We can see the problems for the proposal more clearly by distinguishing the semantic from the metasemantic. Semantics facts are facts of the kind we attempt to systematize in giving a systematic compositional semantic theory for a language, facts as to what its expressions mean. Metasemantic facts are the nonsemantic facts on which the semantic facts supervene. The distinction is rough but clear enough to be workable. Thus the fact that “horse” applies to horses is semantic, not metasemantic; the fact that utterances of “horse” are often caused by horses is metasemantic, not semantic.24 Similarly, the fact that “zzz” means a short sleep is semantic, while the fact that it was introduced by someone saying “A zzz is a short sleep” is metasemantic. The semantic theory takes no notice of the act of stipulation, only of its outcome – that a given expression has a given meaning. The act of stipulation makes the sentence true by making it have a meaning on which it is, in the quite ordinary way, true. My saying “A zzz is a short sleep” did not make a zzz be a short sleep, because that would be to make a short sleep be a short sleep, and my saying “A zzz is a short sleep” certainly did not make a short sleep be a short sleep. In particular, since there were many short sleeps before I was born, there were many zzzes before I was born, independently of my later actions. At best, my saying “A zzz is a short sleep” made “zzz” mean a short sleep, and therefore “A zzz is a short sleep” mean that a short sleep is a short sleep. This is simply the standard semantic contribution of meaning to truth, just as for synthetic truths. The peculiarity of the case is all at the metasemantic level; the use of stipulative definitions as paradigms does not yield a semantic notion of analyticity. Making “zzz” mean a short sleep helps make “A zzz is a short sleep” true only because a short sleep is a short sleep. “A short sleep is a short sleep” is a logical truth, but we have still been given no reason to regard logical truths as somehow insubstantial. The use of stipulative definitions as paradigms of analyticity does not justify the idea that analytic truths are in any way insubstantial.
My stipulation may smooth my path from knowing the logical truth “A short sleep is a short sleep” to knowing the Frege-analytic truth “A zzz is a short sleep,” but of course that does not explain how I know “A short sleep is a short sleep” in the first place.
The metaphysics and semantics of analytic truths are no substitute for their epistemology. If their epistemology is as distinctive as is often supposed, that is not the outcome of a corresponding distinctiveness in their metaphysics or semantics. It can only be captured by confronting their epistemology directly. We therefore turn to epistemological accounts of analyticity.
Notes
1 1 To give just one example, even Jack Smart, whose work robustly engages the nature of the non-linguistic, non-conceptual world and who described metaphysics as “a search for the most plausible theory of the whole universe, as it is considered in the light of total science” (1984: 138), could also write that philosophy is “in some sense a conceptual inquiry, and so a science can be thought of as bordering on philosophy to the extent to which it raises within itself problems of a conceptual nature” (1987: 25), although he admits that he “cannot give a clear account of what I have meant when earlier in this essay I have said that some subjects are more concerned with “conceptual matters” than are others” (1987: 32).
2 2 The overall criticism of Quine’s procedure goes back to Grice and Strawson (1956). Sober (2000) argues that Quine violates his own methodological naturalism in criticizing semantic notions on foundational grounds without considering their use in science.
3 3 Given Kripke’s arguments, defining “analytic” as the conjunction of “a priori” and “necessary” does not yield a natural notion, since a disjunction of an a priori contingency with an unrelated a posteriori necessity will then count as analytic: it is a priori because its first disjunct is and necessary because its second disjunct is. One does somewhat better by defining “analytic” as “a priori necessary,” which excludes that example, although the point of such a combination of epistemological and metphysical elements remains to be explained. The arguments below apply to this notion too. Of course, Kripke’s main concern is the difference between the a priori / a posteriori and the necessary/contingent distinctions; he clarifies their differences from the analytic/synthetic distinction in passing. Nevertheless, the differentiation between the first two distinctions forces the demotion of the third from that of trying to play both the first role and the second.
4 4 See Boghossian (1997) for the distinction between metaphysical and epistemological accounts of analyticity, and Tappenden (1993: 240) for a somewhat similar distinction.
5 5 Etchemendy (1990: 107–24) contrasts “substantive” generalizations with logical ones. The idea is widespread. It occurs in different forms in Wittgenstein’s Tractatus Logico-Philosophicus and in Locke’s “Of trifling propositions” (An Essay Concerning Human Understanding, Book IV, Chapter viii).
6 6 Since analytic truths are standardly taken to be sentences, the term “true” will sometimes be applied to sentences, as well as to thoughts and propositions; where required, the context makes clear what kind of truth-bearer is intended. Talk of knowing or believing a sentence should be understood as elliptical for talk of having knowledge or belief which one can express with the sentence (on its standard meaning). Thus someone who knows “Grass is green” knows that grass is green and can express that knowledge by saying “Grass is green”; this is not to be confused with the metalinguistic knowledge that the sentence “Grass is green” is true.
7 7 Proof: Assume (Taslr), (Faslr) and (Tasrl). To derive (Ta), note that it is equivalent to the conjunction of two claims: (i) if “P” is analytically true, then “P” is analytically true if and only if P; (ii) if “P” is analytically false, then “P” is analytically true if and only if P. Now (i) is logically equivalent to the claim that “P” is analytically true only if P, which follows from (Taslr). Moreover, by (Faslr) “P” is analytically false only if not P; as just seen “P” is analytically true only if P, so “P” is analytically false only if “P” is not analytically true; thus if “P” is analytically false then both sides of the biconditional in the consequent of (ii) fail, so (ii) holds. To derive (Ts), first note that “P” is synthetically true only if P by (Taslr). Conversely, if P then “P” is analytically true or synthetically true by (Tasrl); since by the antecedent of (Ts) it is not analytically true, it is synthetically true. Incidentally, by themselves (Ta) and (Ts) are weak in other ways too; in particular, they do not entail that nothing can be both analytically true and synthetically true.
8 8 For related arguments see Williamson (1994b: 141–2) and Tappolet (1997).
9 9 See Boghossian (1997: 335–6). Quine says that we can say that the logical truth “Everything is self-identical” depends for its truth “on an obvious trait, viz., self-identity, of its subject matter, viz., everything.” However, he claims that it makes no difference whether we say that or say that it depends for its truth “on traits of the language (specifically on the usage of “=”), and not on traits of its subject matter” (1966: 106).