic the analysis really is. Applied more generally, the normativity objection is seen as the most fundamental obstacle to any analysis that attempts to answer the skeptical challenge. The objections to the dispositional analysis based on error and finitude show that my dispositions to calculate with ‘+’ do not correspond perfectly to the addition function. The normativity objection maintains that even if my dispositions did correspond perfectly to the addition function, that would not be enough to show that I meant addition by ‘+’. The mere fact that I am now, and have been in the past, disposed to answer a question in a certain way does not show that I should answer it in that way. But it is a crucial aspect of meaning something by a term that one has adopted a standard to which one tries to adhere, and which provides a basis for judging responses to be correct or incorrect. Kripke contends that the dispositional analysis doesn’t provide such a standard.

9 One possible response to these objections would be to give up the dispositional analysis for ‘+,’ while retaining it for simpler notions like ‘successor,’ in terms of which ‘+’ could then be defined. Certainly, there is less room for error in applying the successor function than in applying the addition function. However, some room for error may remain, and, in any case, the finitude objection still seems to apply. Another possibility, considered by Kripke, is to reformulate the analysis so as to appeal to idealized dispositions—dispositions to answer questions What is n + m?, not in just any circumstances, but in certain idealized circumstances, in which one scrupulously ‘checks’ one’s work, and in which one’s mental capacities have somehow been enhanced to allow one to consider arbitrarily large numbers. Kripke’s criticism of this response is telling. In order for the analysis to be noncircular, one must not characterize the idealized dispositions in a way that presupposes in advance what we mean by ‘+’. For example, they cannot be characterized as dispositions to answer the relevant questions when we have been provided with sufficient mental capacities to allow us to correctly add any two numbers. However, if our characterization of the ideal dispositions is noncircular, then we will have little reason to be confident that they will in fact determine the desired function. It is also worth noting that the finitude and error objections cannot be overcome by appealing to the dispositions of the entire linguistic community of which the speaker is a part. These dispositions are just as subject to the objections as are the dispositions of an individual. These considerations constitute serious obstacles to the dispositional analysis. I do not maintain that they rule out all possible reformulations. Perhaps there are reformulations of the analysis that are capable of avoiding the error and finitude objections. Even if this is so, the normativity arguments will remain, and will, I suspect, require appeal to some factors over and above dispositions. See below.

392 • Essay Fifteen

According to Kripke, the main reason that it doesn’t is that it fails to explain how I can be justified in giving one answer rather than another when presented with a previously unconsidered calculation.10 The picture that emerges from his discussion is something like this: The skeptic gives me a new calculation that I have never done before. He asks me ‘What is 68 + 57?’ I answer that 68 + 57 = 125. The skeptic asks me to cite some past fact that justifies this answer, in the sense of showing that it is the answer I ought to give provided that I am now using ‘+’ with the same meaning as I did in the past. In effect, the skeptic is imposing the requirement that no fact will suffice to answer his challenge unless we can conclude from that fact that I ought to give ‘125’ as answer to the question ‘What is 68 + 57?’, provided that I now intend to use ‘+’ with the same meaning as I did before. Let me formulate this requirement as follows: NE. If the fact that F determined that (in the past) one meant addition by ‘+’, then knowing that F would, in principle, provide one with a sufficient basis for concluding that one ought to give the answer ‘125’ to the question ‘What is 68 + 57?’, provided one intends to use ‘+’ with the same meaning it had in the past. 10

A relevant passage is the following:

To a good extent this reply [the dispositional analysis] immediately ought to appear to be misdirected, off target. For the skeptic created an air of puzzlement as to my justification for responding ‘125’ rather than ‘5’ to the addition problem as queried. He thinks my response is no better than a stab in the dark. Does the suggested reply advance matters? How does it justify my choice of ‘125’? What it says is: “ ‘125’ is the response you are disposed to give, and (perhaps the reply adds) it would also have been your response in the past.” Well and good, I know that ‘125’ is the response I am disposed to give (I am actually giving it!), and maybe it is helpful to be told—as a matter of brute fact—that I would have given the same response in the past. How does any of this indicate that— now or in the past—‘125’ was an answer justified in terms of instructions I gave myself, rather than a mere jack-in-the-box unjustified and arbitrary response? (1982, 23) The emphasis here on justification echoes the way in which Kripke initially sets up the skeptical problem. In the discussion below, the challenge posed by the skeptic takes two forms. First, he questions whether there is a fact that I meant plus, not quus, that will answer his skeptical challenge. Second, he questions whether I have any reason to be so confident that now I should answer ‘125’ rather than ‘5.’ The two forms of the challenge are related. I am confident that I should answer ‘125’ because I am confident that this answer also accords with what I meant. Neither the accuracy of my computation nor of my memory is under dispute. So it ought to be agreed that if I meant plus, then unless I wish to change my usage, I am justified in answering (indeed compelled to answer) ‘125’ not ‘5.’ An answer to the skeptic must satisfy two conditions. First, it must give an account of what fact it is (about my mental state) that constitutes my meaning plus, not quus. But further, there is a condition that any putative candidate for such a fact must satisfy. It must, in some sense, show how I am justified in giving the answer ‘125’ to ‘68 + 57.’ (1982, 11)

Skepticism about Meaning • 393

It is important to understand what must be presupposed in order for this requirement to make sense. The reason one ought to answer ‘125’ is that, if one did mean addition by ‘+’, then ‘125’ is the correct answer to the question ‘What is 68 + 57?’; and the correct answer is the answer one ought to give. The reasoning here is as follows: If knowing that F allows us to conclude that in the past one meant addition by ‘+’, then it allows us to conclude that the sentence ‘68 + 57 = 125’ means that the sum of 68 and 57 is 125, and hence is true iff the sum of 68 and 57 is 125, provided, of course, that one is now using ‘+’ with the same meaning as before.11 Next we appeal to the arithmetical fact that 125 is the sum of 68 and 57, which the skeptic grants, and is not in dispute.12 From these two facts, one about meaning and one about numbers, it follows that the sentence ‘68 + 57 = 125’ is true, and hence that one speaks the truth if one responds ‘125’ to the question ‘What is 68 + 57 ?’. Since the skeptic’s requirement clearly presupposes that one ought to speak the truth in this case, it follows that one ought to answer ‘125’. In this way knowledge of the putative meaning-determining fact that F, together with undisputed knowledge of nonlinguistic facts and the general presumption that one ought to speak the truth, provide a basis for concluding what answer one ought to give to the question ‘What is 68 + 57 ?’. Applying the normativity requirement (NE) to the dispositional view yields the first premise, (P1), of an obvious normativity argument. P1. If one’s past dispositions to answer questions What is n + m? determined that one meant addition by ‘+’, then knowing one’s past dispositions to answer such questions would, in principle, provide one with a sufficient basis for concluding that one ought to give the answer ‘125’ to the question ‘What is 68 +57?’, provided one intends to use ‘+’ with the same meaning it had in the past. 11 I am here taking the meanings of the other symbols—the numerals and identity sign as given. This simplification does not affect the point at hand. 12 Kripke insists, “Neither the accuracy of my computation nor of my memory is under dispute. So it ought to be agreed that if I meant plus, then unless I wish to change my usage, I am justified in answering . . . ‘125,’ and not ‘5’ ” (1982, 11). Similarly, in explaining why he sets up the paradox as a metalinguistic problem about the past use of an expression, Kripke says that when this is not done, “some listeners hear it as a skeptical problem about arithmetic: ‘How do I know that 68 + 57 is 125?’ (Why not answer this question with a mathematical proof?) At least at this stage, skepticism about arithmetic should not be taken to be in question: we may assume, if we wish, that 68 + 57 is 125” (1982, 13). I take these passages to indicate that in Kripke’s formulation of the paradox, my present (true) belief that 68 + 57 = 125 can be appealed to (without further justification) in attempting to meet the skeptic’s demand that I show how I am justified in thinking that ‘125’ is the answer required by my past understanding of the term.

394 • Essay Fifteen

The normativity objection to the dispositional analysis consists in (P2). P2. Knowing one’s past dispositions to answer questions What is n + m? does not, in principle, provide one with a sufficient basis for concluding what answer one ought to give to any such question, provided one intends no change in meaning. Together, (P1) and (P2) entail that no matter how closely my past dispositions to answer questions of the relevant form may, in fact, have corresponded to the results of applying the addition function, those dispositions did not determine that in the past I meant addition by ‘+’. Thus, if (P1) and (P2) are correct, then the dispositional response to the skeptic is incorrect. Much of Kripke’s discussion is aimed at supporting (P2), which is itself quite plausible. Suppose I had amnesia and didn’t remember what I meant by ‘+’ in the past. When asked, at present, to calculate ‘68 + 57’ I might feel inclined to give the answer ‘125’. Still, I might wonder whether this answer was justified, in terms of what I meant by ‘+’ in the past. Would it be sufficient to be told that this was the answer I was disposed to give in the past? It doesn’t seem that it would, for I might wonder whether in the past this was one of the calculations on which I was disposed to make mistakes. A similar observation holds regarding the point of view of someone viewing me from the outside and wondering whether my present answer of ‘125’ is correct. Simply being given a specification of the relevant past dispositions is not enough to allow such an observer to conclude that my present answer accords with my past meaning or intentions. Since (P2) is true, and since (P1) and (P2) entail that the dispositional response to the skeptic is false, the success of the argument against the dispositional view depends on our evaluation of the Normativity Requirement (NE), and the particular instance of it given in (P1). How plausible are these principles? Quite plausible, it seems, provided that we understand what it is for one empirical truth to determine another in accord with the following constraint. DetE. P determines Q only if, given P, one can demonstrate Q without appealing to any other empirical facts—i.e., only if Q is an a priori consequence of P. With this understanding of determination, (NE) and (P1) will hold.13 If the claim that in the past one meant addition by ‘+’ were an a priori consequence of the relevant dispositional truths, then knowing those 13 Is it necessary, in order for the Normativity Requirement (NE) to play its proper role in the skeptic’s argument, for the meaning-determining fact that F to provide the basis for an a priori, demonstrative inference to the relevant conclusion, or would it be enough for the fact that F to provide the basis for any sort of inference to the conclusion? Although the

Skepticism about Meaning • 395

dispositional truths would, in principle, make it possible to conclude that in the past one meant addition by ‘+’. But then, since it is accepted as uncontroversial that 125 is the result of adding 68 and 57, one could conclude that ‘125’ is the correct answer to the question ‘What is 68 + 57 ?’, and hence is the answer that one ought to give, provided one intends to use ‘+’ with the same meaning as it had in the past. It is worth emphasizing that on this account the substantive constraint placed on the meaning-determining fact F by the normativity requirement (NE) is that, in principle, knowledge of F must allow one to demonstrate that in the past ‘+’ was used to mean addition. This is essentially a descriptive matter. The normative prescription about what answer one ought to give in a particular case results from (i) the descriptive fact about what one meant, (ii) the arithmetical facts involving the addition function, and (iii) the normative presumption that one should speak the truth. The substantive requirement on F is just that knowledge of it be sufficient to demonstrate the first of these.14 The advantage of this way of construing the normativity objection is that the resulting argument against the dispositional analysis turns out to be point is arguable, there are reasons for opting for the stronger interpretation. First, Kripke’s skeptic seems to be looking for nonintentional facts that both metaphysically necessitate and epistemologically demonstrate the relevant meaning facts. Nonintentional facts that have meaning facts as a priori consequences would, presumably, do this; whereas nonintentional facts which provide only the basis for an empirical inference to the meaning facts presumably would not. Second, when we cast our net wider, and include nonintentional facts other than the dispositional facts presently under consideration, the skeptic’s claim that meaning facts are not a priori consequences of any such set of facts remains plausible, whereas a corresponding claim, to the effect that meaning facts cannot be inferred from nonintentional facts, even by an empirical inference to the best explanation, say, is not nearly so plausible. A different question about the normativity requirement is whether, as it is currently stated, it is too weak. Perhaps, in addition to being required to have meaning facts as a priori consequences, meaning-determining facts should also be required to be readily accessible to the speaker—so that the speaker is in a position to draw the relevant conclusions from those facts. I have no objection to this strengthening of the requirement; all the philosophical points I want to make regarding (NE) would hold for this strengthening of (NE). My reason for preferring the weaker version in the text is that it leads to a conclusion of greater generality. I am indebted to George Wilson for a discussion of this point. 14 The point here is a general one, and applies to versions of the rule-following paradox involving all sorts of different words. For example, consider a word like ‘fossil,’ which applies to concrete physical objects. Given some object o, the normative prescription regarding whether one should apply ‘fossil’ to o depends on (i) the meaning of the word ‘fossil,’ (ii) the nonlinguistic facts about the nature of o—i.e., whether or not it is a fossil and (iii) the general presumption that one should apply a term to an object only if doing so would involve speaking the truth. As Michael Thau has emphasized to me, examples like this highlight the need for (ii) over and above (i) particularly clearly. There is, however, another issue involving facts like (ii) that is worth mentioning, even though I don’t have sufficient space here to go into detail. Suppose I apply the word ‘fossil’

396 • Essay Fifteen

sound. However, there is a corresponding disadvantage—the conclusion thereby established is weak. Suppose that the proposition that one meant addition by ‘+’ really isn’t an a priori consequence of true propositions about one’s past dispositions to calculate. For all we know it might still be a necessary consequence of those propositions. And if it is, then there is an important sense in which what one means is determined by facts about one’s dispositions. Despite all that has been shown up to now, facts about meaning may yet turn out to be metaphysically necessary consequences of facts about dispositions, even if they are not a priori consequences of such facts. However, this is not all that Kripke has to say about the normativity objection to the dispositional analysis, as is indicated by his discussion of the following example. Assuming determinism, even if I mean to denote no number theoretic function in particular by the sign ‘*’, to the same extent as it is true for ‘+’, it is true here that for any two arguments m and n, there is a to a certain object o, and then I am challenged by a skeptic to justify this new application of the word. Suppose further that part of my response involves citing some fact F about my past use of the word. Imagine for the sake of argument that knowledge of F does allow me to demonstrate that in the past I used the word to mean fossil. Still, this by itself doesn’t justify my application of the word to o, even if I am now using the word with the same meaning as before, and it is granted that I ought in this case to speak the truth. To complete my justification I have to show that I am justified in thinking that o is a fossil. And how am I to do that? Perhaps the explanation of what it is for me to think that o is a fossil is simply for me to understand the word ‘fossil’ and be disposed to apply it to o on the basis of some reasonable, empirical examination of, or inquiry about, o. If so, then the justification for my thinking that o is a fossil may amount to little more than my now understanding the word ‘fossil’ and being disposed, after appropriate investigation and reflection, to apply it to o. If this is the situation, then in some ultimate sense I may have no justification for applying the term to a new item o, other than my own confident, informed, and reflective inclination to do so, plus my status as a competent user of the term. This does seem to be an important part of what Wittgenstein was trying to show in his deployment of the rulefollowing paradox, and it is also present in Kripke’s discussion. Moreover, there are, I think, cases in which something like this point is correct. However, this observation about justification does not undermine the response given in the text to Kripke’s version of the skeptical paradox; nor does it lead to any defensible skepticism about meaning. What it may do is point to an alternative route to one of Wittgenstein’s conclusions about language—a route that does not require any fundamental, skeptical recasting of our ordinary conception of meaning. The conclusion is that, at least in some cases, the explanation of a speaker’s understanding of a term T (including his ability to apply it to newly considered objects) does not involve associating T with an independently apprehended property P (and judging those objects to have P). Understanding a term, or using it to mean a certain thing, is not always like deciding to attach a new proper name to an object with which one is already familiar. For a brief discussion of this idea, see Soames (1989, 587–91), which is essay 5 of volume 1 of this collection. This idea is discussed in more detail in essay 16 in this volume. I am indebted to James Pryor and Michael Thau for helpful discussions of the material in this note.

Skepticism about Meaning • 397

uniquely determined answer p that I would give (I choose one at random, as we would normally say, but causally the answer is determined). The difference between this case and the case of the ‘+’ function is that in the former case, but not in the latter, my uniformly determined answer can properly be called ‘right’ or ‘wrong’ (24).15 Kripke’s idea can be made more graphic by imagining that as a result of some special brain surgery, I might be programmed to verbally respond to an utterance of the interrogative sentence ‘What is x * y?’ by uttering a definite numeral z for each pair of numerals x and y used in the question. For any pair of numerals you give me, using ‘*’, I will respond by uttering a definite numeral. Because of this we may presume that some definite function f on the natural numbers corresponds to the verbal responses I am disposed to give to interrogatives containing ‘*’. Despite this, it might still be that I don’t mean anything by ‘*’. My responses may be nothing more than verbal reflexes. I might notice that I always give determinate responses to queries using ‘*’. But I don’t regard these as correct or incorrect—since I don’t intend to assign any meaning to ‘*’. The point here is that, by themselves, mere verbal dispositions to respond to interrogative sentences involving a symbol, either ‘+’ or ‘*’, are not enough to endow it with meaning—even if the dispositions can be seen as corresponding to a definite function. Since it is possible to have these dispositions without meaning the function, meaning the function is not a necessary consequence of having the specified dispositions. Some other element must be present in order for us to have genuine meaning. So far, we haven’t found it. It should be noticed that in giving this argument, we have quietly replaced Kripke’s original formulation (D) of the dispositional analysis with some nonintentional understanding of it. In the original formulation, the allegedly meaning-determining dispositions were characterized as dispositions to answer arithmetical questions, presumably by asserting that a particular number is the value of the relevant function at a given pair of arguments. By contrast, in Kripke’s ‘*’ example I don’t mean anything by ‘*’; I do not assert anything or answer any question. Here the dispositions— both with ‘*’ and with ‘+’—are characterized simply as dispositions to utter specific numerals when prompted by utterances of particular interrogative sentences containing ‘*’ or ‘+’. When dispositions are characterized in this way, Kripke’s discussion of the ‘*’ example can be seen as constituting a second normativity argument against the dispositional analysis. Like the first, it is based on a general requirement, in this case (NM). 15

See also Kripke’s discussion (1982, nn. 18, 24).

398 • Essay Fifteen

NM. If the fact that F determined that (in the past) one meant addition by ‘+’, then in any possible world in which it is the case that F, one means addition by ‘+’; hence, in any possible world w in which it is the case that F, ‘125’ is the answer one ought to give to the question What is 68 + 57?. Applied to the nonintentional version of the dispositional analysis, this general requirement yields (P1′). P1′. If a certain complex dispositional fact—namely the fact that in the past one was disposed to verbally respond in such and such ways to utterances of interrogative sentences What is n + m?— determines that (in the past) one meant addition by ‘+’, then in any possible world w in which that dispositional fact obtains, one means addition by ‘+’ in w, and hence ought to answer ‘125’ to the question ‘What is 68 + 57 ?’. The second normativity objection to the dispositional analysis (formulated nonintentionally) is (P2′). P2′. It is possible not to mean addition (or anything at all) by ‘+’ even though one is disposed to verbally respond in such and such ways (the ways mentioned in (P1′)) to utterances of interrogative sentences What is n + m? From (P1′) and (P2′) it follows that the fact that in the past one was disposed to verbally respond in the ways mentioned in (P1′) to utterances of interrogative sentences What is n + m? does not determine that (in the past) one meant addition by ‘+’. The argument is clearly valid. Moreover, the discussion of the ‘*’ example shows that (P2′) is true, provided that the dispositions spoken of are nonintentionally stated and restricted simply to the production of numerals in response to utterances of interrogative sentences. (P1′) is also true, provided that the determination relation it speaks of is constrained by the following principle. DetM. P determines Q only if Q is a necessary consequence of P We have, then, two versions of the normativity objection. One version, the necessary-consequence version, incorporates (DetM); the other version, the a priori–consequence version, incorporates the earlier requirement, (DetE), on the determination relation. Both versions are sound, and establish their conclusions, when applied to the nonintentional form of the dispositional analysis. However, recognition of the difference between the two gives rise to an important question. If, in addition to the verbal dispositions already considered, we cast our net wider, so as to include more and more nonintentional facts, will the two versions of the normativity objection continue to apply together, or will they come apart at some point?

Skepticism about Meaning • 399

In order to investigate this question, we let T be the set of nonintentional truths about my dispositions to produce numerals in response to questions What is n + m?. In our discussion of the ‘*’ example, we have seen that it is possible for all members of T to be true even though I don’t mean anything by ‘+’, because my responses are not real answers but only verbal reflexes. Are there other nonintentional truths which, together with T, would change this result? Suppose we formed a new set, T′, which included all members of T together with all additional nonintentional truths about my dispositions to verbal behavior involving ‘+’— for example, dispositions covering cases in which I ‘check and revise’ my work, dispositions to insist on one and only one ‘answer’ for any given question, dispositions to strive for agreement between my own answers and those of others, and so on. Would it be possible for all members of this larger set T′ to be true, even though I didn’t mean anything by ‘+’? I should emphasize that I am not here trying to identify all the relevant verbal dispositions; nor am I trying to characterize them precisely in a thoroughly nonintentional way. Rather I am assuming that some such characterization is possible, and I am asking whether, by taking a broad enough look at such a characterization of our linguistic behavior and dispositions, we would at some point reach a stage in which we were confident that anyone conforming to all those facts meant something by his words? Is there a possible world in which all one’s nonintentionally characterized dispositions to verbal behavior, including those just mentioned, match my dispositions in the actual world, yet one does not mean addition, or anything else, by ‘+’? If so, how is this possible? Would the result change if we enlarged the set of potential meaning-determining truths still further to include not only all truths about my dispositions to verbal behavior, but also all truths about (i) the internal physical states of my brain, (ii) my causal and historical relationships to things in my environment, (iii) my (nonintentionally characterized) interactions with other members of my linguistic community, (iv) their dispositions to verbal behavior, and so on? Is there a possible world in which someone conforms to all those facts—precisely the facts that characterize me in the actual world—and yet that person does not mean anything by ‘+’? I think not. Given my conviction that in the past I did mean addition by ‘+’, and given also my conviction that if there are intentional facts, then they don’t float free of everything else, I am confident that there is no such a world. Although I cannot identify the smallest set of nonintentional facts about me in the actual world on which meaning facts supervene, I am confident that they do supervene. Why shouldn’t I be? I start out convinced that in the past I meant addition by ‘+’. Next, I imagine some large set of nonintentional facts about the past remaining fixed, so that they obtain in some possible world w. I then ask whether the individual who

400 • Essay Fifteen

corresponds to me in w—the individual who is the focus of the fixed nonintentional facts in w—means something by ‘+’. The answer, of course, is that since I meant something in the actual world, he means something in w. And if he means something by ‘+’, surely he means what I meant by ‘+’ in the actual world—addition. Of course, this won’t convince the skeptic. He isn’t willing to grant that I meant anything in the past. He wants to be shown that there is a nonempty set of meaning facts about which the question of supervenience can be raised; and he won’t be satisfied until I specify a set of nonintentional truths from which the existence of such meaning facts can be demonstrated. In effect, he demands that I establish that purported truths about what I meant are a priori consequences of nonintentional truths the existence of which we both accept. But why should I acquiesce in this demand? It would be very interesting if it turned out that I could refute the skeptic using only uncontroversial premises that even he accepts. However, this is not generally required of responses to skeptical problems, and I see no special reason to require it in this case. What is required is that I rebut any skeptical challenge that purports to show that our ordinary, pretheoretic beliefs about meaning conflict with other fundamental beliefs in a way that makes our beliefs about meaning unsustainable. In general, the power of skeptical views lies not in their rejection of what we commonly take for granted, but in their finding reasons for such rejection that arise directly from other fundamental convictions of ours. If the skeptic can find such reasons, he can show that our own most deeply held views are sufficient to undermine the beliefs (about meaning) that he challenges. Has Kripke’s skeptic about meaning done this? I don’t think so. He has insisted that if I meant anything in the past, then what I meant must be determined by nonintentional facts; and I have agreed, provided that the determination relation is one of necessary consequence. I grant that if I meant anything in the past, then what I meant must be a necessary consequence of nonintentional truths about me, my environment, my community, and so on. But it is not evident that there is a problem here, since none of the skeptic’s arguments show that such a relation fails to hold. Indeed, they scarcely even attempt to show this. Instead, they try to establish that no collection of nonintentional truths will allow us to demonstrate the truth of the relevant intentional claims. This, I have suggested, is tantamount to an attempt to convince us that claims about meaning (and propositional attitudes) are not a priori consequences of any set of nonintentional truths. On this point it must be admitted that the skeptic has a strong case. The task of deriving claims about meanings (and propositional attitudes) from nonintentional truths, together perhaps with a priori definitions of intentional notions like meaning and belief, is daunting. I don’t know how to

Skepticism about Meaning • 401

give such a derivation, and I am not sure that any is possible.16 Consequently, I am willing to grant that the skeptic might be right in maintaining that claims about what I meant are not a priori consequences of nonintentional truths.17 If it were clear that any necessary consequence of a set of claims P was also an a priori consequence of P, then this admission would provide the skeptic with just what he needs; for then he could force me to admit that claims about meaning may not be necessary consequences of nonintentional truths. That would conflict with my conviction that meaning facts must supervene on nonintentional facts, and so would threaten my pretheoretic commitment to meaning facts. However, this argumentative strategy fails. Thanks to the work of Kripke and others, it has become clear that many necessary consequences of propositions are not a priori consequences of them. Consequently, my admission that claims about meanings may not be a priori consequences of nonintentional truths need not undermine my belief that they are necessary consequences of those truths. Could the skeptic undermine my claims about past meanings by appealing directly to the thesis that if there are facts about what I meant in the past, then they must be a priori consequences of nonintentional truths? I don’t think so. That thesis is not one of my beliefs about meaning, pretheoretic or otherwise. Since the skeptic has done nothing to establish it in its own right, his challenge fails. To sum up, the general form of the skeptic’s argument is the following: P1. If in the past there was a fact about what I meant by ‘+’, in particular, if there was a fact that I meant addition by ‘+’, then either: (i) this fact was determined by nonintentional facts of such and such kinds—facts about my past calculations using ‘+’, the rules or algorithms I followed in doing calculations involving ‘+’, my past dispositions to respond to questions What is n + m?, the totality of my past dispositions to verbal behavior involving ‘+’, etc. or (ii)

the fact that I meant addition by ‘+’ was a primitive fact, not determined by nonintentional facts.

16 One important reason why it is difficult to be certain on this point is that it is not completely clear what a priori definitions of intentional notions are possible. For a good discussion of this issue, and an argument for the conclusion that facts about meaning (and belief) are not a priori consequences of nonintentional facts, see Byrne (1994). 17 If the required epistemological relationship between the nonintentional truths and the claims about meaning were weakened to include empirical inferences to the best explanation, then the skeptical thesis that meaning claims are not epistemological consequences of nonintentional truths would be far more questionable. In any case, Kripke’s skeptic does not argue in this way. (See note 13 above.)

402 • Essay Fifteen

P2. Nonintentional facts of type (i) did not determine that I meant addition (or anything else) by ‘+’. P3. What I meant by ‘+’ was not a primitive fact. C1. Thus, in the past there was no fact that I meant addition (or anything else) by ‘+’. C2. By parity of reasoning, there never was a fact about what I, or anyone else, meant by any word; ditto for the present. In my view, the argument suffers from equivocation. If the determination relation is one of necessary consequence, then P3 is plausible, but P2 is not.18 If the determination relation is one of a priori consequence, then P2 is plausible, but P3 is not. What makes the argument seductive is the failure to distinguish these alternatives. Once this distinction is made, the argument loses much of its force. There is no interpretation in which P1, P2, and P3 are jointly true; nor do I know of one in which they are even jointly plausible.

Quine’s Indeterminacy Thesis I will now try to show that Quine’s arguments for the indeterminacy of translation (and inscrutability of reference) suffer from a similar defect. We may begin by noting the relationship between Kripke’s skeptical problem and Quine’s. Kripke sets up the challenge as that of specifying nonintentional facts which determine that what I mean now by ‘+’ is the same as what I meant by ‘+’ in the past. This can be rephrased as the challenge of specifying nonintentional facts which determine that a translation from ‘+’ as I used it in the past to ‘+’ as I use it now is correct, whereas a translation from ‘+’ as I used it in the past to, say, ‘quus’ as I use it now (to stand for the quaddition function defined in note 5) is incorrect. Quine’s question is a generalization of this challenge: What nonintentional facts determine that it is correct to translate an expression A, as used by a person p, or a community c, as meaning the same as an expression B, as used by a person p′, or community c′? Like Kripke’s skeptic, Quine answers that in a wide range of cases there are no meaning-determining facts. Quine’s position can be set out as a series of theses. The first of these is the Underdetermination of Translation by Data, which is an instance of his view that empirical theories of all sorts typically are underdetermined by observational evidence. 18

Given an appropriately broad listing of nonintentional facts in P1.

Skepticism about Meaning • 403

The Underdetermination of Translation by Data Let L1 and L2 be arbitrary languages, and let D be the set of all observational truths (known and unknown) relevant to translation from one to the other. For any theory of translation T for L1 and L2, compatible with D, there is a theory T′, incompatible with T, that is equally well supported by D.19 This thesis can be fleshed out by saying a little about what theories of translation are, and what Quine takes the evidence for them to be. First consider theories of translation. A theory of translation for two languages correlates individual words of each language with words or phrases of the other language; this correlation is used in the theory to correlate the sentences of the two languages.20 Any system of establishing such correlations can be counted as a translation manual, or theory. We may take such a theory as yielding (infinitely many) theorems of the form: Word or phrase w in L1 means the same as word or phrase w* in L2. Sentence S in L1 means the same as sentence S* in L2. According to Quine, the empirical data relevant to theories of translation are statements about the stimulus meanings of sentences. The stimulus meaning of a sentence S (for a speaker at a time) is a pair of classes— the class of stimulus situations in which the speaker would assent to S, and the class in which the speaker would dissent from S. Stimulus meanings are particularly important in evaluating translations of what Quine calls ‘occasion sentences’ and ‘observation sentences’. An occasion sentence is one assent to, or dissent from, which depends on the speaker’s current observation. An observation sentence is an occasion sentence for which assent or dissent depends only on observation—with no, or only a 19 The strength of this thesis, as well as the more general thesis of the underdetermination of empirical theories by observational data, depends on one’s conception of what it is for a class of data statements to support a theory. For present purposes I will follow what appears to be Quine’s latitudinarianism on this subject. Theories, together with auxiliary observational statements, make (entail) observational predictions. (Which statements count as observational for this purpose will not be an issue here.) A set of such observational predictions supports a theory to the extent that the theory, supplemented by true auxiliary observation statements, entails the members of the set. Two theories (appropriately supplemented with auxiliary observational statements) that entail the same members of the set, are equally well supported by the set. 20 I am assuming, in order to simplify the argument, that words are the minimal meaningbearing units, that languages contain finitely many such words, and that the translation of the infinitely many phrases and sentences of the two languages is the result of (i) the translation of the words that make them up plus (ii) combinatorial principles specifying the translations of syntactically complex expressions in terms of the translations of their parts.

404 • Essay Fifteen

minimum in the way of, background information required. For example, ‘He is a bachelor’ is an occasion sentence, since assent or dissent in a given case depends on whom the subject is observing; but it is not an observation sentence in Quine’s sense because whether or not one assents in a particular case depends in part on having special background information about the person perceived. ‘That is red’, on the other hand, does not depend (to the same extent) on having such background information, and so counts as an observation sentence for Quine. Quine’s way of approximating this intuitive notion of an observation sentence within his behavioristic framework is to define an observation sentence in a language L to be an occasion sentence for speakers of L the stimulus meaning of which varies little from one speaker to another. The stimulus meanings of observation and occasion sentences play a prominent role in Quine’s conception of the main observational predictions made by translation theories. These predictions are summarized by the following three principles for extracting testable claims from theories of translation (which otherwise wouldn’t entail any such predictions via their form alone):21 (i) if a translation theory states that an observation sentence S in L1 means the same as a sentence S* in L2, then the theory predicts that the two sentences have the same stimulus meanings in their respective linguistic communities; (ii) if S and S′ are occasion sentences of L1, and if a translation theory states both that S means the same in L1 as S* in L2 and that S' means the same in L1 as S*' in L2, then the theory predicts that S and S' have the same stimulus meaning in L1 iff S* and S*' have the same stimulus meaning in L2; (iii) if a theory translates an expression n of a language L as meaning the same as ‘not’ in English, then adding n to sentences of L must reverse stimulus meaning; similar claims are made regarding other truth-functional operators. On this conception, the observational data for theories of translation consist mainly of behavioral evidence regarding the stimulus meanings of occasion sentences. Quine also considers a possible constraint relevant to the translation of what he calls standing sentences—sentences assent to, or dissent from, which is independent of current sensory stimulation. The possible constraint is that sentences assented to (dissented from) in every situation by the community of speakers of L1 must be translated onto those assented to (dissented from) in every situation by the community of speakers of L2. But this constraint is itself problematic, and, in any case, adding it would not significantly change the overall picture. It is clear that if the observational data for theories of translation are restricted to behavioral evidence of the sort Quine has in mind, then profoundly different theories of 21

See Quine (1960, chap. 2).

Skepticism about Meaning • 405

translation will be supported equally well (in his sense) by all observational data, known and unknown, in virtually all interesting cases. For example, the set of all behavioral data concerning the stimulus meanings of sentences for me, past and present, is equally compatible with theories of translation which claim that (i) the term ‘rabbit’ as used by me in the past means the same as the term ‘rabbit’ as used by me now, (ii) the term ‘rabbit’ as used by me in the past means the same as the phrase ‘set of undetached rabbit parts’ as used by me now, or (iii) the term ‘rabbit’ as used by me in the past means the same as the phrase ‘temporal stage of a rabbit’ as used by me now. Since the expressions ‘rabbit’, ‘set of undetached rabbit parts’, and ‘temporal stage of a rabbit’, as used by me now mean, and refer to, different things, alternative translation theories that map these different expressions onto the term ‘rabbit’ as I used it in the past conflict with one another, and can be regarded as incompatible. Quine concludes from this that theories of translation are underdetermined, in the sense defined above, by the observational data for them.22 Since empirical theories of all sorts are typically underdetermined by the observational evidence bearing on them, the extension of this result to theories of translation is not, in itself, a radical development. What makes it striking is Quine’s use of it as the basis for the truly radical doctrine of the Indeterminacy of Translation. 22 When Quine speaks of different, incompatible theories (or ‘theory formulations’) all equally supported by the same possible observational evidence, he seems to have in mind logically incompatible theories (or theory formulations) (see 1970, 179; 1975, 322). However, despite the obvious difference in meanings between ‘rabbit,’ ‘set of undetached rabbit parts,’ and ‘temporal stage of a rabbit’ as I use them now, the following claims are not logically incompatible: (i) the term ‘rabbit’ as I used it in the past means the same as the term ‘rabbit’ as I use it now (ii) the term ‘rabbit’ as I used it in the past means the same as the phrase ‘set of undetached rabbit parts’ as I use it now, (iii) the term ‘rabbit’ as I used it in the past means the same as the phrase ‘temporal stage of a rabbit’ as I use it now. Consequently, translation theories making these different claims need not be logically incompatible with one another. Logical incompatibility will result if translation theories are embedded in larger background theories containing the following claims: (a) Rabbits are not sets of undetached rabbit parts & sets of undetached rabbit parts are not temporal stages of rabbits & rabbits are not temporal stages of rabbits; (b) ‘Rabbit’ (as I use it now) refers to an object iff it is a rabbit & ‘set of undetached rabbit parts’ (as I use it now) refers to an object iff it is a set of undetached rabbit parts & ‘temporal stage of a rabbit’ (as I use it now) refers to an object iff it is a temporal stage of a rabbit; (c) if two words refer to different things then they don’t mean the same; (d) if a means the same as b & a means the same as c, then b means the same as c. Let T1 be a translation theory containing statement (i), T2 be a translation theory containing statement (ii), and T3 be a translation theory containing statement (iii). The union of T1, T2, and a set containing (a)–(d) is logically inconsistent; as are corresponding unions involving the other relevant combinations. The justification for appealing to these auxiliary claims is that (a) states an obvious fact and (b)–(d) are axiomatic to any overall theory that makes significant use of the concepts of meaning and reference.

406 • Essay Fifteen

The Indeterminacy of Translation Translation is not determined by the set N of all truths of nature, known and unknown. For any pair of languages, there are incompatible theories of translation for those languages that accord equally well with N. All such theories are equally true to the facts; there is no objective matter of fact on which they disagree, and no objective sense in which one is true and the other is not.23 This doctrine is a consequence of two more basic Quinean views. Physicalism All genuine truths (facts) are determined by physical truths (facts). The Underdetermination of Translation by Physics Translation is not determined by the set of all physical truths (facts), known and unknown. For any pair of languages, there are incompatible theories of translation for those languages that accord equally well with all physical truths (facts). Quine’s route to the Indeterminacy of Translation is as follows: He begins with the behaviorist premise that since we learn language by observing the linguistic behavior of others, the only facts relevant to determining linguistic meaning must be publicly observable behavioral facts—in particular facts about stimulus meaning. But his discussion of the Underdetermination of Translation by Data shows that these facts do not determine which translations of our words are correct, and so do not determine what our words mean. It follows that no physical facts determine word meaning or correct translation. In the presence of the doctrine of Physicalism this means that claims about what our words mean—e.g., claims like ‘‘Rabbit’ as I used it in the past means the same as ‘rabbit’ as I use it now’—never state genuine facts, and never count as expressing genuine truths. Hence, the Indeterminacy of Translation. How persuasive is this argument? Many object, quite correctly, to its behaviorist premise, which allows Quine to move from the Underdetermination of Translation by Data to the Underdetermination of Translation by Physics.24 However, there is another, more fundamental, point to be made. The contents of Quine’s central claims are unclear. In particular, it is unclear precisely what determination relation is invoked in the three central 23

See Quine (1969b, 303). In this essay I will assume, without argument, that Quine’s behaviorism about language is false. It is noteworthy, however, that the appeal of the indeterminacy thesis, and the challenge posed by it, have been strongly felt by many philosophers who have not been prepared to accept behaviorism independently. My task here is to diagnose the source of that appeal, and defuse the challenge felt by those philosophers. 24

Skepticism about Meaning • 407

theses of Physicalism, the Underdetermination of Translation by Physics, and the Indeterminacy of Translation. Because of this, equivocation threatens. On certain construals of the determination relation Physicalism is plausible; on other construals the Underdetermination of Translation by Physics is plausible; but there is no construal on which both Physicalism and the Underdetermination of Translation by Physics are plausible, and no interpretation in which the Indeterminacy of Translation can be sustained. What is it for one set of claims to determine another? It is not for the claims in the second set to be logical consequences of the claims in the first. Certainly translation theories are not logical consequences of the set of all physical truths. But this is trivial, since whenever an empirical theory of any interest includes vocabulary not found in the truths of physics it will fail to be a logical consequence of those truths. For example, not all the truths of chemistry and biology are logical consequences of the set of true sentences of the language of an ideal physics. But chemistry and biology are supposed by Quine to be determined by physics; so the determination relation cannot be that of logical consequence. If it were, Physicalism would obviously be false. A different way of specifying the determination relation would be to say that a set P of statements determines a set Q of statements iff it would be (metaphysically) impossible for all the statements in P to be true without all the statements in Q being true—i.e., iff Q is a (metaphysically) necessary consequence of P. On this construal physicalism is quite plausible; it states that all genuine truths (facts) supervene on the physical truths (facts). However, the Underdetermination of Translation by Physics now turns out to be implausible. The point here parallels the one made in response to Kripke’s skeptical argument about meaning. Prior to any skeptical argument, Kripke’s or Quine’s, we naturally assume at the outset that there are facts about meaning and translation of the sort the skeptic denies. Given this conviction, we may ask directly whether a physically identical twin—someone (in a physically identical possible world) whose utterances, behavior, brain-states, causal and historical relations to the environment, and interactions with other speakers completely and exactly matches mine (in the actual world)—could mean by ‘rabbit’, what I mean (in the actual world) by, say, ‘set of undetached rabbit parts’. The natural answer to this question is ‘No’. Hence on this interpretation of the thesis of the Underdetermination of Translation by Physics, it should be rejected. The thesis could, of course, be saved if it could be shown both that (i) theories of translation are not a priori consequences of the set of physical truths, and that (ii) whenever a claim is not an a priori consequence of a set of statements it is not a necessary consequence of those statements. As we shall see, something like (i) is reasonably plausible. Thus if one were confused about the relationship between necessity and apriority,

408 • Essay Fifteen

or if one failed to distinguish them, one might wrongly conclude that both Physicalism and the Underdetermination of Translation by Physics are jointly true, when determination is taken to be necessary consequence. The error here is the implicit reliance on the false claim (ii)—that any necessary consequence of a set of statements is an a priori consequence of that set. Once this error is removed there is no plausible route to the theses of the Underdetermination of Translation by Physics and the Indeterminacy of Translation, on the interpretation in which determination is construed as necessary consequence. In light of this, it may be worthwhile to put the notion of necessity aside and examine a more epistemological conception of the determination relation. One might say that a set P of claims determines a set Q iff it is in principle possible, given the claims in P, for one to demonstrate the truth of the claims in Q, appealing only to logic and obvious a priori principles or definitions. The idea here is that for P to determine Q is for P to provide a theoretical basis for establishing Q that is absolutely conclusive, and that rules out any possibility of falsehood. In effect, determination is here construed as a priori consequence. On this construal, the thesis of the Underdetermination of Translation by Physics is both interesting and plausible. Given the total set of behavioral evidence about stimulus meanings of the sort Quine identifies, I cannot absolutely establish that what a speaker means by one of his terms is what I now mean by ‘rabbit’, as opposed to what I now mean by ‘set of undetached rabbit parts’ or ‘temporal stage of a rabbit’ (even if the speaker is me in the past). The claim that a speaker means one of these things rather than another is not an a priori consequence of the total set of claims about Quinean stimulus meanings. Moreover, it is hard to see how adding more behavioral facts (beyond Quinean facts about stimulus meaning), or more physical facts—about the neurological events in the person’s brain or his physical interactions with his environment—would, by itself, change the situation. Thus, it may turn out that theories of translation are epistemologically underdetermined by the set of all physical truths. But how serious would this be? Is there some reason to believe that all genuine truths must be not just necessary consequences, but also purely a priori consequences, of the set of all physical truths? As far as I can tell this is not one of our pretheoretic convictions; nor has Quine given a theoretical explanation of why it must be maintained.25 Hence, if theories of 25 It is important not to confuse what is a genuine pretheoretic conviction—namely that we do know what our words, and those of our neighbors, mean—with what is not such a conviction—namely that we arrive at this knowledge by deriving true claims about the meanings of our words, and those of our neighbors, as a priori consequences of purely physical truths. Whereas the former claim is clearly true, the latter is almost certainly false.

Skepticism about Meaning • 409

translation do turn out to be underdetermined by physics in this epistemological sense, such a result should be taken to show that the corresponding epistemic version of Physicalism is false, even though the metaphysical version of Physicalism, in which the determination relation is that of necessary consequence, remains true. Certainly, such a position is preferable to the radical and paradoxical rejection of facts about meaning and reference implied by Quine’s Indeterminacy thesis.26 If this is right, there may be an interesting epistemological construal of the determination relation according to which the Underdetermination of Translation by Physics is true, while Physicalism and the Indeterminacy of Translation are false. However, this point must be qualified by a complication that I have neglected up to now. In order to derive any empirical theory T from the set of truths of physics, one must appeal to theoretical identifications, or bridge principles, relating the vocabulary of T to the vocabulary of the underlying physics. What conditions must these identifications or bridge principles satisfy in order to be available for such derivations? Is it enough that they be true, or must they also be a priori (or necessary)? In my epistemic characterization of the determination relation, I allowed only obvious a priori truths and definitions. With this understanding the Underdetermination of Translation by Physics is at least plausible (though it has not, of course, been demonstrated). However, with this same understanding, Physicalism is threatened, not just by theories of meaning or translation, but by ordinary instances of theoretical reduction, such as the reduction of the biological concept of a gene to a physical construction involving the concept DNA. The relevant theoretical identity statement relating the two seems to be an empirical, a 26 Quine’s doctrines of Physicalism, the Indeterminacy of Translation, and the Inscrutability of Reference entail that claims of the sort, Person P’s word w refers to rabbits (as opposed to sets of undetached rabbit parts, etc.), are not determined by the totality of physical facts and so do not express genuine truths; ditto for any open formula P’s word w refers to x relative to an assignment of an object as value of the variable ‘x’. Given Quine’s usual understanding of the existential quantifier, one can conclude from this that ∃x P’s word w refers to x never expresses a genuine truth. Supposing that it is nevertheless meaningful, we may conclude that ~∃x P’s word w refers to x will always be true; in effect, no one’s words ever refer to anything. This is eliminativism about reference (as ordinarily understood). Pretty paradoxical, especially for someone who clearly is attempting to use words to refer to, and make claims about, things. Quine nowhere explicitly acknowledges such starkly paradoxical consequences of his views. The closest he comes is in the essay ‘Ontological Relativity” (1969a, 47–51). There he notes the paradoxical consequences of applying the Indeterminacy and Inscrutability doctrines to our (present) selves, and attempts (unsuccessfully) to avoid these consequences by invoking his mysterious doctrine of reference relative to a coordinate system (48) and “acquiescing in our mother tongue” (49). Unfortunately, space limitations prevent me from providing a thorough discussion of these passages here. For discussion, see Soames (2003, vol. 2, chap. 11).

410 • Essay Fifteen

posteriori truth. Thus, if bridge principles relating the biological vocabulary to the physical vocabulary are restricted to a priori definitions, then our theoretical identity statement will be excluded, and the derivation of genetics from physical theory will be placed in jeopardy. Surely no one would conclude from this that genetics in particular, or biology in general, fails to state genuine truths; rather, if a choice has to be made, we will reject the (strong) epistemological version of Physicalism.27 In light of this, we could opt to weaken the epistemological determination relation so as to allow P to determine Q provided that Q is a logical consequence of P together with bridge principles, or theoretical identities, consisting of any truths—contingent, a posteriori, or otherwise—that relate the vocabulary of P to that of Q.28 On this weakened epistemological 27 This result could be avoided if it could be shown that there are genuinely a priori semantic definitions of ‘gene’ and ‘DNA’ from which, together with the set of purely physical truths, the theoretical identification of genes with DNA can be derived. Since I am not certain whether this is possible in principle, I am not certain that the case of genetics provides a genuine counterexample to the strong epistemological version of physicalism. By the same token, I am not certain that no genuinely a priori semantic definitions of notions like meaning and reference exist from which, together with all physical truths, claims about meaning and reference can be derived. 28 This conception of determination is closely related to familiar conceptions of theoretical reduction, which are used by Michael Friedman (1975) to characterize Quine’s theses of Physicalism and the Indeterminacy of Translation. There Friedman recognizes two kinds of reduction, strong and weak. Strong reduction is reduction in the classical sense. A theory T2 is classically reducible to a theory T1 iff the theorems of T2 are logical consequences of T1 together with a set D containing a ‘definition’ for each primitive predicate of T2. A ‘definition’ is a universally quantified biconditional establishing the extensional equivalence of an n-place primitive predicate of T2 with a corresponding formula of arbitrary complexity of the language of T1. The same point can be expressed in another way by noting that the ‘definitions’ appealed to in a reduction can be taken as establishing a mapping D from primitive predicates of T2 onto coextensive open formulas of T1. Given this, we may define the notion of an n-place (primitive) predicate P of T2 being satisfied by an n-tuple in an arbitrary model M relative to a mapping D as consisting in the image of P under D being satisfied by that n-tuple in M. Classical (strong) reduction obtains when there is a mapping D such that every model of T1 is a model-relative-to D of T2 (see Friedman 1975, 357–58). Weak reduction is just like strong reduction except that the mapping D associates each primitive predicate of T2 with a set of corresponding open formulas of the language of T1. The set of formulas D associates with each primitive predicate P must be coextensive with P—i.e., as a matter of fact an n-tuple will satisfy P iff it satisfies at least one formula in the image of P under D. The notion of an n-place (primitive) predicate P of T2 being satisfied by an n-tuple in an arbitrary model M relative to such a mapping D is then defined as consisting in there being at least one member of the set of formulas associated with P by D being satisfied by that n-tuple in M. As before, reduction obtains when there is a mapping D of this sort such that every model of T1 is a model-relative-to D of T2. (Weak reduction differs from strong reduction only in cases in which the sets associated with the primitive predicates of T2 are infinite.)

Skepticism about Meaning • 411

conception the usual reduction of the notion of a gene to DNA will pose no difficulty for the thesis of Physicalism. However, on this weakened conception, even theories of translation will be determined by the set of physical truths. To see this let Sx be some formula specifying a set of physical facts satisfied by me and only me—so that Sx is true relative to an assignment of me as value of ‘x’, and is false relative to other assignments. It doesn’t matter what this formula looks like—whether it is complex, whether we can identify it as applying just to me, etc. It only matters that it exists. For completeness we may suppose that it includes a specification of all physical facts about me relevant to my use of language—facts about my present and past brain-states, facts about my behavioral dispositions, facts about my physical interactions with the environment, including causal and historical relations connecting my uses of individual words with other speakers and things, and so on. Let Lx be a similar formula applying uniquely to a certain Spanish speaker, Luisa. (For the sake of simplicity, suppose that I am a monolingual speaker of English and Luisa is a monolingual speaker of Spanish.) Now imagine a claim of the following sort, exhaustively listing Friedman’s stated reason (358) for allowing weak reduction to count as a genuine type of theoretical reduction is to make room for positions such as functionalist theories of mind which identify each token of a mental type with a particular physical realization, while recognizing arbitrarily many different ways in which the given type might be physically realized. Note, however, the modal notion here. Its use in characterizing the relevant functionalist theories points up a modest puzzle having to do with Friedman’s position. Reduction, as he officially characterizes it, does not require the ‘definitional’ mapping D to pair the predicates of T2 with formulas, or sets of formulas, that are intensionally equivalent to them in any sense. In particular D is not required to produce pairs that are extensionally equivalent in arbitrary counterfactual, or a priori imaginable, circumstances. Because of this the different merely possible, or merely imaginable, ways in which a mental type might be physically realized are, strictly speaking, irrelevant to the existence of ‘definitional’ mappings D satisfying Friedman’s stated conditions for reduction. Since, as far as I know, physicalist functionalists never maintain that there actually exist infinitely many physically different kinds of realizations of a given mental type, they presumably ought to be reasonably confident in asserting the strong reducibility (in Friedman’s official sense) of their theories to physics. Why then is there a need for the notion of weak reducibility? Does Friedman’s use of the notion reflect an implicit desire to require the ‘definitional’ mappings in genuine reductions to provide more than actual coextensiveness? Do they also have to provide coextensiveness in all counterfactual (or in all a priori imaginable) situations as well? If so, then couldn’t we define the determination relations needed to evaluate Quine’s theses directly in terms of necessary, or a priori, consequence, as above? Putting these and other subsidiary issues aside, I would like to acknowledge the essential correctness of some of Friedman’s central points. In particular, he makes a plausible case for interpreting Quine’s thesis of the Indeterminacy of Translation as the doctrine that theories of translation are not reducible (in either his strong or his weak sense) to the set of physical truths. He then argues for the correct (but understated) conclusion that Quine has given no compelling argument for the Indeterminacy Thesis, understood in this way.

412 • Essay Fifteen

the translation of the finitely many individual words of my language into words and phrases of Luisa’s language.29 TSL. ∃x ∃y [Sx & Ly & for all words w (of English) and words or phrases w* (of Spanish), w as used by x means the same as w* as used by y iff (i) w = ‘woman’ and w* = ‘mujer’, or (ii) w = ‘headache’ and w* = ‘dolor de cabeza’, or . . . A corresponding claim lists the translation of the finitely many individual words of Luisa’s language onto words and phrases of my language; similar claims may be imagined for each actual pair of language users— past, present, and future. Since it is extremely plausible to suppose that there are only finitely many such pairs of speakers, it seems safe to assume that some extremely long and complicated general formula of the following sort exists which encompasses all the individual cases. GT. For all speakers x and y, and for all words w in x’s language and words or phrases w* in y’s language, w as used by x means the same as w* as used by y iff (i) Sx & Ly & w = ‘woman’ & w* = ‘mujer’, or w = ‘headache’ & w* = ‘dolor de cabeza’, or . . . ; or (ii) Lx & Sy & (w = ‘semaforo’ & w* = ‘traffic light’, or . . . ; or (iii) Sx & Gy & . . . ; or . . . GT may be regarded as a bridge principle providing a coextensive physicalistic counterpart of the predicate ‘means the same as’ applied to words and phrases in theories of translation. Next we need a general bridge principle that encompasses the combinatorial rules used in theories of translation to combine translations of parts into translations of whole sentences. Presumably there are only finitely many such combinatorial rules for each pair of speakers. If, as we are assuming, there are only finitely many such pairs of actual speakers, these rules can in principle be exhaustively listed. This list together with GT can then be used to formulate another bridge principle that provides a coextensive physicalistic counterpart to the predicate ‘means the same as’ as used between sentences in theories of translation. But then, theories of translation will be derivable from the set of all physical truths together with these bridge principles, and so will count as determined by physics in our weakened epistemological sense. It might, of course, be observed that this trivialization just shows that stronger conditions are needed to characterize the required determination relation. Two possibilities suggest themselves. First, it might be claimed 29 In this discussion I ignore certain practical complications such as the fact that some speakers speak more than one language, the fact that words of the language may be ambiguous, and the possibility that sometimes there may be no translation of a word in one language onto a word or phrase of the other language. Although these are real factors in translation, they are peripheral to Quine’s philosophical claims about translation.

Skepticism about Meaning • 413

that what we want is not a reduction of theories of translation to physical truths, but rather a single reduction to the set of physical truths of all theories making use of semantic notions such as meaning and reference. Surely, if these notions are legitimate they will have significant theoretical uses well beyond theories of translation in the narrow sense considered here. The fact that a trivial reduction is theoretically possible when translation theories are considered in isolation does not show that such a reduction is possible in a context which is properly more inclusive. The point is well taken. But the problem with this suggestion is that it takes us far beyond Quine’s own discussion, and into uncharted waters. Before we can make any progress on the question of whether a single physicalistic reduction of all legitimate uses of semantic notions is theoretically possible, we need a reasonably precise and exhaustive characterization of the range of theoretical uses of semantic notions. Until we have this, we can’t evaluate the case for skepticism about meaning (and reference) because no sufficiently articulated case for skepticism about these notions has even been made. The second possible response to the trivial reduction sketched above is to stick to a physicalistic reduction simply of translation theories, but to claim that what is wanted are bridge principles that provide definitions yielding physical formulas coextensive with their nonphysical counterparts in all counterfactual circumstances. The bridge principles in the trivial reduction do not satisfy this demand, since, for example, someone could have meant the same by a particular term as I do by ‘woman’ even though that person did not satisfy the physical predicate identifying any actually existing individual. I have no quarrel with this strengthened, modal constraint on the determination relation. However, two points should be noted. First, if, unlike Quine, we grant the legitimacy of modal notions, then we can characterize determination directly in terms of necessary (or counterfactual) consequence. But then, as I have maintained above, we have reason to believe that theories of translation are determined by the physical truths. Second, if one insists on characterizing determination in terms of a strengthened reduction relation that requires physicalistic formulas which are necessarily (or counterfactually) coextensive with the predicates and other vocabulary items used in the theories undergoing reduction, then it is not clear what our attitude should be towards the resulting strengthened versions of Physicalism and the Underdetermination of Translation by Physics. On this interpretation the latter thesis requires the existence of physical formulas (or sets of formulas)30 necessarily (or counterfactually) coextensive with the 30 On this interpretation Friedman’s relation of weak reduction, strengthened by the requirement that the mapping from predicates in T2 to sets of formulas of the language of T1 produce

414 • Essay Fifteen

predicates ‘expression E as used by x means the same as expression E' as used by y’ and ‘sentence S as used by x means the same as sentence S' as used by y’. Are there such formulas (or sets of formulas)? There may well be, though it is hard to say for sure. Certainly Quine has given no compelling arguments to the contrary. Thus, on this interpretation, the case for the Underdetermination of Translation by Physics has not been made. A similar point can be made regarding Quine’s other premise—namely Physicalism—in the argument for Indeterminacy. On the present strengthened interpretation of the determination relation, it is not evident what we should think of the resulting version of Physicalism. I have already granted the truth of a weaker version of Physicalism, which states that all genuine truths must be necessary consequences of the physical truths. Is it obvious that we should add the further requirement that they be logical consequences of the Physical truths plus definitions that provide necessarily (or counterfactually) coextensive physical translations of all vocabulary items used in genuine truths? Since this point hasn’t been established, the present interpretation of the determination relation is one in which neither of the two premises for the Indeterminacy Thesis has been secured. In light of this, the conclusion to be drawn is that Quine’s argument for the Indeterminacy of Translation, like Kripke’s skeptical argument about meaning and following a rule, fails to provide a compelling challenge to our pretheoretic convictions about meaning and translation. This does not foreclose the possibility that something could be done to strengthen, or revive such a skeptical challenge. But surely the burden of proof is on those who wish to persuade us to adopt a radically skeptical attitude toward our ordinary semantic notions. Although I do not believe that any skeptical challenge of this sort could succeed, that has not been the burden of my argument. Instead, I have tried to defuse the particular skeptical arguments of Kripke and Quine by showing that their initial power is due in substantial part to an equivocation about what it is for one set of claims to determine another. Once this equivocation is removed, the original skeptical arguments lose their force, and it becomes highly dubi-

pairs that are necessarily (or counterfactually) coextensive (in the sense of note 28), should count as a genuine instance of determination in the sense presently under consideration. 31 The ideas in this essays date back to seminars I gave in the summer of 1988 at the University of Washington, and in the fall of 1988 at Princeton University. These ideas were refined and elaborated in my fall semester seminar of 1993 at Princeton. An early version of the first section of this essay, on Kripke’s account of the rule-following paradox, was presented in June of 1995 at the Instituto de Investigaciones Filosóficas de la Universidad Nacional Autónoma de México. Later versions of the essay were presented at the Department of Philosophy, Wayne State University on October 13, 1995, the Chapel Hill Philosophy Colloquium on October 29, 1995, the Division of Humanities at the California Institute of

Skepticism about Meaning • 415

ous and problematic that any nonequivocal replacements could be found that would provide good reasons for a thoroughgoing skepticism.31

References Byrne, Alex. 1994. “The Emergent Mind.” Ph.D. diss., Princeton University. Friedman, Michael. 1975. “Physicalism and the Indeterminacy of Translation.” Noûs 9:353–73. Kripke, Saul A. 1982. Wittgenstein on Rules and Private Language. Cambridge: Harvard University Press. Quine, W. V. 1960. Word and Object. Cambridge: MIT Press. ———. 1969a. “Ontological Relativity.” In Ontological Relativity, 26–68. New York: Columbia University Press. ———. 1969b. “Reply to Chomsky.” In Words and Objections: Essays on the Work of W. V. Quine, ed. Donald Davidson and Jaakko Hintikka, 302–11. Dordrecht: Reidel. ———. 1970. “On the Reasons for Indeterminacy of Translation.” Journal of Philosophy 67:178–83. ———. 1975. “On Empirically Equivalent Systems of the World.” Erkenntnis 9:313–28. Soames, Scott. 1989. “Semantic Competence.” Philosophical Perspectives 3:575–96. ———. 2003. Philosophical Analysis in the Twentieth Century. 2 vols. Princeton: Princeton University Press.

Technology on November 3, 1995, the Instituto de Investigaciones Filosóficas de la Universidad Nacional Autónoma de México on January 26 and 29, 1996, the Department of Philosophy at Rutgers University on March 7, 1996, the Department of Philosophy at Reed College on March 20, 1996, and the Department of Philosophy at the University of California at Berkeley on April 25, 1996. It was also presented at my fall semester seminar in 1996 at Princeton. I would like to thank the participants at all those presentations, and seminars, for their comments. In addition, I would like to thank James Pryor, Michael Thau, and George Wilson for reading and commenting on earlier drafts of the essay.

ESSAY SIXTEEN

Facts, Truth Conditions, and the Skeptical Solution to the Rule-Following Paradox

Introduction In chapter 2 of Wittgenstein: On Rules and Private Language,1 Saul Kripke presents a skeptical argument, inspired by Wittgenstein’s rulefollowing paradox, for the paradoxical conclusion that there are no facts about what we mean by our words. In chapter 3 Kripke outlines what he calls a “skeptical solution” to the paradox. What makes the position skeptical is that it purports to accept the startling conclusion that there are no facts about meaning; what makes it a solution is that it nevertheless defends the correctness of ordinary meaning ascriptions, such as the claim that I mean, and have meant, addition by ‘+’. On the face of it this is extremely puzzling. If I really have meant addition by ‘+’, then surely it is true that I have meant addition by ‘+’, in which case it is a fact that I have meant addition by ‘+’, and hence there is a fact about what I have meant by my words. This puzzle reflects a deep interpretive problem arising from an (acknowledged) unclarity in Kripke’s discussion regarding what conception of fact is involved in the skeptical argument in chapter 2, and the skeptical solution in chapter 3. I will argue that this unclarity reflects a fundamental tension in the text. On the one hand, there is a very natural and minimal conception of fact talk according to which the skeptical argument presents a powerful and troubling line of reasoning for a truly radical skeptical conclusion. However, on this interpretation the skeptical solution is utterly implausible, even incoherent, thus leaving a “straight solution”—of the sort I argue for in “Skepticism about Meaning: Indeterminacy, Normativity, and the Rule-Following Paradox”2—as the only appropriate response to the paradox. On the other hand, there is a different interpretation in which the facts maintained to be superfluous by the skeptical solution play a contentious explanatory role in theories of meaning and understanding. On this interpretation the skeptical solution 1 2

Kripke (1982). Essay 15 in this volume.

The Rule-Following Paradox • 417

has some important, though circumscribed, lessons to teach—lessons which can, to some extent, be established independently, as I will try to show. Unfortunately, this interpretation also faces a fundamental difficulty. If one adopts it, then it is hard to make sense of the original skeptical argument at all, let alone to see it as presenting any genuinely skeptical, or truly surprising conclusion. In this essay I will flesh out these two interpretations and try to determine to what extent the tension between them can be resolved. I will begin by reviewing the apparent structure of the skeptical argument and the basic strategy of the skeptical solution. I will then articulate two different versions of the skeptical solution, based on two different conceptions of the sorts of meaning facts they are taken to deny. After evaluating and drawing out the positive lessons to be learned from each version, I will close by taking up the question of whether it is possible to give a single coherent interpretation of Kripke’s text.

The Initial Skeptical Argument and Radical Skepticism about Meaning The main skeptical argument developed by Kripke in chapter 2 can be reconstructed as being of the following form:3 P1. If in the past it was a fact that I meant addition by ‘+’, then either: (i) this fact was determined by non-intentional facts of the following sorts: .................... or (ii) the fact that I meant addition by ‘+’ was a primitive fact; that is, a fact not determined by non-intentional facts. P2. It is not the case that non-intentional facts of type (i) determined that I meant addition by ‘+’.

3 Kripke formulates the skeptical paradox, in the first instance, as a paradox about what I meant by my words in the past, as opposed to what I mean by them now. The advantage of this formulation is that it allows us to take the meanings of the words we use in stating the argument for granted while the argument is being given and evaluated. Hence we avoid a potential source of confusion that might result from relying on the meanings of certain words in stating the argument while at the same time questioning what those words mean. Of course, if the argument of Kripke’s skeptic were good, then one could hardly resist extending the conclusion drawn about the past to the present and future as well. Thus nothing of philosophical significance is lost by formulating the skeptical argument as being about the past.

418 • Essay Sixteen

P3. It is not the case that the fact that I meant addition by ‘+’ was a primitive fact. C1. Therefore, in the past there was no fact that I meant addition by ‘+’. Much of chapter 2 is devoted to canvassing the potential meaningdetermining facts needed to fill out premise 1, and to establishing premise 2 by showing that none of these facts really do determine meaning. For example, even if we grant that whatever I meant by ‘+’ must conform to my past calculations using that symbol, these calculations are not sufficient to determine that I meant addition by it, since the values of the addition function far outstrip the limited number of calculations I have made. Nor can we appeal to the beliefs we used sentences containing ‘+’ to express, or to the computational algorithm we associated with the sign. The problem is that beliefs and algorithms, like ordinary words and sentences, are things that have content. As such their contents are as subject to the challenge of Kripke’s skeptic as is the content of ‘+’. This skeptic may, for the sake of argument, grant that in the past we associated ‘+’ with a formula, or set of instructions, whose present content yields the sum of any pair of natural numbers as arguments, but he will question whether that formula, or those instructions, had that same content in the past, and he will challenge us to specify the relevant content-determining facts. Similarly, he may agree that in the past we used the sentence ‘9 + 16 = 25’ to express a belief, but he will question whether the content of that belief was about the addition function (as opposed to some other function that gives the same value for that pair of arguments), and he will not be satisfied until we have identified the facts that determine the content of that belief. In short, if we propose any intentional fact as that which determined what ‘+’ meant in the past, Kripke’s skeptic will respond by reformulating his original problem so as to apply to the content of that fact. Having started by challenging us to find facts that determined the content of our words, the skeptic will continue by challenging us to find facts that determined the contents of all of our mental states. Thus, if we are ever going to be able to answer the skeptic, we must find facts that determine content in general—linguistic and otherwise. Further potential meaning-determining facts discussed by Kripke include my verbal dispositions to use ‘+’—specifically, my past dispositions to answer the question What is m + n? for arbitrary numerals m and n. Although I actually performed relatively few such calculations, I was disposed to perform a great many more. Suppose, for example, that I never performed the particular calculation ‘68 + 57 = 125’. Still, if I had been asked, ‘What is 68 + 57?’, I would have answered ‘125’. In light of this, Kripke considers the view that what one means by ‘+’ is determined by the

The Rule-Following Paradox • 419

answers one is disposed to give to the full range of such questions (the dispositional analysis). On this view, the fact that in the past I meant addition by ‘+’ was determined by the fact I was disposed to answer any question  What is m + n? by giving the sum of the numbers denoted by m and n. This view is rejected for three reasons. First, one can mean addition by ‘+’ even if one is disposed to make some errors in calculation in a few cases (the error objection). Second, one can mean addition by ‘+’ even though, when the numbers get extremely large, one has no dispositions to answer the question What is m + n? (the finitude objection). Third, even if one’s dispositions to calculate with ‘+’ exactly matched the addition function, one could not conclude from that fact that the answers one was disposed to give were the correct answers, and so one could not conclude that one meant addition by ‘+’ (the normativity objection).4 Having rejected verbal dispositions as meaning-determining facts, Kripke’s skeptic considers the suggestion that what one means by a word is determined by the mental images or private, introspectable experiences one associates with it. This suggestion is easily dispensed with since (i) in the case of many words we mean something definite even though no distinctive mental images or experiences are associated with them, and (ii) even if in some cases a private mental occurrence is psychologically associated with a word, such an occurrence can’t determine what we mean by the word unless the occurrence is interpreted in a certain way—i.e., unless the mental occurrence is itself assigned a meaning or content, in which case it becomes the subject of skeptical doubt of precisely the sort it was invoked to resolve. In light of all of this Kripke’s skeptic takes himself to have eliminated all reasonable candidates for facts that might determine what I meant by my words in the past. Since he dismisses as hopelessly mysterious the idea that meaning facts might be wholly undetermined by other facts, he draws the skeptical conclusion that there are no facts about what I meant by ‘+’, or any other word, in the past.5 From here it would seem to be a short step to the conclusion that I didn’t mean addition, or anything else, by ‘+’ in the past. After all, one is 4 5

For discussion see essay 15 in this volume. The text is somewhat equivocal on this point. Kripke says:

Perhaps we may try to recoup, by arguing that meaning addition by ‘plus’ is a state even more sui generis than we have argued before. Perhaps it is simply a primitive state, not to be assimilated to sensations or headaches or any ‘qualitative’ states, nor to be assimilated to dispositions, but a state of a unique kind of its own. Such a move may in a sense be irrefutable, and if it is taken in an appropriate way Wittgenstein may even accept it. But it seems desperate; it leaves the nature of this postulated primitive state of ‘meaning addition by “plus” ’—completely mysterious. (1982, 51)

420 • Essay Sixteen

inclined to suppose, if in the past I meant something by ‘+’, in particular if I meant addition, then surely it was a fact that I meant addition, and so there was a fact about what I meant. However, since precisely this transition is supposed to be put into question by the skeptical solution, it is useful to articulate a line of reasoning that would justify it, and that may serve as a target for the proponent of the skeptical solution to attack. Extension of the Original Skeptical Argument S1. Since there is no fact that I meant addition or anything else by ‘+’, the sentences I meant addition by ‘+’ and I meant something by ‘+’ do not correspond to facts.6 S2. Therefore, I meant addition by ‘+’ and I meant something by ‘+’ are not true. S3. So, I did not mean addition by ‘+’ and I did not mean anything by ‘+’ are true. S4. Thus, I didn’t mean addition by ‘+’; indeed, I didn’t mean anything by ‘+’. The original skeptical argument purports to show that there was no fact that I meant addition by ‘+’, and by parity of reasoning, no fact that I meant anything else either. The extension of that argument uses this putative result to deny all meaning ascriptions that attribute any meaning to my use of ‘+’ in the past. On the assumption that there is nothing special about me, ‘+’, or the past, this result generalizes to all speakers, expressions, and times. Since presumably the argument can be repeated for other contentbearing acts and states, the result also generalizes to assertions, beliefs, and other intentional phenomena. This is radical skepticism about meaning in particular, and content in general. Such thoroughgoing skepticism about intentionality is extremely implausible, even pragmatically self-defeating in a certain clear sense; but it is not internally inconsistent, nor is it unfamiliar. Since according to this radical skepticism no one ever means, asserts, or believes anything, it follows that if one has succeeded in stating (asserting) the skeptical position, or arguing for it, then one has refuted it. The skeptic’s position is The passage continues with more discouraging words on how problematic it would allegedly be to take meaning facts to be primitive. I believe that the most natural interpretation of Kripke’s argumentative strategy here is to see him portraying the skeptic as rejecting the claim that meaning facts are primitive, while holding the possibility open that there may be an interpretation of what such primitiveness would amount to according to which Wittgenstein—the proponent of the skeptical solution—may accept the claim that meaning facts are primitive. I will explore this possibility in the final section of the essay. 6 Italics are used here (and elsewhere) to indicate single quotes. Occasionally they are also used for other purposes—e.g., for emphasis, or to mark the use of a foreign word or phrase. I trust that the context will make it clear which use is which.

The Rule-Following Paradox • 421

self-undermining not in the sense that it is internally inconsistent, but in the sense that it is inconsistent with the supplementary claim (which must be rejected by the skeptic) that he, or anyone else, has stated the skeptical position, argued for it, or believed the conclusion. (Of course, since rejecting that so & so is also a propositional attitude, crucially involving the notion of content, the skeptic must also reject—as we would put it—the claim that he has rejected anything.) What about the notion of truth? It is used by the skeptic in the extension of the original skeptical argument, where certain sentences are said to be true. Can the skeptic consistently make such a claim? It all depends on what one means by calling a sentence ‘true’. A good case can be made that our ordinary notion of truth applies to sentences only derivatively: a sentence is true, as used in a context C, if and only if what it says, the proposition it expresses, as used in C, is true; a sentence is false, as used in C, if and only if what it says, the proposition it expresses, as used in C, is false; and a sentence is not true and not false, as used in C, if it expresses no proposition in C—for example, because it is meaningless, or because, like imperatives and interrogatives, its being meaningful does not consist in its being used to express propositions. Since the claim that a sentence expresses a proposition entails that it has meaning, or content, the skeptic cannot accept the claim that any sentence expresses a proposition. Thus, on this ordinary conception of truth talk, the radical skeptic about intentionality cannot consistently characterize any sentence as true. However, this does not prevent him from replacing our ordinary truth predicate with a Tarski-style truth predicate applying to his own language. With such a predicate the skeptic could, in principle, secure all the advantages of disquotation and semantic ascent, without characterizing any of our ordinary semantic notions as applying to anything. When the truth predicate of the skeptic’s sentences is understood in this way, he can consistently assert S3 in the extension of the original skeptical argument, even if that assertion is extremely implausible and, from our nonskeptical point of view, pragmatically self-undermining.7 7 Another option for the radical skeptic would be to reformulate the extension of the original skeptical argument so as not to appeal to any notion of truth. For example, he might simply accept the premise that if in the past I meant addition by ‘+’ then it was a fact that in the past I meant addition by ‘+’. This plus the original conclusion that there was no such fact, will allow him to conclude that in the past I didn’t mean addition by ‘+’. Similar reasoning would apply to other meaning ascriptions. The internal consistency of the radical skeptical position is a point that appears to be missed by Paul Boghossian (1989, 1990). He considers the radical skeptical position of eliminativism about meaning as involving a kind of “error conception of discourse about meaning.” He asks whether the lesson of the skeptical argument could simply be the adoption of the radical skeptical position, and concludes that it could not. (For Boghossian, what he calls “the error conception” is one of three possible responses to Kripke’s skeptical

422 • Essay Sixteen

Such a position is perversely limited in how it allows one to characterize the linguistic and other intelligent activities of oneself and others. However the position is not merely hypothetical, and is hardly unknown in philosophy. I believe that the best way to make sense of Quine’s various skeptical doctrines, including the indeterminacy of translation and the inscrutability of reference, is to see him as being an eliminativist about all our ordinary semantic and intentional psychological notions, and as proposing to paradox, the other two being the skeptical solution—which Boghossian calls “nonfactualism about content”—and a “straight solution” to the paradox.) Could not the moral of the skeptical argument be understood to consist in an error conception of meaning discourse? It could not, for an error conception of such discourse, in contrast with error conceptions of other regions, is of doubtful coherence. The view in question would consist in the claim that all meaning-attributions are false: (1) For any S: S means that p is false. But the disquotational properties of the truth predicate guarantee that (1) entails (2) For any S: S has no meaning. (1) implies, that is, that no sentence whatever possesses a meaning. Since, however, a sentence cannot be false unless it is meaningful to begin with, this in turn implies that (1) cannot be true: for what (1) says is that some sentences—namely meaning-attributing sentences—are false. (1989, 523) A slight variation of the same argument is given in more detail in Boghossian (1990). Consider first an error conception. As the preceding discussion has argued, this amounts finally to the claim that (4) All sentences of the form “S has truth condition p” are false, where S is to be understood as ranging over sentences in the language of thought, or neural structures, as well as over public-language sentences. But now (4) would seem to have the immediate consequence that no sentence has a truth condition. For whatever one’s conception of “true”—whether robust or deflationary—a sentence of the form “S has truth condition p” will be true if and only if S really does have truth condition p; this is, of course nothing but a reflection of the truth predicate’s disquotational properties, properties it possesses on any conception of truth. And so, since all sentences of that form are held to be false, for no S and for no p does S have truth condition p. Now, however, a problem would seem immediate. For (4) implies, that is, that no sentence whatever has a truth condition. But what (4) says is that all truth-condition-attributing sentences are false. And these sentences cannot be false unless they have truth conditions to begin with. Hence, (4) implies both that truth-condition-attributing sentences have truth conditions and that they don’t have them. This is a contradiction. (174–75) Though I agree with Boghossian that the skeptic’s position is, in a certain sense, “of doubtful coherence,” and hence must be rejected, it seems to me that the conclusion drawn in the quoted passage is too strong; there is no contradiction of the sort maintained in the radical skeptical position arising from Kripke’s skeptical argument. The crux of the matter concerns the notions of truth and falsity employed by Boghossian. If these are our ordinary notions, as I have described them in the main text above, then Boghossian’s contention that it

The Rule-Following Paradox • 423

replace them with various nonintentional and behavioristic substitutes.8 This is not, of course, the position of Kripke’s Wittgenstein.

The Apparent Strategy of the Skeptical Solution Kripke’s Wittgenstein is a proponent of the skeptical solution. The key elements of this putative solution are (i) an acceptance of the conclusion of the original skeptical argument—namely that there are no meaning facts, and hence no facts corresponding to our use of sentences that ascribe meanings to our words, and (ii) a refusal to adopt the radical skeptical position arising from the extension of the original skeptical argument, and an insistence that our uses of some meaning ascriptions, like In is inconsistent to claim that a meaningless sentence is false is correct—since to say of a sentence s that s is true (or false) is to say that s expresses a proposition that is true (or false), and so s is meaningful. However, when this notion of what it is for a sentence to be true is involved, the proponent of the radical skeptical position holds that no sentences are true, and therefore denies the inference from s to the claim that s is true, used in Boghossian’s argument. Moreover, the radical skeptic would not formulate his thesis as the doctrine that meaning ascriptions are false (in the ordinary sense); rather he would simply claim that no sentences mean anything. Thus, if the notions of truth and falsity employed in Boghossian’s argument are the ordinary ones, then the argument misstates the radical skeptical position and employs an inference that would be denied by the radical skeptic. On the other hand, the radical skeptic is free to replace the ordinary notions of truth and falsity with Tarski-type substitutes for his own language that do not presuppose any intentional notions. Let ‘true’ and ‘false’ be such predicates. A skeptic who says that there is a true sentence written on the blackboard says something trivially equivalent to the following: either ‘Snow is white’ is written on the blackboard and snow is white, or ‘Grass is green’ is written on the blackboard and grass is green, etc. (Sentences of other languages could not be dealt with using the same predicate.) When this notion of truth is employed the skeptic will accept the inference from s to the claim that s is true, and vice versa. He will characterize meaning ascriptions as not true and, depending on how falsity is defined, he may even characterize them as false. However, now there is no contradiction in claiming that a sentence is both meaningless and true (or false). Similarly, the claim that there are conditions under which a sentence is true, as well as the claim that the sentence has truth conditions, are not conceptually tied to any claim about the meaning of the sentence. Thus, if Boghossian’s argument is taken to involve notions of truth and falsity that the radical skeptic is willing to embrace, then a key premise in the argument must be rejected. As a result, no contradiction of the sort Boghossian is looking for can be derived from the skeptic’s position. 8 This includes replacing the ordinary notion of truth, which is tied to content, with a Tarski-style predicate, which can be defined using only logic, set theory, syntax, and nonintentional object language concepts that Quine is willing to countenance. Ordinary reference is similarly replaced with Tarski-style disquotational reference for one’s own present language, while talk of the reference of other speakers and languages is understood to involve indeterminate translation into one’s own language plus disquotational reference in the home language. See chapter 11 of Soames (2003), vol. 2 for discussion.

424 • Essay Sixteen

the past I meant addition by ‘+’, are fully justified and correct, whereas potential uses of other ascriptions, like In the past I meant quaddition by ‘+’, are not.9 What is supposed to make the combination of (i) and (ii) possible is a conception of meaning ascriptions in which their uses are not seen as attempts to state facts, and so do not require facts for their justification. Since meaning ascriptions have legitimate uses, they are perfectly meaningful, even though their meaning does not consist in standing for possible facts the existence of which would make them true. Instead, they are said to have a kind of meaning that does not conform to the classical truth-conditional conception of meaning. Kripke gives a brief overview of this position in the following passages: I choose to be so bold as to say: Wittgenstein holds, with the skeptic, that there is no fact as to whether I mean plus or quus. But if this is to be conceded to the skeptic, is this not the end of the matter? What can be said on behalf of our ordinary attributions of meaningful language to ourselves and to others? Has not the incredible and selfdefeating conclusion, that all language is meaningless, already been drawn? In reply we must say something about the change in Wittgenstein’s philosophy of language from the Tractatus to the Investigations. Although in detail the Tractatus is among the most difficult of philosophical works, its rough outlines are well known. To each sentence there corresponds a (possible) fact. If such a fact obtains, the sentence is true; if not, false. . . . The simplest, most basic idea of the Tractatus can hardly be dismissed: a declarative sentence gets its meaning by virtue of its truth conditions, by virtue of its correspondence to facts that must obtain if it is true. For example, “the cat is on the mat” is understood by those speakers who realize that it is true if and only if a certain cat is on a certain mat; it is false otherwise. The presence of the cat on the mat is a fact or condition-in-the-world that would make the sentence true (express a truth) if it obtained. So stated, the Tractatus picture of the meaning of declarative sentences may seem not only natural but even tautological. Nonetheless, as Dummett says, “the Investigations contains implicitly a rejection of the classical (realist) Frege-Tractatus view that the general form of explanation of meaning is a statement of the truth conditions.” In place of this view, Wittgenstein proposes an alternative rough general picture. (1982, 71–73) 9 Quaddition is an arithmetical function that agrees with the addition function on all pairs of arguments to which I have applied ‘+’, and which gives the value 5 to all other pairs of arguments.

The Rule-Following Paradox • 425

Kripke describes this alternative general picture of how a sentence gets its meaning as one in which truth conditions are replaced with justification conditions, or conditions of proper use.10 The application of this picture to meaning ascriptions, and to the skeptical paradox, is sketched in the following passage. Now the replacement of truth conditions by justification conditions has a dual role in the Investigations. First, it offers a new approach to the problems of how language has meaning, contrasted with that of the Tractatus. But second, it can be applied to give an account of assertions about meaning themselves, regarded as assertions within our language. Recall Wittgenstein’s skeptical conclusion: no facts, no truth conditions, correspond to statements such as “Jones means addition by ‘+’ ” . . . Now if we suppose that facts, or truth conditions, are of the essence of meaningful assertion, it will follow from the skeptical conclusion that assertions that anyone ever means anything are meaningless. On the other hand, if we apply to these assertions the tests suggested in Philosophical Investigations, no such conclusion follows. All that is needed to legitimize assertions that someone means something is that there be roughly specifiable circumstances under which they are legitimately assertable, and that the game of asserting them under such conditions has a role in our lives. No supposition that ‘facts correspond’ to those assertions is needed. (1982, 77–78)11 10

See Kripke (1982, 73 and 74). A word of caution is needed here regarding Kripke’s talk about assertions that someone means something, and about substituting conditions of legitimate assertability for truth conditions. Although Kripke here speaks as if, according to the skeptical solution, meaning ascriptions are used to assert propositions, as opposed to performing other nonassertive speech acts, earlier he has warned us that in talking about assertions and assertability conditions in connection with the skeptical solution he is adopting an oversimplified terminology. Thus he says: 11

Of course Wittgenstein does not confine himself to declarative sentences, and hence to assertion and denial, as I have just done. On the contrary, any reader of the earlier parts of Philosophical Investigations will be aware that he is strongly concerned to deny any special primacy to assertion, or to sentences in the indicative mood. (See his early examples “Slab!”, “Pillar!”, etc.) This in itself plays an important role in his repudiation of the classical realist picture. Since the indicative mood is not taken as in any sense primary or basic, it becomes more plausible that the linguistic role even of utterances in the indicative mood that superficially look like assertions need not be one of ‘stating facts’. Thus, if we speak properly, we should not speak of conditions of ‘assertion’, but rather, more generally, of the conditions when a move (a form of linguistic expression) is to be made in the ‘language game’ [my emphasis]. If, however, we allow ourselves to adopt an oversimplified terminology more appropriate to a special range of cases, we can say that Wittgenstein proposes a picture of language based, not on truth conditions, but on assertability conditions or justification conditions: under what circumstances are we allowed to make a given assertion? (1982, 73–74)

426 • Essay Sixteen

With this in mind we may look back to the extension of the original skeptical argument to see how the radical skeptical conclusion may be blocked. Whereas S1 may be accepted (and perhaps even S2), S3 may nevertheless be rejected. S1. Since there is no fact that I meant addition or anything else by ‘+’, I meant addition by ‘+’ and I meant something by ‘+’ do not correspond to facts. S2. Therefore, I meant addition by ‘+’ and I meant something by ‘+’ are not true. S3. So, I did not mean addition by ‘+’ and I did not mean anything by ‘+’ are true. Moreover, uses of Soames meant addition by ‘+’, and Soames meant something by ‘+’, may be regarded as correct and justified, whereas uses of the negations of these sentences are not. On this interpretation, the proponent of the skeptical solution rejects the skeptic’s radical claim that no one means anything by their words as based on an incorrect conception of meaning.

The Tension in the Skeptical Solution: What Is a Fact? The skeptical solution denies that there are meaning facts. But in what sense of “fact” is there supposed to be no fact that I meant addition by ‘+’? Kripke’s Wittgenstein does not deny that one can correctly say things like It is a fact that Soames meant addition by ‘+’ or Jones was unaware of the fact that Soames meant addition by ‘+’. One can correctly say these things provided one doesn’t misconstrue them as involving some mistaken, philosophically contentious sense of “fact”. The problem is that we are not told what this mistaken sense of “fact” is. Kripke addresses this point immediately after describing Wittgenstein’s skeptical solution as a position that accepts the hypothetical skeptic’s conclusion that there are no meaning facts, and in particular no fact that constitutes my having meant addition by ‘+’. Kripke says: Admittedly, I am expressing Wittgenstein’s view more straightforwardly than he would allow himself to do. For in denying that there is any such fact, might we not be expressing a philosophical thesis that doubts or denies something everyone admits? We do not wish to doubt

In light of this warning we are well advised, at this point, to leave it as an open question whether, according to the skeptical solution, utterances of meaning ascriptions are to be regarded as assertions proper, or as performances of some other speech act.

The Rule-Following Paradox • 427

or deny that when people speak of themselves and others as meaning something by their words, as following rules, they do so with perfect right. We do not even wish to deny the propriety of an ordinary use of the phrase ‘the fact that Jones meant addition by such-and-such a symbol’, and indeed such expressions do have perfectly ordinary uses. We merely wish to deny the existence of the ‘superlative fact’ that philosophers misleadingly attach to such ordinary forms of words, not the propriety of the forms of words themselves. (1982, 69)12 Our problem in understanding the skeptical solution lies in understanding the sense of “fact” in which meaning ascriptions do not state facts, and in which one can correctly deny instances of (1) If S, then it is a fact that S. such as (2) If I meant addition by ‘+’, then it is a fact that I meant addition by ‘+’. In trying to spell out this crucial sense of “fact”, one is pulled in two directions. On the one hand, we may build a substantial amount into the notion of a fact. For example, we might restrict the facts under consideration to facts intrinsic to me. If we do this, then it is not unreasonable to deny meaning ascriptions like (2). Under this interpretation such denials do not amount to the claim that there are no meaning facts, but only to the claim that if there are meaning facts, then these facts include social, causal, or historical elements external to the individual. Unfortunately, although this claim about meaning facts is coherent and plausible, it does not provide a good interpretation of the skeptical solution outlined by 12

The passage continues:

It is for this reason that I conjectured above (p. 5), that Wittgenstein’s professed inability to write a work with conventionally organized arguments and conclusions stems at least in part, not from personal and stylistic proclivities, but from the nature of his work. Had Wittgenstein . . . stated the outcomes of his conclusions in the form of definite theses, it would have been very difficult to avoid formulating his doctrines in a form that consists in apparent skeptical denials of our ordinary assertions. . . . If on the other hand, we do not state our conclusions in the form of broad philosophical theses, it is easier to avoid the danger of a denial of any ordinary belief. . . . Whenever our opponent insists on the perfect propriety of an ordinary form of expression . . . we can insist that if these expressions are properly understood, we agree. The danger comes when we try to give a precise formulation of exactly what it is that we are denying—what ‘erroneous interpretation’ our opponent is placing on ordinary means of expression. It may be hard to do this without producing yet another statement that, we must admit, is still ‘perfectly all right, properly understood’. (1982, 69–70)

428 • Essay Sixteen

Kripke. First, the view that meaning facts include social or other elements extrinsic to the individual does not, in and of itself, threaten the orthodox, truth-conditional construal of sentences as getting their meanings by virtue of standing for possible facts. Even meaning ascriptions may be treated as fitting this model, as long as it is recognized that the possible facts they stand for are at least partially social, or extrinsic to the individual. But it is a central feature of the skeptical solution that it is supposed to constitute a radical alternative to the truth-conditional conception. Second, in giving the skeptical argument, Kripke did not invoke anything like the now familiar Twin-Earth or arthritis examples that are designed to show that meaning is not individualistically determined.13 On the contrary, the considerations adduced by Kripke’s skeptic to show that facts about me don’t determine meaning carry over to social facts of the same type. For example, the error, finitude, and normativity objections to the view that my dispositions to calculate with ‘+’ determined that I meant addition also apply to the view that the dispositions of the community to calculate with ‘+’ determined its meaning. Thus, if the original skeptical argument succeeded in showing that facts about me don’t determine meaning, then analogous reasoning would show that facts about me and my community don’t either.14 But it is the essence of the skeptical solution that it should survive, and in fact be motivated by, the reasoning of the skeptical argument. Perhaps some other substantive conception of meaning facts can be found for the skeptical solution to deny. Suppose, for example, one admits both that meaning ascriptions express propositions, and so are proper objects of assertion and belief, and that there is a certain minimal sense of “true” and “fact” according to which for any sentence S that expresses a proposition, M. S iff it is true that S iff it is a fact that S iff the proposition that S is true 13 14

For examples of this sort see Putnam (1975), and Burge (1979). Kripke explicitly recognizes this:

Finally, the point made in the last paragraph, that Wittgenstein’s theory is one of assertability conditions, deserves emphasis. Wittgenstein’s theory should not be confused with a theory that, for any m and n, the value of the function we mean by ‘plus’, is (by definition) the value that (nearly) all the linguistic community would give as the answer. Such a theory would be a theory of the truth conditions of such assertions as “By ‘plus’ we mean such-and-such a function,” . . . The theory would assert that 125 is the value of the function meant for the given arguments, if and only if ‘125’ is the response nearly everyone would give, given these arguments. Thus the theory would be a social, or community-wide, version of the dispositional theory, and would be open to at least some of the same criticisms as the original form. (1982, 110–11)

The Rule-Following Paradox • 429

is regarded as unproblematic, and perfectly correct.15 On this view, there is a certain minimal conception of truth and facts according to which meaning ascriptions have truth conditions, and many such ascriptions state facts. However, there is at least one way of understanding the classical truth-conditional conception of meaning in which it makes claims that are considerably stronger than this. According to this strong version of the truth-conditional conception, the facts correlated with sentences play an explanatory role in meaning and understanding: a sentence s is meaningful because it stands for a potential fact the obtaining of which would make s true: a person understands s in virtue of knowing that s is true iff the fact in question obtains. Perhaps the philosophically contentious sense in which the skeptical solution denies that there are meaning facts is one in which it is assumed that any potential meaning fact must be explanatory in this way.16 On this interpretation, the skeptical solution doesn’t deny that there are facts about meaning, in the perfectly ordinary and minimal sense of (M); it simply denies that these facts play a certain explanatory role. As before, such a denial is not patently unreasonable. However, as an account of the skeptical solution, this interpretation faces a difficulty of its own. In giving the original skeptical argument Kripke seemed to be willing to consider any nonintentional fact as a candidate for being a meaningdetermining fact. Moreover, the argument did not focus on the explanatory role posited by the strong version of the truth-conditional conception of meaning. Thus, if the interpretation is to succeed, care must to taken to show precisely how the denial of explanatory meaning facts connects with the skeptic’s original argument. 15 Propositions, as I am conceiving of them here, are objects of belief, assertion, and other attitudes, as well as being potential bearers of truth-value. So in cases in which it is correct to report someone as believing or asserting (the proposition) that so and so, the sentence one uses in ascribing the content of that belief or assertion should typically be regarded as expressing a proposition. 16 Note, on this account, the potential fact associated with a sentence plays two conceptually distinct (though related) explanatory roles. In evaluating the correctness or incorrectness of strong versions of the truth-conditional conception of meaning, the possibility should be left open that there may be cases in which the potential fact in question plays one of these explanatory roles but not the other. For example, there might be a sentence s (or other expression e) that was originally endowed with meaning by an initial speaker or group of speakers who antecedently grasped a certain potential fact (or property/object) and stipulated that s (or e), was to stand for it—even though later speakers are able to understand s (or e) without having any such antecedent grasp of the potential fact (or property/object) in question. In such a case it would be plausible to maintain that the potential fact associated with the sentence played the initial meaning-endowing explanatory role, but not the role of explaining subsequent understanding. There may be other situations in which neither role is played by such a potential fact.

430 • Essay Sixteen

This brings us to the other direction in which one is pulled in trying to make sense of the skeptical solution. Instead of trying to build some contentious philosophical substance into the notion of a fact, we might simply accept the minimal conception (M), and see the skeptical solution as denying that there are meaning facts in even this minimal sense. The virtue of this interpretation is that it fits very well with the original skeptical argument. When, in giving the argument, we raised the question of whether one or another set of facts determined meaning facts, we didn’t explicitly assume anything about these facts beyond their being, or corresponding to, true claims. Thus, the import of the original skeptical conclusion was naturally taken to be tantamount to the claim that there are no true propositions to the effect that I meant so and so in the past. Consequently, if we interpret the skeptical solution as denying that there are meaning facts even in the minimal sense of (M), then the skeptical solution will be in apparent agreement with the conclusion of the original skeptical argument—as it is supposed to be. However, now we face a different problem, since it is not easy to see how one can characterize a use of the meaning ascription Soames meant addition by ‘+’ as perfectly correct and justified, while rejecting the claim that it is true, or a fact, that Soames meant addition by ‘+’. Given (M), we can do this only if we deny that meaning ascriptions like this express propositions. Such a position would be genuinely surprising. We know, of course, that some meaningful sentences, such as imperatives and interrogatives, do not express propositions. But meaning ascriptions seem quite different from these. Perhaps they could be thought of as performatives, along the lines of ‘I hereby name this barge the Hillary’, or ‘I hereby pronounce you husband and wife’. According to some analyses, utterances of these sentences are not attempts to describe any act of naming a barge or marrying people. In general, according to these analyses, utterances of performative sentences are not seen as expressing propositions, or purporting to describe states of affairs, and so they are not taken to be candidates for truth and falsity. Rather than describing certain acts, performative utterances in appropriate circumstances are seen as constituting the performance of those acts—in our examples, naming a barge and marrying two people. With this in mind, one might suggest that saying Jones means addition by ‘+’, is performative in the following way: to utter such a sentence in appropriate circumstances is to take Jones into one’s linguistic community, to certify him as a competent user of ‘+’, and to license him to use ‘+’ to do what we call “adding”. One way of interpreting the skeptical solution is to view it as treating meaning ascriptions roughly along these lines. In the sections that follow I will examine different interpretations of the skeptical solution. I will begin with the interpretation according to

The Rule-Following Paradox • 431

which meaning ascriptions are said not to express propositions, or to state facts in even the minimal sense (M), but rather are claimed to have some other nondescriptive, perhaps performative, kind of meaning. I will argue that this interpretation fails in several respects—in particular, the position outlined lacks crucial features that Kripke attributes to the skeptical solution, it is implausible on its face, and it can be shown to be incoherent. Next, taking George Wilson’s insightful interpretation of Kripke as a point of departure,17 I will articulate a view which grants that meaning ascriptions do express propositions, and may state facts in the sense of (M), but denies that these facts play the explanatory role demanded by a certain strong version of the truth-conditional conception of meaning. Accordingly, I will distinguish strong and weak versions of the truthconditional conception and present an independent argument that the strong version cannot be maintained for all proposition-expressing sentences. This will result in a modest, but I hope reasonable, skepticism about the role of truth conditions in explaining meaning and understanding. Finally, I will take up the question of how this modest skepticism relates to Kripke’s skeptical solution. I will argue that my modest skepticism about truth conditions lacks certain essential features of the skeptical solution as Kripke describes it. In particular, its rejection of explanatory truth conditions does not apply to all sentences, and is substantially more limited in scope than Kripke’s skeptical solution is supposed to be; moreover, Kripke’s skeptical argument cannot be seen as providing an argument for the modest skepticism I describe. In the last section of the paper I will assess where the apparent failure of these two interpretations of the skeptical solution leaves us, and I will try to bring together the positive lessons to be learned from Kripke’s discussion.

The Minimal Conception of Facts and the Nondescriptive Version of the Skeptical Solution According to this interpretation, meaning ascriptions like in the past I meant addition by ‘+’ do not express propositions, or purport to state facts in even the minimal sense (M), but rather are claimed to have some other, nondescriptive kind of meaning. The suggestion that they might be given a performative analysis is meant to illustrate this possibility, without ruling out other potential alternatives. The important point is that utterances of meaning ascriptions are not seen as attempts to describe 17

Wilson (1994, 1998).

432 • Essay Sixteen

any state of affairs, or to make any claim about the way the world is. Rather, since they have some other, nondescriptive function, they may be perfectly correct, even though they do not correspond to facts, and are not properly assessable as true or false. Because of this, one who took this position could, in principle, accept the conclusion of the original skeptical argument, while blocking the extension of that argument, thereby avoiding the need to deny such ascriptions as In the past Soames meant addition by ‘+’.18 Unfortunately, there several intractable problems that make adopting this picture untenable. First, as I tried to show in “Skepticism about Meaning: Indeterminacy, Normativity, and the Rule-Following Paradox,”19 the original skeptical argument is flawed. Given our ordinary, minimal conception of facts (M), one cannot establish that there are no facts about meaning. But if one can’t establish this conclusion, then one doesn’t need some special nondescriptive analysis of meaning ascriptions in order to be able correctly to characterize my past use of ‘+’ as meaning addition.20 Second, there is reason to think that the nondescriptive analysis of meaning ascriptions is indefensible on its face. If the skeptical argument motivated the claim that meaning ascriptions are nondescriptive, then a parallel argument would motivate the claim that propositional attitude ascriptions are too. But surely, sentences ascribing beliefs, assertions, and intentions to agents purport to describe those agents in ways that may be true or false. All of these sentences, along with meaning-ascribing sentences, are routinely characterized by ordinary speakers as being assessable for truth; all are unproblematically embeddable in ordinary truthconditional constructions, such as antecedents of conditionals, in ways in which interrogatives, imperatives, and pure performatives—like ‘Hello’— are not; and all may themselves appear as objects of propositional attitude verbs, like ‘know’ and ‘believe’, which are used to indicate our cognitive relations to facts, and to propositions generally. Satisfaction of prosaic conditions like these is all that is required for a sentence to be truth-assessable, and to belong to fact-stating discourse, when the notions of truth and facts are understood in the ordinary, minimal sense characterized by (M). Since meaning and propositional attitude ascriptions satisfy these conditions, they qualify as fact-stating.

18 An interpretation of the skeptical solution along roughly these lines is given by Crispin Wright (1984) and by Paul Boghossian (1989). 19 Essay 15 in this volume. 20 Kripke’s Wittgenstein might, of course, be forgiven for not noticing this. Hence this is not a criticism of the interpretive hypothesis that attributes the nondescriptive analysis of meaning ascriptions to him. However, it is a criticism of that analysis when put forward as the proper response to the skeptical argument.

The Rule-Following Paradox • 433

Third, in addition to rendering the skeptical solution implausible, the non-fact-stating interpretation of meaning-ascribing sentences robs the skeptical solution of its intended philosophical power. The goal of Kripke’s Wittgenstein was to replace what he regarded as a radically defective philosophical conception of meaning with a fundamentally different conception. It is no defect of the truth-conditional conception that it does not apply to interrogatives, imperatives, or some performatives. If, for some reason, one thought that it did not apply to meaning and propositional attitude ascriptions, because they too do not purport to be fact-stating, this would not discredit the truth-conditional conception, but simply limit its scope. A skeptical solution that amounted to no more than this would tell us nothing about the vast range of non-contentascribing, yet philosophically significant, discourse. For all that we would be told by such a solution, perception sentences, ascriptions of feelings like pain, and mathematical sentences might continue to be analyzed as fact-stating, and full-bloodedly truth conditional. But if that were the case, then the skeptical solution could not play the crucial role intended by Kripke’s Wittgenstein in the private language argument, for example, or in the philosophy of mathematics.21 21 Crispin Wright and Paul Boghossian, who interpret the skeptical solution along roughly the nondescriptive lines sketched above, provide arguments—not found in Kripke’s text—that a nondescriptive analysis of meaning ascriptions leads immediately to a nondescriptive analysis of all sentences. See Wright (1984, 769) and Boghossian (1989, 524–25). The basic idea may be illustrated by considering Wright’s argument in the following form (which is here expressed in a slightly different way than the way he presents it): Suppose that s is a descriptive sentence—e.g., ‘snow is white’—that expresses a proposition and purports to state a fact. Now consider the material biconditional whose left side attributes truth to s and whose right side is s itself—e.g., ‘Snow is white’ is true iff snow is white. It would certainly seem that this biconditional ought to be accepted. Suppose further that we assume, independently of the skeptical solution, that if B is a descriptive sentence that expresses a proposition, then the sentence A iff B is acceptable only if A is descriptive, and expresses a proposition. It then follows that the left side of our original biconditional, which attributes truth to s, must also be taken to be descriptive, and hence to express a proposition. However, this is problematic. For if we are here employing the ordinary notion of truth, according to which a sentence is true if and only if it expresses a true proposition, then a sentence ‘S’ is true is equivalent to a sentence There is some proposition p such that ‘s’ expresses p and p is true. Since a meaning ascription ‘S’ expresses a proposition follows from this sentence, there are grounds for the proponent of the nondescriptive version of the skeptical solution to classify the corresponding sentence ‘S’ is true as being nondescriptive. (We are here tacitly assuming that any sentence from which a nondescriptive sentence follows is itself nondescriptive.) But now we have characterized the claim that s is true as both descriptive and nondescriptive. If one takes the correctness of the skeptical solution as given (and also accepts the supplementary principles appealed to in the argument), one may regard this as a reductio of the claim that there are any nondescriptive sentences that the skeptical solution doesn’t apply to, and hence that its application is universal.

434 • Essay Sixteen

Finally, as if all this were not enough, the nondescriptive version of the skeptical solution is incoherent. One way to bring out this incoherence is with the following argument: Consider the sentence, ‘no meaning ascription expresses a proposition’, used in stating this version of the skeptical solution. It makes a theoretical claim about a certain class of sentences, and so must itself be proposition-expressing and fact-stating. Thus, its negation, ‘some meaning ascription expresses a proposition’, ought have the same status. This sentence is an existential generalization whose instances are sentences of the form, ‘s is a meaning ascription and s expresses a proposition’. Surely if the existential generalization expresses a proposition, then some of its instances do too, in which case some sentences of the form ‘s expresses a proposition’ must express propositions. But these are meaning ascriptions of a certain sort, which, by the nondescriptive version of the skeptical solution, do not express propositions. Thus we have it that some sentences both express propositions and do not express propositions. This contradiction follows from the nondescriptive version of the skeptical solution plus a very plausible set of supplementary assumptions (3). (3) a. Since the sentence ‘no meaning ascription expresses a proposition’ is used to make a theoretical claim about a certain class of sentences, it must express a proposition, and purport to be fact-stating. b. If s and s′ are negations of one another, then one expresses a proposition (and so purports to be fact-stating) iff the other does so as well. c. An existential generalization expresses a proposition only if some of its instances do. d. A conjunction expresses a proposition only if both of its conjuncts do. e. A sentence of the form ‘s expresses a proposition’ is a meaning ascription. It is hard to see how any of these could plausibly be denied. The nondescriptive version of the skeptical solution, as I have presented it, doesn’t I don’t think this is the best way to view the matter, however. What this reconstruction of Wright’s argument shows is that the nondescriptive version of the skeptical solution, together with (i) a certain set of supplementary assumptions and (ii) the claim that there are some descriptive sentences that express propositions, is incoherent. However, the nondescriptive version of the skeptical solution can be shown to be incoherent on independent grounds, as I argue below. Thus the incoherence reached in this reconstruction of Wright’s argument is better blamed on this version of the skeptical solution than on the claim that there are sentences it doesn’t apply to. Wright seems to share this view, since he goes on to express skepticism that we can make sense of a universal nondescriptivism (1984, 769–70).

The Rule-Following Paradox • 435

deny that some sentences are used descriptively, express propositions, and purport to state facts. But if some sentences have these features, then ‘no meaning ascription expresses a proposition’ should be one of them, in which case (3a) is acceptable. Since (3b–e) are too, the argument can plausibly be taken to be a reductio of the nondescriptive version of the skeptical solution. It is also possible to give a different argument with the same result, which does not appeal to any supplementary premises beyond the nondescriptive version of the skeptical solution itself. The first step in the argument is simply a statement of the view that meaning ascriptions are nondescriptive, and so do not express propositions. (i) No meaning ascriptions—e.g., Soames means addition by ‘+’, ‘+’ means addition, ‘2 + 2 = 4’ means (expresses the proposition) that the sum of two and two is four, ‘2 + 2 = 4’ means something/expresses some proposition—express propositions. Next we observe that, according to the skeptical solution, sincere utterances of the particular sentences mentioned in (i) are correct, even though they do not state facts and so are not true in the ordinary sense. The reason for this is that they have some other, nondescriptive function, and so are not even attempts to state facts. Indeed, the skeptical solution tells us that it would be incorrect to seriously utter the negations of the sentences mentioned in (i). This gives us (ii). (ii) It would be incorrect to say: Soames does not mean addition by ‘+’, ‘+’ doesn’t mean addition, ‘2 + 2 = 4’ does not mean (express the proposition) that the sum of two and two is four, ‘2 + 2 = 4’ does not mean anything, or express any proposition. Now consider another class of meaning ascriptions—e.g., sentences of the form s means something, s expresses a proposition, s means that p, s expresses the proposition that p, where ‘s’ is replaced by a name of one of the meaning ascriptions mentioned in (i). These are meaning ascriptions about meaning ascriptions. So, by (i) they do not express propositions and are not candidates for truth or falsity in the ordinary sense. This gives us a variety of results of the sort (iii). (iii) The following sentences do not express propositions: “‘+’ means addition means something”, “‘+’ means addition expresses a proposition”, and so on. But the fact that these meaning ascriptions about meaning ascriptions don’t express propositions doesn’t give us any more reason to reject them than there is to reject the original meaning ascriptions about sentences that are not themselves meaning ascriptions. According to the skeptical

436 • Essay Sixteen

solution, there is a role in our lives for meaning ascriptions about ordinary sentences like ‘2 + 2 = 4’ which justifies our use of those meaning ascriptions. But surely if this is so, then there is a role in our lives for meaning ascriptions about some sentences that themselves contain the word ‘meaning’, or the phrases ‘expresses a proposition’, and ‘expresses the proposition that’. Hence there ought to be a role for meaning ascriptions of the sort given in (iii) that justifies their use. In particular, the skeptical solution should be construed as legitimating (iv). (iv) Serious uses of (a) are correct, and corresponding uses of (b) are incorrect. (a) ‘+’ means addition means something ‘+’ means addition expresses a proposition (b) ‘+’ means addition does not mean anything ‘+’ means addition does not express a proposition Thus we have from (i) that ‘+’ means addition does not express a proposition; whereas from (iv) we are licensed to say: “‘+’ means addition does express a proposition”. Similarly, from (iv) we have it that the characterization gotten from (i) is a mistake. So we have it that ~S, but it is incorrect or mistaken for one seriously to utter that very sentence. In addition, the proponent of the skeptical solution must say something of the form s does not express a proposition on the basis of (i) and something of the form s does express a proposition on the basis of (iv). According to the nondescriptive version of the skeptical solution, this may not be a case of asserting contradictory propositions. Presumably, though, it is still incoherent—like ordering someone both to close the door and not to close it. For all these reasons the nondescriptive version of the skeptical solution must be rejected. As a philosophical position to be evaluated on its own merits it is incorrect and indefensible. As an interpretation of Kripke’s text it fails to make sense of important aspects of his discussion— such as the claim that it is only when one employs a philosophically contentious sense of “fact” that it is denied that there are facts,22 the claim that the skeptical solution represents a radical rejection of the truth-conditional conception of meaning and a proposal to replace it with something else, and the claim that the proposal in question has far-reaching consequences for other philosophically significant regions of discourse. These failures give us good reason to look for another interpretation.

22

This point is made by Alex Byrne (1996).

The Rule-Following Paradox • 437

The Explanatory Conception of Meaning Facts: An AntiTruth-Conditional Version of the Skeptical Solution What we need is an interpretation according to which the skeptical solution grants that there are facts about meaning, in the minimal sense (M) of “fact” discussed above. The needed interpretation should grant that the meaning ascription, In the past Soames meant addition by ‘+’ expresses a proposition, and is true iff in the past I meant addition by ‘+’. It should also grant that I did mean addition by ‘+’, and hence that it is a fact (and it is true) that in the past I meant addition by ‘+’. Nevertheless, such an interpretation must be skeptical in the sense of providing us with a reason to revise some of our fundamental views about meaning. In order to determine whether such an interpretation can be constructed, we must answer two questions: (i) If the original skeptical argument is not intended to show that there are no meaning facts, in the minimal sense of “fact”, what is it supposed to show? (ii) If the skeptical solution is to be seen as compatible with the claim that sentences, including meaning ascriptions, have truth conditions in the minimal sense just illustrated, in what respect does it represent a radical rejection of the truth-conditional conception of meaning, and a proposal to replace it with something else? In order to investigate these questions we will begin by examining George Wilson’s illuminating interpretation, presented in “Kripke on Wittgenstein and Normativity.”23 There Wilson argues that the lesson of Kripke’s original skeptical argument is not that there are no meaning facts, but rather that if we accept a certain natural, but decidedly nontrivial conception of meaning, then—because of the considerations in the skeptical argument—there are no meaning facts. Anyone who accepts the relevant conception of meaning will have to accept that there are no meaning facts, and will be under great pressure to acquiesce in the conclusion that no one ever means anything. Since this is not the position of Kripke’s Wittgenstein, the skeptical argument is viewed as a reductio ad absurdum of the initial conception of meaning, and the skeptical solution is put forward as a conception of meaning that replaces the one rejected. For Wilson the target of the reductio is something called “classical realism,” according to which what it is for a sentence to be meaningful is for it to have “classically realist truth conditions.” What these are turns out not to be fully specified, and it will be important to make this notion as clear as possible. However, certain things are evident from the outset. First, Wilson regards the reductio as being designed to establish the global conclusion 23

Wilson (1994).

438 • Essay Sixteen

that no sentence has classically realist truth conditions. Second, it is not supposed to follow from this conclusion that no sentences are true, or that none state facts, in the ordinary, minimal sense illustrated by (M). Moreover, since the conclusion of the reductio is global, it is seen as automatically applying to different domains of discourse—meaning ascriptions, mathematical sentences, psychological sentences, and so on. Thus, on Wilson’s interpretation, conclusions established by the rule-following considerations are properly seen as being directly relevant (at least potentially) to philosophical analyses across all domains. Wilson first gives a brief overview of his interpretation, as it applies to Kripke’s example of the plus sign, and then provides a generalized statement of the argument, as it might apply to any predicate. The situation with ‘+’ is characterized as follows: The skeptic, it seems, begins with a skeletal but plausible conception of what it is to mean something by a term. Restricting ourselves, first of all, to the exemplary case of addition, the conception runs as follows. If, during a certain period, I meant addition by “+”, then, during that time, I must have adopted some standard of correctness for my actual and potential applications of the term. Meaning something by “+” essentially involves having a policy about what is to count as a correct and incorrect application of the expression. . . . What I must do to establish the standard for myself, the skeptic maintains, is to adopt a rule or a commitment to the effect that correct answers to queries of the form j + k = ? are given by the values of the addition operation for the pairs of numbers that, query by query, are in question. Having adopted this linguistic rule for “+”, correct applications of the term are governed by the infinite table of values that addition generates. Alternatively, we can say that, in meaning addition by “+”, I must successfully ‘single out’ the addition function and, correlatively, form the intention or purpose, concerning just that function, that it is to determine correct applications of my use of the term. . . . It is precisely at this juncture that the skeptic interposes his disturbing challenge. It is my contention that the skeptic does not doubt that the skeletal conception of what is involved in meaning addition by “+” is right. What he does doubt is that this conception can be intelligibly filled in with an account of how I (or any other speaker) achieve the adoption of the specified linguistic policy or commitment. What, he asks, does my adopting such a conception concretely consist in? In particular, what are the facts about me in virtue of which it is addition that governs correct applications of “+” for me and not some other initially similar but divergent arithmetic operation?24 24

Wilson (1994, 368–69).

The Rule-Following Paradox • 439

Wilson generalizes this discussion by formulating how the skeptical argument should be construed, when applied to any predicate. His “classically realist” conception of meaning is seen as incorporating the following conception of normativity.25 The Classically Realist Conception of Normativity Ns. If x means something by a (general) term ϕ, then there is a set of properties P1–Pn, that govern the correct application of ϕ for x. Next comes the characterization of what it is for a set of properties to govern the correct application of a predicate.26 Government Gov. A set of properties P1–Pn governs the correct application of a term ϕ for x iff x has the intention (has adopted the linguistic commitment) that ϕ correctly applies to an object o iff o has P1–Pn. At this point, Wilson assures us that “the skeptic insists that, if certain properties are to function as conditions of correct applicability for x’s use of ϕ, then there must be concrete facts about x which establish that it is just those properties which she has successfully singled out and about which she has formed a proper semantic commitment. The existence of conditions of applicability for a term must be intelligibly grounded in facts about the speaker’s psychology and/or social history.”27 This leads to the “grounding constraint.”28 Grounding G. If there is a set of properties, P1–Pn, that govern the correct application of ϕ for x, then there are facts about x that constitute [establish, determine] P1–Pn as the conditions that govern x’s use of ϕ. According to Wilson, the skeptic then closes the trap by arguing that there are no facts that ground the speaker’s use of ϕ. Wilson calls this the “basic skeptical conclusion.”29 The Basic Skeptical Conclusion BSC. There are no facts about x that constitute [establish, determine] any set of properties as conditions that govern x’s use of ϕ. 25

Wilson (1994, 370). Wilson (1994, 369–70). 27 Wilson (1994, 370). 28 Wilson (1994, 370). 29 Wilson (1994, 370). 26

440 • Essay Sixteen

The skeptic concludes from Ns, Gov, G, and BSC that x doesn’t mean anything by ϕ. Since the argument is supposed to apply for arbitrary choices of speaker and term, the radical skeptical conclusion is that no one ever means anything by any term. Of course, Kripke’s Wittgenstein is no skeptic. Thus, Wilson interprets him as rejecting these conclusions, and as treating the skeptic’s argument as a reductio ad absurdum of the classically realist conception of normativity, Ns, which is supposed to follow from the classically realist conception of the meaning of a sentence as consisting of its truth conditions. What is the classically realist conception of meaning as truth conditions, such that Ns may properly be seen as following from it? Consider the following, minimal interpretation of the truth-conditional conception. A Minimal Truth-Conditional Conception A. (i) A normal declarative sentence is meaningful iff it expresses a proposition; it is true iff the proposition it expresses is true. The proposition that P is true iff P; hence if a sentence s expresses the proposition that P, then s is true iff P. (ii) A speaker understands a sentence s that expresses the proposition that P iff the speaker knows that s expresses the proposition that P, and hence knows that s is true iff P. B. (i) A predicate is meaningful iff it expresses a property; it (correctly) applies to an object iff the object has that property. An object o has the property of being F iff o is F; hence if a predicate ϕ expresses the property of being F, then ϕ (correctly) applies to an object o iff o is F. (ii) A speaker understands a predicate ϕ that expresses the property of being F iff the speaker knows that ϕ expresses the property of being F, and hence knows that ϕ (correctly) applies to an object o iff o is F. C. Similarly for other categories of expressions. This statement of the truth-conditional conception fits well with Wilson’s formulation of Ns and his characterization of what it is for properties to govern the use of a predicate. Putting these two together—Ns and Gov—we get the claim that if x means something by a (general) term ϕ, then there is a set of properties P1–Pn, such that x has the intention (has adopted the linguistic commitment) that ϕ correctly applies to an object o iff o has P1–Pn. This is close to what is stated in B above. Presumably, a speaker who knows that ϕ correctly applies to an object o iff o is F, and who intends to use ϕ in accord with its meaning, will intend to use it in such a way that it correctly applies to an object iff that object is F. Thus Ns would appear to be a natural consequence of the

The Rule-Following Paradox • 441

minimal truth-conditional conception of meaning—in which case a denial of Ns would be a denial of the minimal truth-conditional conception. However, this cannot be what Wilson has in mind. Putting aside ontological issues about the nature and existence of properties and propositions, which do not play a role in the argument, the minimal truthconditional conception is virtually undeniable.30 It is obvious that sentences and predicates do have truth and application conditions of this sort, and speakers do have the relevant beliefs and intentions. Try this for yourself. All of you who understand the predicate ‘red’ (and also what it means for a predicate to apply to an object), ask yourselves whether you know that the predicate ‘red’ correctly applies to an object iff it is red. The answer, of course, is that you do. The skeptical solution cannot plausibly be seen as denying this triviality. If the minimal truth-conditional conception is at stake, then the proper response to the skeptical argument is to deny principle G, provided that the determination relation is taken to be something strong and restrictive like a priori consequence, and to deny BSC, if the determination relation is construed as something weaker, like necessary consequence.31 Consequently, if one is trying to construct an in30 In Soames (1989, 585–87; essay 5 in vol. 1 of this collection) I raise a limited, somewhat technical, objection that could be extended to even the minimal truth conditional conception. I argue that in certain unusual cases one can know that s expresses the proposition that P (this could be extended to cases in which one can know that s both expresses the proposition that P and is true iff P), without understanding s—though it still holds that if s means that P and one understands s (as well as what it is for something to mean that, and be true iff, P), then one will be in a position to conclude that s means that P and is true iff P. The argument makes some controversial, but I believe defensible, assumptions about the proper semantic analysis of various constructions which are not directly relevant to the issues discussed here. Since the correctness or incorrectness of the argument does not affect the main issues related to the interpretation of the skeptical solution, I will put the argument aside for present purposes. Another technicality that I will put aside in discussing the minimal truth-conditional conception involves liar paradoxical sentences, and other sentences that contain partially defined predicates. I have argued (Soames 1999, chaps. 6 and 7) that some sentences of this type express propositions, but cannot correctly be characterized either as true or as not true. Such sentences are indeterminate—i.e., they are not determinately true and not determinately untrue. If S is such a sentence, and S iff S and ‘S’ is true iff S are normal material biconditionals, then it is not unreasonable to regard these biconditionals as being indeterminate as well. If this is right, then when such a sentence is in question, A(i) and A(ii) of the Minimal Truth Conditional Conception will express claims that are indeterminate, and so cannot correctly be asserted, provided that the biconditionals they contain are treated as material biconditionals. Fortunately, this complication need not affect our attempt to make sense of the skeptical solution. For one thing, on the present interpretation of that solution, meaning ascriptions are true, and hence not indeterminate. Thus the complication does not come into play. For another thing, we could always reinterpret the biconditionals used in the statement of the minimal conception so that such a biconditional is true just in case both sides are true, both sides are not true, or both sides are indeterminate. Such biconditionals may not be the usual material biconditionals, but they are perfectly legitimate. 31 See essay 15 in this volume.

442 • Essay Sixteen

terpretation of the argument according to which a skeptical solution to the paradox is to have any plausibility, one must come up with a stronger, more substantial, and more plausibly deniable interpretation of the “classically realist conception of normativity” and “classical realism about truth conditions” than we have come up with so far. Although Wilson does not provide fully explicit formulations of the needed principles, his discussion does contain strong hints that point in a promising direction. In the passage quoted above about the skeptic’s conception of what is necessary if I am to mean addition by ‘+’, Wilson speaks of my adopting a certain policy, which involves my singling out a certain function to govern the application of the term, and thereby establishing a standard of correctness.32 The picture he paints seems to be one in which my meaning addition by ‘+’ is explained by my having the relevant intentions about the addition function; I mean addition by ‘+’ because I have independently singled out the addition function and decided to use ‘+’ to stand for it.33 In general the idea seems to be that in order to bring it about that I mean something by a word, I must first pick out a property and form the right sort of intention. There is even a hint of this explanatory picture in one of the passages from Kripke quoted above, which is also cited by Wilson. The simplest, most basic idea of the Tractatus can hardly be dismissed: a declarative sentence gets its meaning by virtue of its truth conditions, by virtue of its correspondence to facts that must obtain if it is true. (Kripke 1982, 72; my underlined emphasis) Let us put these hints together in an attempt to extend the minimal truth-conditional conception of meaning into something more substantial. We may take the expanded conception to incorporate everything included in the minimal truth-conditional conception given above, plus the following claims. The Explanatory Version of the Truth-Conditional Conception of Meaning AE. (i) A speaker x who understands a sentence s, which expresses the proposition that P, does so because x knows that s expresses the proposition that P, and hence is true iff P. X’s knowledge of these facts explains, and is prior to, x’s understanding of s. 32 33

Wilson (1994, 368). This is spelled out explicitly, and in greater detail, by Wilson (1998; see in particular n. 8).

The Rule-Following Paradox • 443

BE.

CE.

(ii) If a sentence s, as used by a speaker x, means that P, then s has this meaning for x because x has the intention that s is to express the proposition that P, and hence to be true iff P. In general, a sentence that expresses a proposition is meaningful because speakers have a system of intentions the content of which assigns that proposition to the sentence. These intentions explain, and are prior to, the fact that the sentence means what it does. (i) A speaker x who understands a predicate ϕ that expresses the property of being F does so because x knows that ϕ expresses the property of being F, and hence (correctly) applies to an object o iff o is F. X’s knowledge of these facts explains, and is prior to, his understanding of ϕ. (ii) If x uses ϕ to mean the property of being F, then x does so because x has the intention that ϕ is to be understood as expressing the property of being F, and hence as (correctly) applying to an object o iff o is F. In general, a predicate is meaningful because speakers have prior intentions that the predicate is to be understood as standing for a certain property. Similarly for other categories of expressions.

If this is what classical realism is, then, presumably, the normativity premise used in Wilson’s version of the skeptical argument will have to be strengthened as well. This can be done by retaining the original wording of Ns, while strengthening the definition—Gov—of what it is for a set of properties to govern the correct application of a predicate. GovE. A set of properties P1–Pn governs the correct application of a term ϕ for x iff x has a grasp of P1–Pn which is prior to x’s understanding of ϕ, and ϕ has the meaning it does for x because x has the intention (has adopted the linguistic commitment) that ϕ is to be understood as (correctly) applying to an object o iff o has P1–Pn. With this understanding of Wilson’s notion of the government of a predicate by a set of properties, his “classically realist conception of normativity,” Ns, is strengthened so as to express the following claim: if x means something by a (general) term ϕ, then x does so because there is a set of properties P1–Pn, of which x has a prior grasp, and x has the intention (has adopted the linguistic commitment) that ϕ is to apply to an object o iff o has the properties P1–Pn. Since this claim is a consequence of the expanded truth-conditional conception of meaning, denying it amounts to rejecting

444 • Essay Sixteen

that conception. Moreover, on this strengthened interpretation, both Ns and the truth-conditional conception are substantive proposals that can be denied without absurdity.

An Argument against the Explanatory Version of the Truth-Conditional Conception of Meaning I will now present an independent argument against the explanatory version of the truth-conditional conception of meaning. The argument is limited in the sense that it does not purport to show that there are no sentences, expressions, and speakers that fit the explanatory account, but only that the account fails for some sentences, expressions, and speakers. Later, I will examine the scope of this failure, and extract the lessons to be learned from it, including the relationship between the position arrived at and Kripke’s skeptical solution to the problem posed by his original skeptical argument. The argument is based on examples like (4).34 (4) a. Pluto is a distant planet. b. ‘Pluto’ refers (in English) to Pluto. c. ‘Pluto is a distant planet’ is true (in English) iff Pluto is a distant planet. d. ‘Pluto is a distant planet’ means (in English) that Pluto is a distant planet. I believe all of these things. The reason I do is not that I have seen or had direct contact with the planet Pluto. On the contrary, my only contact with Pluto has been indirect, mediated by representations of it. The most important of these has been the name ‘Pluto’. In general, my beliefs about Pluto are linguistically mediated. I believe that Pluto is a distant planet because I have read or been told that it is. I have encountered the sentence ‘Pluto is a distant planet’ in reading or in conversation; I have understood the sentence, and I have accepted it. It should be noted that this understanding did not consist in my associating an identifying description with the name ‘Pluto’, and a descriptive proposition with the entire sentence. In acquiring the belief, I may well have come to believe descriptive propositions as a result of accepting the sentence. Still, my belief that Pluto is a distant planet cannot be identified with any such descriptive belief. Even if the description I associate with 34 The argument that follows is taken from my discussion of this issue in Soames (1989; essay 5 in vol. 1 of this collection). The next several paragraphs paraphrase the material found there (1989, 588–90).

The Rule-Following Paradox • 445

the name turns out to be inaccurate and doesn’t, in fact, pick out Pluto, my belief that Pluto is a distant planet is about Pluto (though the descriptive beliefs acquired at the same time may not be). Similarly, even if the descriptive beliefs turn out to be false, my belief that Pluto is a distant planet may remain true. Finally, the proposition I express by ‘Pluto is a distant planet’ is the same as the proposition that others express by the sentence, even if the descriptions we associate with the name are different. In short, the explanation of my belief that Pluto is a distant planet is based on three facts: (i) the fact that I accept the sentence ‘Pluto is a distant planet’, (ii) the fact that the sentence means (expresses the proposition) that Pluto is a distant planet, and (iii) the fact that I am a competent speaker, and hence understand the sentence. Moreover, my understanding the sentence is not a matter of my using it to express descriptive propositions that I might have come to believe on independent grounds. Analogous points can be made regarding other attitudes I might have taken toward the proposition. If I had come to wonder whether Pluto was a distant planet, or doubt that it was, my having that propositional attitude would have involved my having a certain attitude toward a sentence that expressed it. The point illustrated here is that the only epistemic connection I have with certain propositions is mediated through sentences that express them. This point can be extended to the other examples (4b–d). For instance, I believe the propositions expressed by the semantic claims (c) and (d). However, these beliefs do not explain my understanding of sentence (a). I don’t understand the sentence ‘Pluto is a distant planet’ because I have those beliefs. Rather, it is the other way around. My belief in the propositions expressed by (c) and (d) is (in large part) due to my understanding and accepting sentences (c) and (d), and to the fact that they mean what they do. Moreover, part of what it is for me to understand the semantic sentences (c) and (d) is for me to understand the nonsemantic sentence (a), which is a constituent of both. Thus, the direction of explanation in this case is not from semantic beliefs to linguistic competence, but from linguistic competence to semantic beliefs. If this is right, then a natural picture of language and linguistic competence associated with the explanatory truth-conditional conception is mistaken. According to the picture, we have the ability, prior to acquiring language, to apprehend objects, grasp properties, entertain propositions, and form beliefs. In setting up a language, we adopt certain conventions according to which individual expressions and sentences come to stand for these antecedently apprehended objects, properties, and propositions. Learning the language involves learning for each expression, and for each sentence, which antecedently available object, property, or proposition it stands for.

446 • Essay Sixteen

The most fundamental thing wrong with this picture is that in the case of at least some terms and sentences, we do not apprehend the objects, properties, or propositions they stand for prior to understanding the terms or sentences themselves. As a result, coming to understand these individual terms and sentences does not consist in searching through our stock of previously encountered objects, properties, and propositions to find the ones assigned to them. Rather, coming to understand the terms and sentences is a matter of satisfying conventional standards regarding their use. Just what these standards are is something that needs to be filled out in greater detail. However, whatever they are, once they are satisfied, one is counted not only as understanding new expressions and sentences, but also as cognitively relating to new objects, properties, and propositions.

The Scope of the Failure of the Explanatory Version of the Truth-Conditional Conception The above argument against the explanatory version of the truthconditional conception of meaning is modest, and does not pretend to be global in scope. In the example about Pluto, my beliefs are linguistically mediated, but they need not have been. Presumably, these beliefs could have been acquired through direct, nonlinguistic contact with the planet. Many of my beliefs about individuals are like this. For example, my beliefs about Plato are linguistically mediated; but they are also dependent on the fact that others had direct contact with the man, and passed the name down to me. Some of my beliefs about natural kinds also fit the above pattern. For example I believe that tungsten is a metal with some industrial use, which is found somewhere in South America, among other places. However I can’t give an accurate description that uniquely picks it out. Nevertheless, I satisfy whatever minimal criteria there are for understanding the term, in the sense of being able to use it with its conventional meaning. When I do use it in a sentence, the proposition I express is one that directly involves the natural kind itself, despite the fact that I don’t have any direct epistemic contact with the kind. In general, my beliefs about tungsten, including semantic beliefs analogous to (4b–d), are linguistically mediated, and are explained, in part, by my understanding and accepting various sentences containing the word ‘tungsten’. Thus, the explanation of my understanding of this word posited by the explanatory version of the truth-conditional conception of meaning is incorrect. How far, one wonders, could this critique be expanded? Are there linguistically mediated beliefs about objects with which no one is, or could be, in direct epistemic contact—e.g., mathematical objects and

The Rule-Following Paradox • 447

unobservable, theoretical entities in science? It seems reasonable to suppose that there are. After all, there are such objects, and we do have beliefs about them. It is not farfetched to suppose that we acquire these beliefs in virtue of understanding and accepting sentences about them. How this all comes about is, of course, a large and perplexing question. However, it is not implausible to think that, somehow, using a mathematical or scientific term in a certain way counts as referring to one or another mathematical or theoretical object, and expressing a belief about it, despite the fact that the referent of such a term cannot be “read off” our use of it, as well as the fact that we have no direct apprehension of the object in question. The upshot of these considerations is (i) that the explanatory version of the truth-conditional conception of meaning sketched above definitely is incorrect for certain sentences, expressions, and speakers (the Pluto and tungsten examples) and (ii) that it seems reasonable to suppose that the range of discourse for which the conception fails may turn out to be large and substantial (the scientific and mathematical cases). This is not to say that sentences in this range—whatever its precise boundaries— don’t have truth conditions; nor is it to say that speakers who understand those sentences don’t know what they are. The sentences in question do have truth conditions, and speakers who understand the sentences know them. But, in the cases we are concerned with, knowledge of truth conditions doesn’t explain understanding, and intentions to use sentences in such a way that they have certain truth conditions don’t explain how we endow sentences with meaning. Corresponding conclusions hold for predicates and other types of expressions. This raises the question of how linguistic competence and understanding is to be explained in the types of cases we have been considering. Here I must confess that I know of no satisfying general answer to this question that is clear and illuminating. Nevertheless, it seems plausible to suppose that, at least in some cases, one’s position in a larger linguistic community, and one’s mastery of the conditions of use imposed by that community, rather than any privately arrived at semantic beliefs, play a primary role in explaining one’s linguistic understanding. In this respect, the position we have arrived at resembles the skeptical solution proposed by Kripke’s Wittgenstein. However, the position also differs from that of Kripke’s Wittgenstein in that our critique of the explanatory version of the truth-conditional conception of meaning is not global, and does not call for a wholesale rejection of that conception. In fact, our critique does not apply to Kripke’s paradigmatic example of the plus sign. On the contrary, my understanding of ‘+’ conforms to the explanatory version of the truth-conditional conception.

448 • Essay Sixteen

According to that conception, if in the past I meant addition by ‘+’, I did so because I had a prior grasp of the addition function, and I formed the intention to associate the symbol with that function. What did that prior grasp of the function consist in, and how did I succeed in associating it with ‘+’? The obvious answer—which clearly guided Kripke in selecting this example in the first place—is that in the past I associated the plus sign with a rule, or algorithm, that determines the sum of any two numbers. For most of us, the rule was probably one we learned in the first grade, involving stipulated values for combining single digits, rules about carrying, and the like. However, the details are not important, and for present purposes we might as well assume that the rule I memorized was the standard number theoretic characterization (+). (+) For all natural numbers x and y, x + 0 = x & x + y'= (x + y)' Of course, we may assume that I understood the rule (+) as having a certain content. I interpreted it as saying that for any natural numbers x and y, x plus zero equals x and x plus the successor of y equals the successor of x plus y. In effect, I defined the function designated by ‘+’ to be that binary numerical function on the natural numbers which when applied to the pair consisting of a number x and zero always gives the value x, and when applied to the pair consisting of a number x and the successor of y always gives the successor of the value it gives to the pair of x and y. If this is right, then in the past I meant addition by ‘+’ because (i) I had the intention to use ‘+’ to stand for the function uniquely determined by the content of the rule (+), (ii) the content of that rule determined the addition function, and (iii) I understood that content. All of this is fully in accord with the expanded truth-conditional conception of meaning. Thus, what that conception tells us about the meaning of ‘+’ is correct. The plus sign did, and does, stand for the addition function; and it did, and does, so because of beliefs and intentions regarding its use which explain, and are prior to, my understanding of the symbol itself. It is undeniably true that this explanation of my past use of ‘+’ presupposes that the rule (+) had a certain content, which I understood. In particular, it presupposes that in the past I understood the successor symbol ‘'’ to stand for the successor function, ‘=’ to stand for the identity relation, the numeral ‘0’ to denote the number zero, and the quantifiers to range over the natural numbers. It is also true that we can raise the same question about my understanding of these terms that we did regarding ‘+’—namely, did my understanding of them conform to the explanatory truth-conditional conception of meaning? The crucial point to notice here is that for the purpose of determining whether my use of ‘+’ conformed to that conception, it doesn’t matter how we answer this question.

The Rule-Following Paradox • 449

Perhaps no further rules guided my use of the successor symbol, ‘'’. Perhaps in addition, there is no other way in which I could be credited with an independent, prior grasp of the successor function, and a decision to use ‘'’ to stand for it. If so, then my understanding of the successor symbol cannot be explained on the model of the explanatory truth-conditional conception of meaning, but rather must be explained in some other way.35 But no matter how it is explained, it remains true that I did mean successor by ‘'’. That much must be granted by any nonrevisionary account of our ordinary meaning ascriptions—any account that acknowledges that they express facts in the minimal sense (M). But then, given that the terms occurring in the rule used to define ‘+’ did mean what I took them to mean, and were used to express beliefs independent of, and prior to, my use of ‘+’, one must grant that my understanding of the plus sign did meet the conditions imposed by the explanatory truth-conditional conception. This point illustrates an obvious, but important lesson. No defensible grounds exist for a global rejection of the explanatory role of truth conditions for sentences, or of the explanatory character of the reference and application conditions for individual terms. Some sentences do have explanatory truth-conditions in the sense required by the explanatory truth-conditional conception, and some terms have explanatory reference or application conditions. In light of this we need to ask, “What, if anything, does the limited skepticism that we have arrived at regarding truth conditions have to do with Kripke’s skeptical solution to the rulefollowing paradox?”.

The Anti-Truth-Conditional Version of the Skeptical Solution Revisited According to our reconstruction of Wilson, Kripke’s skeptical argument is a reductio of the explanatory version of the truth-conditional conception of meaning, and the skeptical solution is an alternative conception according to which our meaning what we do by our words is not explained by 35 For Kripke’s Wittgenstein, it is part of the skeptical solution that a certain range of linguistic behavior, together with a certain pattern of activity relating that behavior to my nonlinguistic activities, simultaneously constitutes my understanding the successor symbol, my knowing that it stands for the successor function, and my possessing various elementary arithmetical beliefs involving the function. The proponent of this view rejects the idea that understanding the symbol, and possessing the relevant arithmetical and metalinguistic beliefs, are themselves mental states that causally explain the corresponding linguistic and nonlinguistic behavior. Instead, on the picture he advocates, the linguistic and nonlinguistic behavior provide the only warrant possible for the claim that one understands the successor symbol, and has the relevant metalinguistic and arithmetical beliefs.

450 • Essay Sixteen

any prior intentions to use them to stand for antecedently apprehended objects, properties, or propositions. Having shown that this interpretation of Kripke has the virtue of locating the target of the skeptical solution as a contentious view about meaning that really does have to be rejected, we are now in a position to see the problems with the interpretation. First, Wilson quite reasonably interprets Kripke’s Wittgenstein as arguing that no sentences have “classically realist truth conditions” and no expressions have “classically realist” reference, application, or satisfaction conditions.36 However, as we have just seen, some expressions, such as the symbol ‘+’ and the word ‘addition’, do have reference conditions that play a significant role in explaining how they are used and understood—i.e., conditions that fit the explanatory truth-conditional conception. By extension, some sentences containing these expressions have similarly explanatory truth conditions.37 This suggests either that the skeptical solution fails in an utterly obvious way to establish what it sets out to achieve, or that its real target is not the explanatory truth-conditional conception of meaning characterized above, but something else. Second, Kripke’s Wittgenstein does not attempt to provide the sort of argument that is required to show the incorrectness of the explanatory version of the truth conditional conception. The key claim made by this conception is that one kind of intentional fact—namely, semantic beliefs and intentions relating sentences and expressions to propositions, properties, and objects—is prior to, and part of the explanation of, another kind of intentional fact—namely, understanding sentences and expressions, and meaning something by them. To show the incorrectness of this claim one must argue, as I did above, that in at least some cases the alleged explanatory priority does not exist (either because both types of fact are on a par, or because the order of priority is reversed). Kripke’s Wittgenstein does not do this. Instead, when it is suggested that my past meaning of ‘+’ may have been determined by beliefs or algorithms I associated with it, he, along with the skeptic, dismisses the suggestion by pointing out that the same skeptical question can be raised about those beliefs or algorithms, since they too have content. But, from the point of view of the explanatory version of the truth-conditional conception of meaning, this is beside the point. Since anything intentional has content, one can always ask for the nonintentional basis of that content. This observation, so prominent in Kripke’s skeptical argument, is irrelevant to 36

See Wilson (1994, 374 and 375). For example, I understand the sentences 2 + 2 > 3 and In the past I meant addition by ‘+’ because I know that they express the propositions that two plus two is greater than 3, and that in the past I meant addition by ‘+’; moreover my knowledge that they express these propositions is due in part to the fact that I have an independent grasp of the addition function that explains my understanding of ‘+’ and ‘addition’ which occur in these sentences. 37

The Rule-Following Paradox • 451

the issue of whether some intentional facts are explanatorily prior to other intentional facts. Moreover, no attempt is made anywhere else in the skeptical argument to show that semantic beliefs and intentions fail to have the priority required by the explanatory version of the truthconditional conception. The reason for this is that Kripke’s skeptic is simply not concerned with such questions of priority. He is looking for the nonintentional basis of all intentional facts, and suggesting that there isn’t any. Because of this the skeptical argument cannot be seen as a reductio ad absurdum of the explanatory truth-conditional conception—or, at any rate, it cannot be seen as a reductio of the special explanatory part of the conception that distinguishes it from the minimal truth-conditional conception. But to say this is just to say that Kripke’s actual skeptical argument does not conform to our reconstruction of Wilson’s interpretation. The third problem with the interpretation is that it cannot make sense of the extended discussion of the dispositional analysis. According to the interpretation, the skeptic takes the explanatory version of the truthconditional conception of meaning for granted, and relies on it for the following strengthened version of Wilson’s “classically realist conception of normativity.” The Strengthened Version of the “Classically Realist Conception of Normativity” If x means something by a (general) term ϕ, then x does so because there is a set of properties P1–Pn, of which x has a prior grasp, and x has the intention (has adopted the linguistic commitment) that ϕ is to apply to an object o iff o has the properties P1–Pn. (This condition is stated for predicates, but a similar condition would be stated for other types of expression—e.g., function signs, like ‘+’.) According to our reconstruction of Wilson, the skeptic treats this as a central, unargued assumption in the skeptical argument. However, if that were so, then he would have no need to argue against the dispositional analysis at all. On the one hand, the only candidates for meaningdetermining facts allowed by the assumption are those that determine explanatorily prior semantic beliefs and intentions governing ‘+’. On the other hand, the dispositional analysis does not even attempt to specify such facts. Thus, if the skeptic were taking the assumption for granted, he could simply dismiss the dispositional analysis as not addressing the problem of meaning, as he conceives it. Kripke’s skeptic does not do this; rather he sees the dispositional analysis as requiring extensive, independent refutation. This is evidence that his argument does not assume the strengthened version of the “classically realist conception of normativity,” and cannot be

452 • Essay Sixteen

regarded by the proponent of the skeptical solution as a reductio ad absurdum of the explanatory version of the truth-conditional conception of meaning.

Is There a Single, Coherent Interpretation of Kripke’s Text? So far we have considered two interpretations of the skeptical argument and the skeptical solution. According to the first interpretation, the skeptical argument purports to show that there are no meaning facts, in even the minimal sense (M) of “facts”; the skeptical solution maintains that meaning ascriptions are nondescriptive, and do not purport to express truths. On this interpretation, the skeptical argument proves to be a highly interesting, though fundamentally flawed, attempt to establish a truly radical conclusion; the skeptical solution, on the other hand, is incoherent, and represents no serious alternative to the truth-conditional conception of meaning. According to the second interpretation, the skeptical argument assumes from the outset that if there is such a thing as meaning, then facts about meaning must conform to the explanatory version of the truth-conditional conception; the argument purports to show that, on this assumption, there is no such thing as meaning. On this interpretation the skeptical solution treats the skeptical argument as a reductio ad absurdum of the explanatory version of the truth-conditional conception of meaning, and tries to sketch an alternative conception. Unlike the first interpretation, this interpretation is able to view the skeptical solution as containing at least a modicum of sense; it is seen as an attempt to refute a contentious conception of meaning that really is incorrect. Unfortunately, as we have seen, this interpretation cannot, in the end, be sustained. Whatever may be assumed about meaning in the skeptical argument, and whatever may be the target of the attempted refutation in the skeptical solution, it is not the explanatory version of the truth-conditional conception of meaning. Both interpretations we have considered make sense of part of the text, while failing to make sense of the text as a whole. Can we do better? Let us consider a third interpretation that combines important elements of the previous two. Recall the formulation of the skeptical argument according to the first interpretation. P1. If in the past it was a fact that I meant addition by ‘+’, then either: (i) this fact was determined by nonintentional facts of the following sorts: ....................

The Rule-Following Paradox • 453

or (ii) the fact that I meant addition by ‘+’ was a primitive fact; that is, a fact not determined by nonintentional facts. P2. It is not the case that nonintentional facts of type (i) determined that I meant addition by ‘+’. P3. It is not the case that the fact that I meant addition by ‘+’ was a primitive fact. C1. Therefore, in the past there was no fact that I meant addition by ‘+’. On the analysis given in my “Skepticism about Meaning: Indeterminacy, Normativity, and the Rule-Following Paradox,” this argument suffers from equivocation about what is meant by ‘determine’. On the one hand, we may say that a set F of truths determines a claim M iff it is possible to demonstrate M from F—i.e., iff M is an a priori consequence of F. On this interpretation I argue that P1 and P2 may be true, but P3 is false. On the other hand, we may say that a set F of truths determines a claim M iff M is a necessary consequence of F. I argue that on this interpretation P1 and P3 are true, but P2 is false. Let us concentrate on the epistemological version of the argument—in which the determination relation requires the ability to demonstrate something from the underlying facts. One of the main reasons for thinking that, on this interpretation, P2 may be true is brought out by the difficulty one encounters in specifying a set F of nonintentional facts that would satisfy the corresponding (epistemological) version NE of the normativity requirement. NE. If a set F of facts determined that (in the past) one meant addition by ‘+’, then knowing the facts in F would, in principle, provide one with a sufficient basis for concluding (demonstrating) that one ought to give the answer ‘125’ to the question ‘What is 68 + 57 ?’, and similarly for all other questions of this type, provided one intends to use ‘+’ with the same meaning it had in the past. What makes this constraint so powerful is that for every set of nonintentional facts that one can think of, it appears that knowing those facts is not sufficient for demonstrating which answers to the relevant questions about the application of ‘+’ are correct, and hence ought to be given. With this in mind, let us reformulate the skeptical argument so that it explicitly assumes both NE and the claim that meaning facts are not (epistemologically) primitive. The argument begins with a supernormativity premise, Nplus, incorporating these assumptions. Nplus. If in the past I meant addition by ‘+’, then there is a set of nonintentional facts such that knowledge of those facts

454 • Essay Sixteen

would, in principle, provide me with a sufficient basis for concluding (demonstrating) that I ought to give the answer ‘125’ to the question ‘What is 68 + 57?’, and similarly for all other questions of this type, provided I intend to use ‘+’ with the same meaning I did in the past. On the new interpretation we are constructing, Kripke’s skeptic assumes Nplus from the outset. In accord with Nplus, he insists that the facts he is looking for must be nonintentional, since otherwise his skeptical worries would apply to them. Having limited the field in this way, he spends most of chapter 2 arguing for the following antireductionist premise AR. AR. There are no nonintentional facts such that knowledge of them would, in principle, provide me with a sufficient basis for concluding (demonstrating) that I ought to give the answer ‘125’ to the question ‘What is 68 + 57?’, and similarly for all other questions of this type, provided I intend to use ‘+’ with the same meaning I did in the past. Since the skeptic accepts both Nplus and AR, he concludes that I didn’t mean addition by ‘+’ in the past. By analogous reasoning he concludes that I didn’t mean anything else by ‘+’ either, and that, in general, no one ever means anything by any word. The proponent of the skeptical solution, on the other hand, does not draw any of these skeptical conclusions. Although he agrees with the skeptic about AR, he rejects the skeptical conclusions, and treats the argument as a reductio ad absurdum of the supernormativity premise Nplus. This interpretation has two main virtues; it makes reasonable sense of the skeptical argument, and it treats the skeptical solution as rejecting a misguided conception of normativity that fails not only for ‘+’, but for all expressions. However, it also suffers from two fundamental defects: (i) there is no reasonable version of the truth-conditional conception that is committed to the supernormativity principle Nplus; hence rejection of that premise does not motivate the sort of attack on the truth-conditional conception, and construction of an alternative account of meaning, that is advertised in Kripke’s skeptical solution. (ii) the claim Nplus, for which we have a plausible reductio, has very little independent plausibility; hence a solution that rejects it hardly qualifies as even mildly surprising, let alone skeptical. So, in the end, we have three interpretations. Each has important virtues; each has fundamental defects. None is completely satisfying. The reason for this, I suspect, is that, like many leading works in philosophy, there is no single, coherent line of argument that runs through the whole of the text. Rather, Kripke’s text is a fabric through which the different

The Rule-Following Paradox • 455

threads of these three interpretations are woven together. If this is right, then the proper response is give up as illusory the attempt to find a single unifying interpretation, to distinguish instead several illuminating partial interpretations, and to discover what there is to learn from each. That is what I have tried to do.38

References Boghossian, Paul. 1989. “The Rule-Following Considerations.” Mind 98:507–49. ———. 1990. “The Status of Content.” Philosophical Review 99:157–84. Burge, Tyler. 1979. “Individualism and the Mental.” In Studies in Metaphysics, ed. Peter A. French, Theodore E. Uehling, and Howard K. Wettstein, 73–121. Minneapolis: University of Minnesota Press. Byrne, Alex. 1996. “On Misinterpreting Kripke’s Wittgenstein.” Philosophy and Phenomenological Research 56:339–43. Kripke, Saul A. 1982. Wittgenstein on Rules and Private Language. Cambridge: Harvard University Press. Putnam, Hilary. 1975. “The Meaning of ‘Meaning’.” In Language, Mind, and Knowledge, ed. Keith Gunderson, 358–98. Minneapolis: University of Minnesota Press. Soames, Scott. 1989. “Semantics and Semantic Competence.” In Philosophy of Mind and Action Theory, ed. James E. Tomberlin, 575–96. Philosophical Perspectives 3. Atascadero, Calif.: Ridgeview. ———. 1999. Understanding Truth. New York: Oxford University Press. ———. 2003. Philosophical Analysis in the Twentieth Century. 2 vols. Princeton: Princeton University Press. Wilson, George. 1994. “Kripke on Wittgenstein and Normativity.” Midwest Studies in Philosophy 19:366–90. ———. 1998. “Semantic Realism and Kripke’s Wittgenstein.” Philosophy and Phenomenological Research 58:99–122. Wright, Crispin. 1984. “Kripke’s Account of the Argument against Private Language.” Journal of Philosophy 81:759–78.

38 Early versions of this essay were given in January 1996 at the Instituto de Investigaciones Filosóficas, Universidad Nacional Autónoma de México, in Mexico City, and in March 1996 at Reed College. The essay was later revised and expanded, and much of the new material was presented in my graduate seminar at Princeton in the fall of 1996. I would like to thank Chrys Gitsoulis, Jim Pryor, and Alex Byrne for reading and commenting on a draft of this essay.

This page intentionally left blank

Index

actuality operator, 9, 17–19, 157–161, 206n24, 208, 220, 255, 256, 263, 272, 277–80, 283; and the adverb ‘actually’, the adjective ‘actual’, and the predicate ‘is actual’, 279–80, 291–92; and the contingent a priori, 281, 288–291; and the necessary a posteriori, 221–22, 249–50, 269, 280–81, 285–87. See also transparency/nontransparency Almog, Joseph, 33, 50n17 analytic truth, 186; relation to a priori and necessary truth, 166–67, 183, 186–87, 203–4, 243 anaphoric pronouns, 7–8, 111–32; as bound variables, 113, 117–20, 132–34; and the predicate-abstraction analysis, 60, 118–20, 132–34; and reference or sense inheritance, 115–17 assertion, 14–16, 211–41, 335; analysis of, 59n29; and knowledge, 367–68; the norm of, 25, 365–68; and pragmatic enrichment, 241; rules governing, 236–41; and semantic content, 34–36, 111, 211–12, 233, 241, 292; Stalnaker’s rules for, 214–16, 223–24, 225, 252; and truth, 365–68 Barwise, Jon, 33, 40, 41n4, 44, 45n7, 58n28, 73 Bealer, George, 275 Benacerraf, Paul, 301 Boghossian, Paul, 421–23n7, 432n18, 433n21 Braun, David, 161n29 Buhler, Alex, 301 Burge, Tyler, 428n13 Burger, Alan, 313n18 Burgess, John, 301, 306n6 Byrne, Alex, 195n9, 401n16, 436n22, 455n38 Carnap, Rudolf, 38, 40n3, 87, 282, 283, 297, 307n7, 316n23 causal descriptivism, 249

c-command, 112, 117, 118, 120–22, 132–34 Chalmers, David, 14, 183–85, 187, 207n25, 252, 255n14, 258, 264, 270–73 Church, Alonzo, 98n21 conceivability, 10, 12–13, 182; and the de re, 181, 237; and possibility, 167, 168–70, 179–81, 184, 185, 186, 237 consequence, 5; a priori, 27, 238, 286–87, 290–91, 296–97, 324–29, 394–96, 398, 400–402, 407–8; and argumentative commitment, 336–38; and deflationism about truth, 337–38; logical, 39, 78, 337–38, 407; model theoretic, 5, 76–79; necessary, 27, 34–35, 39, 79, 238, 324–29, 335–37, 340–41, 349, 397–402, 407–8; truth-conditional, 5, 78–80. contingent a priori, 18–19, 247, 248, 252, 281, 288–91, 297; and the actual worldstate; 288–91; two-dimensionalist explanation of, 256–57, 259–60, 269–70, 272, 275 Davidson, Donald, 21, 315–16n23, 317 Davies, Martin, 122, 222n8 descriptions, 8; definite, 41, 43–45; indefinite, 41, 45; reference-fixing, 49–50, 178–79, 182, 245–49, 264–69, 271, 274; rigidified, 9, 156–62, 183, 184, 208–9, 226n11, 249, 255, 264, 268, 269–70, 272, 274; wide-scope, 8–9, 142–56 determination, 27 direct reference, 3, 36, 37, 38, 46, 55n24, 58, 61, 64–69, 72–73, 76, 115–17, 123, 166, 196, 234, 245–47, 293, 294 Donnellan, Keith, 59n29, 149n11 ‘dthat’ operator, 9n3, 162, 183, 197, 206n24, 208, 255, 256, 269–70, 272, 274 Dummett, Michael, 24, 141n2, 142–43n5, 147–49n11, 149n12, 150n13, 152n14, 154n16, 315n23, 364–68, 370, 424

458 • Index Edelberg, Walter, 4–5, 74–79 essentialism, 10, 15, 166–67, 168, 170–71, 181–82, 185, 187, 227, 237, 243, 247, 252, 260, 262, 282, 285 Evans, Gareth, 122, 148–49n11, 152n14 Field, Hartry, 20–21, 307n7, 308, 309–13, 316–20 Fitch, G. W., 161n29 Fodor, Jerry, 5–7, 81–109 Friedman, Michael, 410–11n28, 413n30 Frege, Gottlob, 68, 247 Frege’s Puzzle, 4, 247, 269, 270–71, 275 Gadd, Nathan, 298n19 Gitsoulis, Chrys, 435n38 Glanzberg, Michael, 25, 364–68, 370 Gomez-Torrente, Mario, 192n4 Grice, Paul, 292n16 Harman, Gilbert, 301, 313n16, 338n12 Hempel, Carl, 307n7 Hilbert, David, 195n9, Humberstone, Lloyd, 222n8 indeterminacy of reference, 310, 319–20, 402; and eliminativism about reference, 409n26 indeterminacy of translation, 26, 402–18; and eliminativism about meaning and mind, 422–23; and equivocation about determination, 406–14, and indeterminacy of meaning, 402, 406; and observation sentences, 403–4; and occasion sentences, 403–4; and physicalism, 406–11, 413–14; and standing sentences, 404; and stimulus meaning, 402–6, 408; and underdetermination by evidence, 402–5; and underdetermination by physics, 406–9, 411–14. See also skepticism about meaning. Jackson, Frank, 16, 184, 185, 207n25, 248n7, 252, 258, 264, 265, 267–69 Jubian, Michael, 141n3 Kaplan, David, 8n3, 33, 37, 39, 40n3, 62n33, 66n36, 90, 98n21, 103, 146n8, 158, 160, 162, 165, 197, 203n19, 215, 243–45, 251, 253, 255, 268, 269, 278 Kazmi, Ali, 33, 81, 89n9, 95, 187n19, 298n19

King, Jeff, 380n21 Kripke, Saul, 8–13, 26–29, 48, 49n11, 50, 53n22, 60n31, 98n21, 139, 141, 147–49n11, 156, 158, 162, 163, 165–87, 190, 197, 199n13, 209, 218, 223, 237, 243–45, 247, 249, 265–67, 269, 281, 316n24, 317n25, 320, 369n12, 385–402, 407, 414, 416–65 language of thought, 5–6, 86, 92 Lewis, David, 182, 249, 265n21, 301, 316n24, 317 Linsky, Bernard, 191n4 MacFarlane, John, 380n21 Manley, David, 191n5, 298n19 Marks, Charles, 81 meaning, 2, 78–79; and analyticity, 166, 186–87; conceptions of, 33–38, 67–71; and conceptual analysis, 2, 10, 166, 186–87, 243; as source of apriority and necessity, 166–67, 183, 186–87; and truth conditions, 2–5, 20–21, 28–29, 63, 314–15, 320; and truth-supporting circumstances, 33–71, 72–80. See also assertion; semantic content; structured propositions; transparency/nontransparency. modal argument, 8–9, 139–63, 245, 249 Montague, Richard, 117, 118n6 Moravcsik, Julius, 33 Moruzzi, Sabastiano, 380n21 names, 3, 7, 36–37, 72–73, 80, 89n9, 90, 92, 103, 114n5, 115, 171; acquaintance constraint on, 274; and descriptions, 8–10, 139–63, 183, 184, 243–45, 247, 249, 255, 263–74; fixing the reference of, 49–50, 192, 197–98, 202–3, 245; as quantifiers, 116–17; and substitutivity, 46–47, 50n18, 54–57, 125, 161. See also direct reference; rigid designation natural kinds, 13–14, 171, 182, 189–91, 363–64; essential properties of, 181–82; and explanatory kinds, 194; individuation of, 195; knowledge of, 14, 198–201, 205, 207, 208; manifest, 189; and substances, 193 natural kind terms, 13–14, 67, 90–92, 103, 166, 171, 175, 177–79, 189, 245, 363–64; and descriptions, 183, 184, 190, 196, 205, 207–9, 255, 265–72, 274, 376n14;

Index • 459 fixing the reference of, 190–92, 197–98, 207–9, 363–64; and the necessary a posteriori, 189–90, 192–201, 204–5; and predicates containing them, 171, 189–92, 196; rigidity of, 178, 182, 196; semantic analysis of, 196, 205 Neale, Stephen, 122 necessary a posteriori, 9–18, 167–87; and the actual world-state, 221–22, 285–87; essentialist route to, 170–72, 180–81, 183, 184, 185, 186, 187, 247; and mind-body identity, 177–82, objection to, 174, 248; second route to, 172–77, 183–84; two-dimensionalist explanation of, 184–86, 223–36, 252–55, 256–57, 259–60, 261–62, 269–70, 275; two routes to, 9–11, 175 Neurath, Otto, 307n7 normativity requirement, 27–28 partially defined predicates, 23–26, 338, 340–41, 355; and absolute undefinedness, 357; the alleged impossibility of, 366–68, 370; and context-sensitivity, 374–77; (determinate) extensions and antiextensions of, 23, 340–41, 362, 375–76; and the “laws” of classical logic, 26, 370–74, 379–80; the possibility of, 25, 362–80; and properties, 372, 375–78; and relative undefinedness, 356–57; rules governing, 340; and superdeterminateness, 358–59; and superundefinedness, 358–59; and supervaluationism, 371–72; and the truth predicate, 338; undefined instances of, 340–41, 346n6, 368–74; and weak partiality, 356–57; and the unknowability of the undefined, 368–70. See also vague predicates Perry, John, 33, 40, 41, 44n4, 45n7, 57, 58n28, 68, 73, 98n21 Popper, Karl, 307n7 propositional attitude ascriptions, 2–8, 72; and anaphora, 111–34, and disquotation, 48, 114, 173–77; and explanation, 57, 96–100, 101–2, 108–10; interaction with modality, 154–56, 159–61, 162–63, 263–64; and linguistic competence, 67, 98, 445–46; and mental representations, 5–7, 81–109; and mental states, 58; quantification into, 46–47, 125–33, 246;

and semantic content; 33–71, 72–80, 113–14, 160–61, 246; and sentential attitudes, 56–61, 98, 114. See also transparency/nontransparency propositional concept, 214–15, 224–26, 227–28, 230, 236, 241, 253, 255, 260, 262 propositions, 4, 145, 335; about the actual world-state, 221–22, 278, 281, 283, 285–87, 292–95; annotated, 126–33; and partially structured intensions, 64–71; and truth conditions, 63; and truthsupporting circumstances, 3–5, 33–71, 72–80, 97, 211, 213, 236, 253, 257; singular, 59n29, 114–15, 118–19, 147, 150, 163n31, 176, 177, 204, 206n24, 235–36, 239, 281, 295; structured (Russellian), 4, 38, 61–64, 69–70, 97, 118n6, 126, 147, 150, 240, 246, 282 Pryor, Jim, 200, 207n25, 396n14, 455n38 Putnam, Hilary, 90n11, 104n27, 165, 202–3, 243–44, 315n23, 428n13 Quine, Willard, 26, 166, 244, 247, 304, 310, 313, 319, 385, 402–14, 422 reference borrowing, 266 Reichenbach, Hans, 307n7 Rieber, Steven, 234n14 Richard, Mark, 33, 50n15, 53–54, 55n24, 56n24, 57, 66n36, 126–33, 154n16, 329–31, 336–38 rigid designation, 139–42, 147–49n11, 152–53, 156–63, 166–67, 178–80, 182, 223, 231, 236, 246, 253, 264; and simulated rigidity, 142–43, 247 Russell, Bertrand, 246, 247, 269 Salmon, Nathan, 33, 55n24, 56n24, 57n27, 60n30, 62, 97n20, 98n21, 104n27, 118n7, 149n11, 158, 234n14, 246, 281, 283n10, 368n10, 379n19, 380n21 semantic content, 3–5; attempted reduction of, 87n7, 92, 103n26; broad, 7, 102–3; and character, 49–52, 55n24, 66n36, 103, 104n27, 203n19, 205–6, 215, 225–26, 245, 251, 253, 255, 268, 278; narrow, 7, 101–9; and truth conditions, 33–71, 72–80. See also propositions semantic theory, 34–38, 72, 77–79; two senses of, 78–79

460 • Index Sinnott-Armstrong, Walter, 301 situations, 3, 40n3, 41, 73 skeptical solution, 28–29, 416, 423–55; the anti-truth-conditional version of, 437–44, 449–52; and the classical truth-conditional conception of meaning, 424, 428, 433, 436, 437, 440, 442; and an explanatory conception of truth-conditions as meaning facts, 428–29, 442–52; and a minimal conception of truth conditions and meaning facts, 428–36, 440–41; the nondescriptive version of, 430–36; and the radical skeptical conclusion, 426; and the social conception of meaning facts, 428; and the ultimate disunity of Kripke’s text, 452–55 skepticism about meaning, 26–28, 385–414; and the burden of proof, 400–401, 414; and the determination relation, 385–88, 392–94, 396–98, 401–2; and the dispositional analysis, 389–99, 418–19, 451–52; and elminativism about meaning and intentional content, 420–23; and equivocation about determination, 385, 402, 414; and the error and infinitude objections, 390–91, 419; and images or experiences, 419; and Kripke’s skeptical argument, 385–402, 416–23, 449; and the normativity objection, 391–99, 419, 439–41, 443, 451, 453–54; as pragmatically self-defeating, 420–23; and skepticism about mental content, 387–88, 389n6, 418; and supervenience of the intentional on the nonintentional, 399–401; and truth, 421. See also indeterminacy of translation sorites paradox, 343–46; and the strengthened sorites, 359–60 Sosa, David, 141n2 Speaks, Jeff, 187n19 Stalnaker, Robert, 14–16, 211–41, 248n9, 252–55, 260–63, 281, 284 Stanley, Jason, 141n3, 149n11 Strawson, Peter, 292n16 Sterelny, Kim, 92n14 Strong Descriptive Origin and Justification, 177, 184 Strong Disquotation and Justification, 173–77, 184 Tarski, Alfred, 19–21, 301–2, 306–13, 316, 317–21, 334

Thau, Mike, 156n20, 395–96n14 transparency/nontransparency, 165, 176, 186, 213, 233–34, 238–41, 243 truth, 21–22, 365; and argumentative commitment, 337–38; deflationism about, 21–22, 304, 306, 308, 321, 323–38; determinate, 338, 341; inflationism about, 22, 333–34; and meaning 20–21; and nonfactualism, 329–38; and paradox, 306, 320, 323, 326, 338; and physicalism, 307, 309–13, 317–20; and semantic ascent, 304–6, 326, 421; supervaluationist, 372; Tarski’s theory of, 19–21, 301–2, 306–7, 308–13, and truth-value gaps, 338, 340, 364–66, 371 truth conditions, 2–3, 21; and meaning, 2–4, 20–21, 28–29, 314–15, 424–25; of propositions, 63–64; and truth-supporting circumstances, 3–5, 33–71, 72–80 Twin-Earth cases, 90–91, 101–6, 428 two-dimensionalism, 13, 183–86, 244; ambitious, 252; and attitude ascriptions, 256, 258, 259, 263–64, 270–74; benign, 251; and the diagonal proposition, 15, 223–24, 225, 228, 254, 261; hybrid, 252, 270–74; and mind-body identity, 183; the pragmatic (discourse) version of, 15–17, 211–41, 252–55, 260–63; and primary intension (proposition), 183, 184, 187; and the problem of the de re, 205, 207, 213, 217–21, 227, 229–30, 234–36, 250–51, 261–62, 270, 272, 274; and secondary intension (proposition), 183, 184, 187, 250–51; strong, 209, 250–51, 255–58, 263–64; weak, 205–7, 251, 258–60, 264–70 vague predicates, 1, 23–26, 340–61, 362–64; and context sensitivity, 25–26, 341–45, 358–60, 362–63, 377–80; default settings of, 341, 362, 377–78; (determinate) extensions and antiextensions of, 23, 340–41, 344–48, 362; and the ‘determinately’-operator, 346–55; higherorder, 23–24, 346–61; and ignorance, 26, 377–79; and indeterminacy, 24–26, 348–57, 377; and the “laws” of classical logic, 379–80; and linguistic competence, 24, 360; and penumbral truths, 380; and rules governing, 341–43, 348–51, 362–64; and sharp lines, 243–345, 347, 358–61,

Index • 461 376–77; and the ‘superdeterminately’operator, 358–59; super-undefined instances of, 358–59; undefined instances of, 340–44, 346, 348–53, 355–57; and weak vagueness, 357. See also partially defined predicates Wettstein, Howard, 33 what-is-said, 2n1, 33–34, 48, 51n19, 67, 78, 106, 335, 421; indeterminacy of, 372. See also assertion Williamson, Timothy, 25, 367–68, 377–78 Wilson, George, 395n13, 431, 437–43, 449–51 Wittgenstein, Ludwig, 26, 186, 385n1, 396n14, 416, 424–25

world-state, 1, 10, 167, 236; actual, 157–62, 170, 213–14, 219–21, 222, 252, 269, 277, 281, 282, 288–91, 292; epistemically possible, 10, 15, 17, 18, 167, 168–70, 181, 184, 185, 186, 213, 230–33, 237–38, 248, 262, 277, 281–83, 285, 288–91, 296; knowledge of, 17–19, 216–21, 222, 281, 285–91, 293–96; logically possible, 39, 45n8, 68; metaphysically possible, 9, 10, 15, 17, 18, 35, 37, 39, 43, 44, 68, 73, 76, 78, 157–62, 167, 168–70, 180–82, 184, 185, 186, 213, 229, 237–38, 248, 261–62, 277, 281–83, 285, 288–91, 296; nature of, 281–85; and state descriptions, 38–40, 282, 297–98 Wright, Crispin, 432n18, 433–34n21