There Are No Recognitional Concepts; Not Even RED*
Let it be that a concept is recognitional if and only if:
(1) It is at least partially constituted by its possession conditions; and
(2) Among its possession conditions is the ability to recognize at least some things that fall under the concept as things that fall under the concept. For example, RED is a recognitional concept iff it numbers, among its constitutive possession conditions, the ability to recognize at least some red things as red.
In this paper, I propose to argue --- indeed, I propose to sort of prove --- that there are no recognitional concepts; not even RED.
Lots of philosophers are sympathetic to the claim that there are recognitional concepts. For one thing, insofar as recognitional capacities are construed as perceptual capacities, the claim that there are recognitional concepts preserves the basic idea of Empiricism: that the content of at least some concepts is constituted, at least in part, by their connections to percepts. For philosophers who suppose that Empiricism couldn't be all wrong, recognitional concepts can therefore seem quite a good place to dig in the heels. More generally, the claim that there are recognitional concepts is a bastion of last resort for philosophers who think that semantic facts are constituted by epistemological facts, a doctrine that includes, but is not exhausted by, the various forms of Empiricism. If you have ever, even in the privacy of your own home among consenting adults, whispered, hopefully, the word `criterion', then probably even you think there are recognitional concepts.
Philosophers who hold that there are recognitional concepts generally hold that it's important that there are; for example, a familiar line of anti-skeptical argument turns on there being some. The idea is that, if a concept is recognitional, then having certain kinds of experience would, in principle, show with the force of conceptual necessity that that the concept applies. If, for example, RED is a recognitional concept, then having certain kinds of experience would, in principle, show with the force of conceptual necessity that there are red things. Ditto, mutatis mutandis, SQUARE, CHAIR, IS IN PAIN and BELIEVES THAT P, assuming that these are recognitional concepts. So, if you think that it's important that skepticism about squares, chairs, pains, beliefs or red things be refuted, you are likely to want it a lot that the corresponding concepts are recognitional. Nevertheless, it's sort of provable there aren't any recognitional concepts; so, at least, it seems to me.
I pause to mention a kind of argument against there being recognitional concepts to which I am sympathetic, but which I am not going to pursue in what follows. Viz. that it's truistic that the content of ones experience underdetermines the content of ones beliefs, excepting only ones beliefs about one's experiences. No landscape is so empty, or so well lit ---so the thought goes--- that your failure to recognize that it contains a rabbit entails that you haven't got the concept RABBIT. So, it couldn't be that your having the concept RABBIT requires that there are circumstances in which you couldn't but recognize a rabbit as such.
I think this is a good argument, but, notoriously, lots of philosophers don't agree; they think, perhaps, that the connection between concept possession and recognitional capacities can be relaxed enough to accommodate the truisms about rabbits without the claim that there are recognitional concepts lapsing into vacuity. I propose, in any event, not to rely upon this sort of argument here.
Compositionality. The considerations I will appeal to are actually quite robust, so a minimum of apparatus is required to introduce them. It will, however, be useful to have on hand the notion of a satisfier for a concept. The satisfier(s) for a concept are the states, capacities, dispositions, etc. in virtue of which one meets the possession condition(s) for the concept.1 So, if the ability to tell red from green is a possession condition for the concept RED, then being able to tell red from green is a satisfier for the concept RED. If a disposition to infer P from P&Q is a possession condition for the concept &, then being so disposed is a satisfier for the concept &. And so forth. Since, by assumption, concepts have their possession conditions essentially, and possession conditions have their satisfiers essentially, the exposition will move back and forth among the three as convenience dictates.
I propose to argue that there are no concepts whose satisfiers are recognitional capacities; hence that there are no recognitional concepts. I need a premise. Here's one:
Premise P: S is a satisfier for concept C if and only if C inherits S from the satisfiers for its constituents concepts.2 (I'll sometimes call this the `compositionality condition' on concept constitution.)
Consonant with my general intention not to have the argument turn on its details, I leave it open how `inherited from' is construed in premise P, so long as fixing the satisfiers for constituent concepts is necessary and sufficient for fixing the satisfiers for their hosts.
Why premise P is plausible: Unless P is true, we will have to give up the usual account of why concepts are systematic and productive; and, mutatis mutandis, of how it is possible to learn a language by learning its finite basis. Consider, for example, the concept-constitutive possession conditions for the concept RED APPLE. If Premise P is false, the following situation is possible: The possession conditions for RED are ABC and the possession conditions for RED APPLE are ABEFG. So denying P leaves it open that one could have the concept RED APPLE and not have the concept RED.
But, now, the usual compositional account of productivity requires that one satisfy the possession conditions for complex concepts, like RED APPLE, by satisfying the possession conditions for their constituent concepts. That is, it requires that one's having a grasp of the concept RED is part of the explanation of one's having a grasp of the concept RED APPLE. So accepting the usual compositional account of productivity is incompatible with denying premise P.
Likewise the other way around. The usual compositional account of productivity requires that if one satisfies the possession conditions for the constituents of a complex concept, one thereby satisfies the possession conditions for the concept.3 But, suppose that premise P is false, and consider, once again, the concept RED APPLE. Denying P leaves it open that the concept-constitutive possession conditions for RED APPLE are not exhausted by the concept-constitutive possession conditions for RED and APPLE. For example, the former might be ABCDE and the latter might be AB and CD respectively. But then grasping the concepts RED and APPLE would not be sufficient for grasping the concept RED APPLE, and, once again, the standard account of conceptual productivity would be undermined.
So much for the bona fides of premise P. The next point is that the condition that compositionality imposes on concept constitution is highly substantive. A brief digression will show the kind of theory of concepts that it can rule out.
Consider the idea that concepts are (or are partially) constituted by their stereotypes, hence that knowing its stereotype is a satisfier for some concepts. Premise P says that this idea is true only if, if you know the stereotypes for the constituents of a complex concept, then you know the stereotype for that concept. Which, in some cases, is plausible enough. Good examples of RED APPLES are stereotypically red and stereotypically apples. Let's assume that that's because the stereotype of RED APPLE is inherited from the stereotype for RED and the stereotype for APPLE. (This assumption is concessive, and it may well not be true. But let's let it stand for the sake of the argument.) So then, as far as RED APPLE is concerned, it's compatible with premise P that knowing their stereotypes should be possession conditions for RED and APPLE.
Still, concepts can't be constituted by their stereotypes (knowing its stereotype can't be a satisfier for a concept). That's because RED APPLE isn't the general case. In the general case, complex concepts don't inherit their stereotypes from those of their constituents. So, in the general case, stereotypes don't satisfy premise P.
Consider such concepts as PET FISH, MALE NURSE, and the like.4 You can't derive the PET FISH stereotype from the FISH stereotype and the PET stereotype. So, if stereotypes were constitutive of the corresponding concepts, having a grasp of FISH, and having a grasp of PET (and knowing the semantics of the AN construction; see fn. 3) would not suffice for having a grasp of PET FISH. So, the usual story about how PET FISH is compositional would fail.
So much for stereotypes. If premise P is true, it follows that they can't be concept-constitutive. I will now argue that, if premise P is true, then it likewise follows that there are no recognitional concepts.
In fact, most of the work is already done since, for all intents and purposes, the notion of a recognitional concept is hostage to the notion that concepts are constituted by their stereotypes. Here's why. Nobody could (and nobody does) hold that the possession of a recognitional concept requires being able to identify each of its instances as such; if that were the requirement, then only God would have any recognitional concepts. So, the doctrine must be (and, as a matter of fact, it always is) that possession of a recognitional concept requires the ability to identify good instances as such in favorable conditions. (There are various variants of this in the literature; but it doesn't matter to what follows which you choose.)5
But, now, unsurprisingly, the ability to recognize good instances of Fs doesn't compose, and this is for exactly the same reason that knowing the stereotype of F doesn't compose; good instances of F&Gs needn't be either good instances of F or good instances of G. See PET FISH, once again: Good instances of PET FISH are, by and large, poorish instances of PET and poorish instances of FISH. So a recognitional capacity for good instances of PET and good instances of FISH is not required for, and typically does not provide, a recognitional capacity for good (or, indeed, any) instances of PET FISH.
Somebody who is good at recognizing that trouts are fish and that puppies are pets is not thereby good at recognizing that goldfish are pet fish. The capacity for recognizing pet fish as such is not conceptually, or linguistically, or semantically connected to capacities for recognizing pets as such or fish as such. The connection is at best contingent, and it's entirely possible for any of these recognitional capacities to be in place without any of the others.
This doesn't, of course, show that the semantics of PET FISH are uncompositional. What it shows is that recognitional capacities aren't possession conditions for the concepts that have them. If recognitional capacities were possession conditions, PET FISH would not inherit its satisfiers from those of PET and FISH. So if recognitional capacities were possession conditions, PET FISH would fail premise P. So recognitional capacities aren't possession conditions. So there are no recognitional concepts.6
Q1: You're, in effect, taking for granted not only that compositionality is needed to explain productivity, but that it is therefore a test for whether a property is constitutive of the concepts that have it. Why should I grant that?
A1: I suppose I could just dig my heels in here. Compositionality is pretty nearly all that we know about the individuation of concepts. If we give that up, we will have no way of distinguishing what constitutes a concept from such of its merely contingent accretions as associations, stereotypes and the like.
But though I do think it would be justifiable to take that strong line, I really don't need to in order to run the sort of argument I'm endorsing. If push comes completely to shove, the following exiguous version will do for my polemical purposes:
If you know what `pet' and `fish' mean, you thereby know what `pet fish' means. But you can be able to recognize pets and fish as such but be quite unable to recognize pet fish as such. So recognitional capacities can't be meanings, and they can't be constituents of meanings, all varieties of Empiricist semantics to the contrary notwithstanding.
As far as I can see, that formulation doesn't need much more than the distinctness of discernibles, so it seems to me that it cuts pretty close to the bone.
Q2: But couldn't an Empiricist just stipulate that recognitional capacities, though they demonstrably don't satisfy P, and are thus demonstrably not constituents of meanings, are nevertheless to count as essential conditions for the possession of primitive concepts?
A2: Sure, go ahead, stipulate; and much joy may you have of it. But nothing is left of the usual reasons for supposing that any of the concepts we actually have comply with the stipulation; in particular, nothing is left of the idea that the content of our concepts is constituted by our recognitional capacities. Whatever content is, it's got to be compositional; so it's got to come out that the content of RED APPLE includes the content of RED, and the content of PET FISH includes the content of PET.
It may be worthwhile to reiterate here an argument I gave for the plausibility of premise P. Suppose a primitive concept has a possession condition which is not inherited by one of its complex hosts; suppose, for example, that being able to recognize good instances of pets is a possession condition for PET but is not a possession condition for PET FISH. Then it is presumably possible that someone who has the concept PET FISH should nonetheless not have the concept PET. I take this to be a reductio, and I think that you should too.
Here's a closely related way to make the same argument: Perhaps you're the sort of philosopher who thinks it's a possession condition for RED APPLE that one is prepared to accept the inference RED APPLE ---> RED (i.e. that one finds this inference `primitively compelling').7 If so, that's all the more reason for you to hold that the possession conditions for RED APPLE must include the possession conditions for RED. Hence, it's all the more reason for you to hold that the satisfiers for RED are inherited under composition by RED APPLE. But if that's right, then, once again, it couldn't be that a recognitional capacity is a satisfier for RED unless it's a satisfier for RED APPLE. But the capacity to recognize pets as such is not a satisfier for the concept PET FISH, so it can't be a satisfier for PET. Since sauce for the goose is sauce for the gander, the ability to recognize red things is likewise not a satisfier for RED.
Q3. Couldn't we split the difference? Couldn't we say that the satisfiers for the primitive concepts include recognitional capacities, but that the satisfiers for complex concepts don't?
A3. Simply not credible. After all, people who have the concept PET FISH do generally have a corresponding recognitional capacity; for example, they are generally good at recognizing goldfish as pet fish. And, surely, being able to recognize (as it might be) a trout as a fish stands in precisely the same relation to having the concept FISH as being able to recognize a goldfish as a pet fish does to having the concept PET FISH. So, how could that relation be constitutive of concept possession in the one case but not in the other? Is it, perhaps, that the concepts FISH and PET FISH have content in different senses of `content'?
This sort of point is probably worth stressing. Some philosophers have a thing about recognitional capacities because they want to tie meaning and justification close together (see the remarks earlier about antiskeptical employments of the idea that there are recognitional concepts). But if recognitional capacities are constitutive only of primitive concepts, then the connection between meaning and justification fails in infinitely many cases. It will thus be concept-constitutive that (ceteris paribus) it's troutlooking is evidence for `it's a fish', but not concept constitutive that (ceteris paribus) it's goldfishlooking is evidence for `it's a pet fish'. What possible epistemological use could such a notion of concept constitutivity be put to?
Q4. FISH and PET are only relatively primitive (they're only primitive relative to PET FISH). What about absolutely primitive concepts like RED? Surely the concept RED is recognitional even if neither FISH nor PET FISH is.
A4. It's just more of the same. Consider RED HAIR, which, I will suppose, is compositional (that is, not idiomatic) and applies to hair that is red as hair goes. This view of its semantics explains why, though red hair is arguably not literally red, still somebody who has RED and has HAIR and who understands the semantic implications of the syntactic structure AN, can figure out what `red hair' means. So, prima facie, RED HAIR is compositional and the demands of productivity are satisfied according to the present analysis.8
But notice, once again, that the productivity/compositionality of the concepts does not imply the productivity/compositionality of the corresponding recognitional capacities. Somebody who is able to say whether something is a good instance of HAIR and whether something is a good instance of RED is not thereby able to recognize a good instance of RED HAIR. Well then, what does `red' contribute to the semantics of`red hair'? Just what you'd suppose: it contributes a reference to the property of being red (as such). It's just that its doing that isn't tantamount to, and doesn't entail, its contributing a recognitional capacity for (good instances of) redness.
One's recognitional capacity for RED doesn't compose. So one's recognitional capacity for red things is not a satisfier for the concept RED. So not even RED is a recognitional concept.
Q5. What do you say about intentional concepts?
A5. Nothing much for present purposes. They have to be compositional, because they are productive. If they are compositional, then there are, to my knowledge, three theories (exhaustive but not exclusive) of what they inherit from their constituents:
-they inherit the extensions of their constituents.
-they inherit the senses of their constituents
-the inherit the shapes of their constituents. (Notice that shape is compositional; `red hair' contains `red' as a morphosyntactic part; and the shape of `red hair' is completely determined given the shape of `red', the shape of `hair', and the morphosyntactics of the expression.)
But, invariably a theory of intentional concepts that says that any of these are inheritable properties of their constituents will also say that they are constitutive properties of their constituents. So, as far as I can tell, nothing that anybody is likely to want to say about intentional concepts will deny my argument the premise it requires.
Conclusion. The moral of this paper is that recognitional capacities are contingent adjuncts to concept possession, much like knowledge of stereotypes; a fortiori, they aren't constitutive of concept possession. How, indeed, could anyone have supposed that recognitional capacities are satisfiers for concepts, when recognitional capacities patently don't compose and concept satisfiers patently do?
I think what went wrong is, after all, not very deep, though it's well worth attending to. Content, concept-constitutivity, concept possession, and the like, are connected to the notion of an instance (i.e. to the notion of an extension). The notion of an instance (/extension) is semantic, hence compositional, through and through; idioms excepted, what is an instance of a complex concept depends exhaustively on what are the instances of its parts. The notion of a recognitional capacity, by contrast, is connected to the notion of a good (in the sense of a typical, or a reliable) instance; the best that a recognitional capacity can promise is to identify good instances in favorable conditions. It's a mistake to try to construe the notion of an instance in terms of the notion of a good instance;9 unsurprisingly, since the latter is patently a special case of the former, the right order of exposition is the other way around.
Recognitional capacities don't act like satisfiers: What's a satisfier for a complex concept depends on what's a satisfier for its parts; but what's a good instance of a complex concept doesn't depend on what's a good instance of its parts. Why should it? What's a good instance of a concept, simple or complex, depends on how things are in the world.10 Compositionality can tell you that the instances of PET FISH are all and only the pet fish; but it can't tell you that the good instances of pet fish are the goldfish; which is, all the same, the information that pet fish recognition (as opposed to mere PET FISH instantiation) is likely to depend on. How could you expect semantics to know what kind of fish people keep for pets? Likewise, what counts as red hair depends, not just on matters of meaning, but also on what shades hair actually comes in (i.e. because red hair is hair that is relatively red.) How could you expect semantics to know what shades hair actually comes in? Do you think that semantics runs a barber shop?11 How, in short, could you expect that relations among recognitional capacities would exhibit the compositionality that productivity requires of semantic relations?
Oh, well; so what if there are no recognitional concepts?
For one thing, as I remarked at the outset, if there are no recognitional concepts we lose a certain class of antiskeptical arguments; ones that depend on the connection between percepts and concepts being, in some sense, constitutive of the latter. We will no longer be able to say to the skeptic: `If you don't think that this experience shows that that's a chair, then you don't have the concept CHAIR'. But maybe this isn't a great loss; I've never heard of a skeptic actually being convinced by that kind of argument. I sure wouldn't be if I were a skeptic.
I'm not, however, meaning to deny that the issue about recognitional concepts goes very deep. To the contrary, I'm meaning to claim that it goes very much deeper than (mere) epistemology. Close to the heart of the last hundred years of philosophy is an argument between a Cartesian and a Pragmatist account of concept possession. Though the details vary, the essentials don't: According to Cartesians, having the concept X is being able to think about Xs; according to Pragmatists, its being able to respond differentially or selectively to Xs (for short: it's being able to sort Xs.) I doubt that there's a major philosopher, anyhow since Peirce ---and including, even, the likes of Heidegger---, who hasn't practically everything at stake on how this argument turns out.
Notice that the issue here isn't `Naturalism'. Sorting is just as intentional as thinking, and in the same way: Neither coextensive thoughts nor coextensive sorts are ipso facto identical. A Pragmatist who isn't a behaviorist can (and should) insist on this. The issue, rather, is whether the intentionality of thought derives from the intentionality of action. Roughly, Pragmatists think that it does, whereas Cartesians think that the metaphysical dependencies go the other way around. It's the difference between holding, on the one hand, that whether you are sorting Xs is a matter of how you are thinking about what you are doing; or, on the other hand, that whether you are thinking about Xs depends on (possibly counterfactual) subjunctives about how you would sort them.
Well, the minimal Pragmatist doctrine (so it seems to me) is the claim that there are recognitional concepts; i.e. that at least some concepts are constituted by one's ability to sort their instances. And the present arguments (so it seems to me) show that even this minimal Pragmatist doctrine isn't true. Thinking centers on the notion of an instance; recognitional capacity centers on the notion of a good instance. Unless you are God, whether you can recognize an instance of X depends on whether it's a good instance of an X; the less good it is, the likelier you are to fail.12
But you can always think an instance of X; viz. by thinking an instance of X. So thinking is universal in a way that sorting is not. So thinking doesn't reduce to sorting. That is bed rock. To try to wiggle out of it, as so many philosophers have drearily done, by invoking ideal sorts, recognition under ideal circumstances, the eventual consensus of the scientific community, or the like, is tacitly to give up the defining Pragmatist project of construing semantics epistemologically. Being an ideal sort always turns out not to be independently definable; it's just being a sort that gets the extension right.
Or, to put the point with even greater vehemence: The question whether there are recognitional concepts is really the question what thought is for; whether it's for directing action or for discerning truth. And the answer is that Descartes was right: The goal of thought is to understand the world, not to sort it. That, I think, is the deepest thing that we know about the mind.
Afterward. This paper was presented at the 1997 meeting of the Central Division of the American Philosophical Association. Stephen Schiffer commented, and what he said was typical of the reaction I've had from a number of philosophical friends. So I include here my reply to Steve's reply:
Steve asked, in effect: `What's wrong with a mixed view, according to which recognitional capacities are constitutive for (some) primitive concepts but not for their complex hosts?' Steve thinks that my reply must be either aesthetic (mixed theories are ugly) or an outright appeal to the `agglomerative principle' that if the conjuncts of a conjunctive proposition are recognitional, then so too is the conjunctive proposition. Since Steve takes this principle to be not better than dubious, he thinks that I haven't a better than dubious argument against there being recognitional concepts.
Now, it does seem to me that somebody who holds that there are primitive recognitional concepts should also hold the agglomerative principle (see the discussion of Q3). But my thinking this isn't an essential part of my argument. I tried to make clear in the text what the essence of my argument is; but, evidently, I didn't succeed. This hasn't been my century for making things clear.
Here it is again:
A theory of compositionality should explain why, in the standard case, anybody who has a complex concept also has its constituent concepts. (Why anybody who has RED TRIANGLE has RED and TRIANGLE; why anybody who has GREEN HAIR has GREEN and HAIR.... and so forth.) This is tantamount to saying that a compositionality principle should be so formulated as to entail that satisfying the possession conditions for a complex concept includes satisfying the possession conditions for its constituents.
Now look at Steve's proposal, which is that "It is reasonable to hold that the possession conditions of complex concepts are determined by those of their constituents concepts. But for the case at hand this simply requires F-&-G to be such that to possess it the possessor must be able to recognize good instances of F and good instances of G."
As stated, Steve's theory is wrong about PET FISH: It's true, of course, that to have the concept PET FISH you have to have the concept FISH. But it's certainly not true that to have the concept PET FISH you have to have a recognitional capacity for good instances of fish. To have a concept, conjunctive or otherwise, you have to have the concepts that are its constituents. But you don't have to have recognitional capacities corresponding to its constituents; not even if, by assumption, the complex concept is itself recognitional. So, having the constituents of a concept can't require having a recognitional capacity in respect of their instances. If it did, you could have a complex concept without having its constituents; which is not an option. So concepts can't be recognitional capacities.
Can Steve's proposal be patched? Well, if he is to get the facts to come out right, he'll presumably just have to stipulate that for some F-&-G concepts (RED TRIANGLE) "the possessor must be able to recognize good instances of F and G", but that for others (PET FISH, MALE NURSE) that's not required. And he'll have to say, in some general and principled way, which concepts are which. But, surely, you don't want to have to stipulate the relations between the possession conditions for a complex concept and the possession conditions for its constituents; what you want is that they should just fall out of the theory of compositionality together with the theory of concept constitutivity.
Which, indeed, they do if you get these theories right. What's constitutive of FISH, and hence what PET FISH inherits from FISH, is (not a capacity for recognizing fish but) the property that FISH expresses: viz. the property of being a fish. Likewise, mutatis mutandis, what's constitutive of RED, and hence what RED TRIANGLE inherits from RED, is (not a recognitional capacity for red things but) the property that RED expresses, viz. the property of being RED. Because what is constitutive of a concept determines its possession conditions, it follows that you can't have the concept PET FISH unless you know that pet fish are fish, and you can't have the concept RED TRIANGLE unless you know that red triangles are red. This is, of course, just what intuition demands.
Explanations are better than stipulations; and they're a lot better than stipulations that misdescribe the facts. So there still aren't any recognitional concepts.13
NOTES TO CHAPTER 4
*Thanks to Ned Block, Paul Horwich, Chris Peacocke, Stephen Schiffer and Galen Strawson for helpful comments on an earlier draft.
1. Whereas, by contrast, the satisfiers of a concept are just: whatever is in its extension. This is not, admittedly, a very happy way of talking, but it's no worse, surely, than intention/intension, and nothing better came to mind.
2. If C is a primitive concept, the condition is trivially satisfied.
3. This isn't quite right, of course; you also have to know how the constituents are `put together'. Suppose a satisfier for RED is being able to identify red things and a satisfier for SQUARE is being able to identify square things. Then, presumably, the corresponding satisfier for RED SQUARE is being able to identify things in the intersection of RED and SQUARE. The fact that it's the intersection rather than, say, the union, that's at issue corresponds to the structural difference between the concept RED SQUARE and the concept RED OR SQUARE.
4. The immediately following arguments are familiar from the cognitive science literature on stereotypes, so I won't expand on them here. Suffice it to emphasize that the main point ---that stereotypes don't compose--- holds whether stereotypes are thought of as something like exemplars or as something like feature sets. For reviews, see Fodor and Lepore, 19xx; Fodor, forthcoming, 1997.
5. The intended doctrine is that having the recognitional concept F requires being able to recognize good instances of F as instances of F, not as good instances of F. It's concessive of me to insist on this distinction, because it requires the Empiricist to defend only the weaker of the two views.
6. This assumes, of course, that what holds for PET and for FISH holds likewise for any candidate recognitional concept: viz. that there will always be some complex concept of which it is a constituent but to which it does not contribute its possession condition. The reader who doubts this should try, as an exercise, to find a counterexample. For starters, try it with RED and APPLE.
7. Even conceptual atomists like me can hold that inferences which relate a complex concept to its parts are typically analytic and concept constitutive. See Fodor and Lepore, 1992.
8. Let it be that an AN concept is `intersective' if its extension is the intersection of the As with the Ns. The standard view is that being intersective is sufficient but not necessary for an AN concept to be compositional: RED HAIR is compositional but not intersective, and PET FISH is both. (For a general discussion, see Kamp and Partee, (1995)). Actually, my guess is that RED HAIR, BIG ANT, and the like are the general case. Excepting the `antifactive' adjectives (`fake', `imitation', etc.) AN usually means A FOR (an) N, and the intersectives are just the limiting case where things that are A for (an) N are A.
But it doesn't matter for present purposes whether this is so.
9. Indeed, it's a venerable mistake. I suppose the Platonic theory of Forms was the first to commit it.
10. It also depends on how things are with us. What the good instances of RED are almost certainly has to do with the way the physiology of our sensory systems is organized (see, for example, Rosch, 1973; Berlin and Kaye, 1969). Likewise, it's no accident that the good instances of ANIMAL are all big enough for us to see (i.e. big enough for us to see). It does not follow that a creature whose range of visual acuity is very different from ours would thereby have a different concept of animals from ours.
11. The cases in the text are not, of course, exceptional. What counts as an average income depends not only on what `average' and `income' mean, but also on what incomes people actually earn. Semantics tells you that the average income is in the middle of the income distribution, whatever the distribution may be. But if you want to recognize an average income, you need the facts about how many people actually earn how much. Semantics doesn't supply such facts; only the world can.
12. Analogous remarks hold for other epistemological capacities like, e.g., drawing inferences. Unless you are God, whether, in a particular case, you are disposed to infer P from P&Q depends, inter alia, on whether the logical form of the proposition is perspicuous. This strongly suggests that the considerations that rule out recognitional capacities as concept-constitutive will apply, mutatis mutandis, to rule out any epistemological candidate.
13. Steve also suggested that maybe PAIN is a recognitional concept, even if RED is not. I won't, however, discuss the notion that sensation concepts might be recognitional since I guess I don't really understand it. Does one recognize one's pains when one has them? Or does one just have them? If I can, indeed, recognize good instances of MY PAIN, I suppose it follows that I have the concept PAIN. Does it follow, as compositionality would require if PAIN is a recognition concept, that I can also recognize good instances of YOUR PAIN?
Hard cases make bad laws. Sensation concepts are too hard for me.