Computation and Intentional Psychology

Murat Aydede

The University of Chicago
Department of Philosophy
1010 E. 59^th Street
Chicago, IL 60637, USA

ABSTRACT. The relation between computational and intentional psychology has always been a vexing issue. The worry is that if mental processes are computational, then these processes, which are defined over symbols, are sensitive solely to the non-semantic properties of symbols. If so, perhaps psychology could dispense with adverting in its laws to intentional/semantic properties of symbols. Stich, as is well-known, has made a great deal out of this tension and argued for a purely "syntactic" psychology by driving a wedge between a semantic individuation of symbol tokens and their narrow functional individuation. If the latter can be carried out, he claimed, we do not need semantic typing. I argue that since a narrow functional individuation cannot type-identify symbol tokens across organisms, a semantic account of typing must be the only option given that interpersonal physical individuation of tokens is not to be taken seriously.

In the beginning of his celebrated book From Folk Psychology to Cognitive Science (1983), Stephen Stich asked a certain question and built his argumentative strategy against intentional psychology around it. The question was put in the form of a theoretical demand: "[F]or a Fodor-style account of belief sentences to hang together, we must have some workable notion of what it is for two distinct people, speaking different languages, to have in their heads distinct tokens of the same sentence type" (Stich 1983, pp. 43-4)

This paper will be a meditation on this question: what is it for two symbol tokens of Mentalese in different heads to be of the same type? I will first lay out the background and the presuppositions of the question, together with the range of possible answers it might receive. I will indicate why the question is important and why we need an adequate answer to it. I will then show that if the question has an answer, then the answer will also be a good one for an equally important and vexing question: if psychological processes are computational, how can psychological laws be intentional?

This is what Fodor has recently called in his (1994) book the "Eponymous Question" (EQ). This second question has in fact been around, constantly popping up here and there and haunting people working in the field, for more than fifteen years, mostly thanks to Stich and Fodor's early "methodological solipsism".[1] I will show that this question has a very simple answer to the extent that Stich's question has one.

The Computational/Representational Theory of Mind (CRTM) is the doctrine that postulates a system of mental representations physically realized in the brain of sufficiently sophisticated cognitive organisms. This system of representations has a combinatorial syntax and semantics, and the processes that operate on the representations are causally sensitive to their syntax, but not to their semantics. Although CRTM is generally offered as a theory covering almost all aspects of cognition, the core domain of application is standardly conceived to be the explication of propositional attitudes like beliefs and desires. Roughly, according to CRTM, for any subject S and for any propositional attitude A toward any proposition P, S has A that P if and only if (a) S stands in a computational/functional relation A* to a symbol #P# in her Language of Thought (LOT) and (b) #P# means [P]. It is this doctrine that Stich has in mind when he raises his question.

Stich is asking about the identity conditions of symbols like #P# across cognitive organisms. It certainly seems fair to make such a demand: if you are advancing a theory that quantifies over certain entities like symbols realized in the brain, you owe an account of their individuation conditions. What is not clear, however, is whether the individuation conditions must be stated across organisms. The answer essentially depends on what you want to do with CRTM. If you want CRTM to state psychological laws or generalizations whose projectible predicates essentially denote symbol-types in the Mentalese of organisms covered by these laws, then an account of individuation across organisms is needed. For it is in terms of these posits CRTM is meant to cast its laws, i.e. to taxonomize its domain.[2] I will take CRTM for the purposes of the present paper as a claim about the proper scientific taxonomy of mental states (partly) in terms of underlying symbol types. That is, I want to take Stich's question as legitimate. This is how Stich and all the others who were interested in CRTM took the doctrine.[3]

When the stage is set this way, it is easier to see the significance of Stich's question. For suppose that it is possible to say what it is for symbol tokens in different brains to belong to a given type without ever mentioning their semantic properties, and then to state all the laws of cognition that advert to symbol types so individuated. This would mean at a minimum that it is possible to construct a scientific psychology that is free of intentional properties. And this in turn would mean that the intentional states posited by folk psychology are unnecessary for scientific purposes, hence raising the possibility for their elimination. This is the familiar strategy that Stich himself deploys in his (1983) book. He attempts to show that it is not only possible to come up with such semantic-free identity criteria for the explanatory posits of a scientific psychology, but also, given that semantic criteria are scientifically unfit for the job (because semantic typing allegedly yields vague and unstable theoretical posits whose attribution is observer relative and thus holistic), it is extremely desirable that we do so if we want psychology to be scientifically respectable. In what follows, assuming that Stich's question has an adequate answer, I will argue that there is no semantic-free way of typing symbol tokens across different organisms basically by arguing that none of the other suggested methods can possibly do the job.[4]

As Stich himself indicates, there are three standard ways to answer the question: what might be called the Physical Account (PA), the Narrow Functional Account (NFA) and the Semantic Account (SA).

According to the Physical Account (PA), two symbol tokens in different brains are of the same symbol type if and only if they have the same physical, or quasi-physical,[5] properties. Even if we leave aside the problem of stating exactly what physical properties of tokens must enter into typing them (since there are obviously so many of them and it is unlikely that all are relevant), PA is committed to a very strong version of type-physicalism across organisms with respect to particular propositional attitudes, like the belief that snow is white. In what follows, I will assume without argument that type-physicalism of this sort is false, whatever its prospects may be with respect to other ("less-cognitive") mental states such as feeling pain in one's left foot or seeing red, etc. I take it that this assumption is not very controversial, so we can safely put PA aside in order to concentrate on NFA, which is what people usually have in mind as the only significant alternative to SA.

According to the Narrow Functional Account (NFA), two symbol tokens in different brains are of the same type if and only if they have the same narrow functional role. Functional role is constituted by the causal relations of a symbol token, both actual and counterfactual. It is therefore customary to think of the functional role of a token as fixed by the Ramsey sentence of all the law-like generalizations involving the symbol type the token belongs to.[6] This is as it should be since we want the causal roles of tokens to come out (more or less) the same across systems. This could be secured by assuming that the law-like generalizations in question are interpersonally applicable. It will be important to keep in mind these two points. First, the causal roles are meant to be fixed by generalizations that apply across systems. Second, they are counterfactual supporting (and lawlike), covering not only the actual etiologies of tokens but also their potential ones. So, these generalizations go beyond summaries of what typically causes what, where the notion of typicality, as Stich notes, "is an utterly pedestrian one: a typical effect is just an effect that arises in a substantial majority of relevant cases" (1983: 27). The generalizations in question are also ceteris paribus ones.

What makes this account narrow is that the generalizations would detail only the causal relations among proximal stimuli, other symbol tokens, and a proprietary set of behaviors like basic motor-gestures. The network of causal relations described thereby totally supervenes on the brain (or central nervous system; at any rate, inside the skin) of organisms. As such, nothing outside the skin (i.e. the physical, social, historical or cultural environments in which the systems are embedded) is relevant to the type individuation of symbol tokens. It is crucially important for the functional account to be narrow. For if the functional roles are allowed to extend beyond the skin of organisms, the individuation of symbol tokens, as many theorists believe, cannot be semantic-free.[7]

Can NFA succeed in typing symbol tokens across organisms for all or even most of the symbol types that are likely to figure in psychological laws? Does it have the technical resources to do that? I will show that the answer to both of these questions is "no."

First, notice that what is being called for is not -- not only, at any rate -- the type identification of the syntactic structure or logical form of symbol tokens, e.g. their being conjunction, a conditional, a universal quantification, or the property of having a certain parsing tree such as [S -> NP N+VP V]. These may be called syntactic (or, formal) properties proper, and can perhaps be identified on the basis of the narrow functional (NR) profile of symbol tokens that possess them. Rather, we need to type individuate particular symbol tokens, so that e.g. #Fa# and #Fb#, or for that matter, #Fa# and #Ga#, belong to different symbol types. What is called for, in other words, is the individuation of symbol tokens as they differ in their particular lexical constituents. And this is to be done on the basis of their NF profile, as specified above.[8] This would require at a minimum having enough interpersonally applicable lawlike generalizations to detail the causal interactions of lexically specific symbol tokens with other such tokens, proximal stimuli, and basic motor-gestures, such that the generalizations would secure a unique NF role for each symbol type.[9] Failing to secure a unique role would be tantamount to failing to provide identity conditions for symbol types, i.e. for the primary theoretical posits of a psychology in terms of which psychological explanations are to be given. Are there enough generalizations like that? I do not think there are.

Let me first introduce some conventional notation in order to avoid long and cumbersome ways of expressing the same thing. I will mark an intentional expression with a '*' to express whatever its syntactic parallel may be. Also, I will hedge a content sentence or symbol with '#'s to indicate that I intend its syntactic parallel, i.e., whatever syntactic object or sentence (general or specific) might go in its stead.

Now take, first, the causal generalizations that are supposed to connect lawfully a set of proximal stimuli to, say, #Clinton is not faring well#, or any similarly specific sentence. Whatever the laws of psychophysics may tell us with respect to a very restricted range of psychophysically available properties, they will certainly be silent for the vast majority of symbol types figuring in full-blown propositional attitudes*. The problem partly stems from stimuli being proximal. There are certainly no scientifically well-delineated sets of proximal stimuli nomically correlated with the objects of beliefs*. This is to say that no such set could constitute a natural kind which would lawfully correlate with the objects of beliefs*. The other part of the problem is the holism involved in belief* fixation. Which proximal stimuli will cause which symbol(s) to be tokened in the belief* box is determined by what other symbols actually happen to be there and by the overall internal organization of the belief* box (simplicity, conservatism, etc.).

The history of behaviorism also provides an overwhelming inductive evidence that there are no such laws to be stated. No one has ever succeeded actually stating a single such law! Similarly for the supposed generalizations that would lawfully connect basic motor-gestures to particular symbol types in the belief* and desire* boxes. To be sure, behaviorists were after lawful stimuli/behavior connections, which is different. But the moral must be the same, since their failure partly stemmed from an inability to find projectible predicates to apply to all and only those proximal stimuli under physical descriptions that lawfully govern a given piece of behavior. They assumed that such stimuli directly and lawfully control the relevant piece of behavior: they wanted to bypass mediating internal states. They failed partly because of the holism problem again. Very little would change, however, if you assume that it is particular beliefs*, rather than behavior, that are directly under the lawful control of proximal stimuli: the routes from stimuli are equally holistic in each case.

Perhaps I am laboring this point needlessly. It should be clear that there are no laws to be stated with respect to proximal inputs/outputs for the full range of particular symbol types deployed in central cognitive processing as direct objects of propositional attitudes*. And even if there may happen to be some, they will be so few and fragile that they will be of very little help in type individuating all the symbol types we may need in psychological explanations

In fact, it is curious that more or less the same criticism is given by Stich himself for the claim made by content functionalists that there are such generalizations:

There is generally no characteristic environmental stimulus which typically causes a belief. There is no bit of sensory stimulation which typically causes, say, the belief that the economy is in bad shape, or the belief that Mozart was a freemason... Nor do most beliefs have typical behavioral effects. My belief that Ouagadougou is the capital of Upper Volta does not cause me to do much of anything. (1983, p. 24).

Later on, he argues (1983, pp. 180-1), on familiar grounds, that there can be no principled distinction between beliefs whose content is "observational" and those whose content is "theoretical." So, according to Stich, even for allegedly "observational" beliefs there seems not to be any particular set of stimuli nomologically connected to them. I wonder why and how Stich could think that the parallel case of beliefs* with particular symbol types as their objects is immune to the parallel criticism. (See below.)

It is clear, then, that the heaviest burden for the individuation of symbol tokens must be carried by the generalizations that hold among particular symbol types. This means that there must be enough interpersonally applicable and lawlike generalizations detailing the causal interactions among lexically specific symbol tokens to secure a unique NF role for each symbol type adverted to in our explanations of behavior. Are there such generalizations? Even if we put aside the problem of commitment to a principled distinction parallel to the analytic/synthetic one (a/s), there are other more serious ones.

Now, the thing to watch here is that the generalizations must be both lawlike and interpersonally applicable. Intuitively, the only hope of making them lawlike is by relying on the parallel of the historical/traditional idea of analytic connections among symbols. If the generalizations are to be lawlike, then, intuitively, they must resemble the following:

(1) For all subjects S and for all x, if S comes to believe* that #x is a cow#, then, ceteris paribus, S will come to believe* that #x is an animal#;

(2) For all subjects S and for all x, if S comes to believe* that #x is a bachelor#, then, ceteris paribus, S will come to believe* that #x is unmarried#;

(3) For all subjects S and for all x, if S comes to believe* that #x is assassinated#, then, ceteris paribus, S will come to believe* that #x is dead#.

Now, even these may not be lawlike if Quine is even remotely right about analyticity. But put that worry aside for the moment and assume that some parallel but weaker notion of analyticity can be sustained, especially if the epistemic claims usually accompanying the a/s distinction are dropped.[10] So there are "syntactic" parallels of analytic connections which can be specified in the form of generalizations, and these generalizations are lawlike if any generalization (causally connecting particular symbol types) is. But that is not enough: there must be enough of these to secure unique causal/functional roles on the basis of such generalizations. For given the scarcity of (interpersonally valid) analytic connections understood along the lines of the historical/traditional conception of analyticity, it is obvious that there is no hope of picking out a unique functional role for each possible symbol type on the basis of its analytic* connections.[11] Intuitively, making such a demand is tantamount to claiming that each possible concept* be definable* in terms of other symbols, i.e. be given necessary and sufficient conditions for its applicability. But even if it were possible to give a few necessary connections for some concepts* or even for each concept*, it would be very implausible to claim that each concept* could be given both necessary and sufficient conditions, especially after observing, as Fodor once put it, the failure of philosophy to define any term of any significance after two millennia.

On the other hand, it may be thought that if analytic* connections (as historically/traditionally conceived) will not suffice for defining unique roles, perhaps other kinds of connections would help, e.g., ordinary beliefs* about things that specify their contingent but quite common properties and relations. For instance, even if #all men are mortal# does not, strictly speaking, characterize an analytic* connection, it is nevertheless commonly held by relevantly equipped cognitive organisms. So perhaps, we may hope to be able to secure unique functional roles with generalizations such as

(4) For all subjects S and for all x, if S comes to believe* that #x is a star#, then, ceteris paribus, S will come to believe* that #x is a celestial object#

(5) For all subjects S and for all x, if S comes to believe* that #x is a bird#, then, ceteris paribus, S will come to believe* that #x has feathers#

(6) For all subjects S and for all x, if S comes to believe* that #x is a tiger#, then, ceteris paribus, S will come to believe* that #x is dangerous#

This is in fact the usual move made by many (narrow) functional role semanticists like Block (1993), and it lies at the source of Fodor's well-known pessimism that such a view leads to a destructive sort of holism. One way to see why this is troublesome is to ask whether these generalizations are lawlike. I do not think they are. I do not know exactly what needs to be true of a generalization for it to count as lawlike beyond merely being counterfactual supporting. The following rough and ready condition would have to suffice in our context:[12] a lawlike generalization (in the form of a conditional) is true just in case in all the (nomologically possible) nearby worlds where its antecedent is true its consequent is also true.

Take, for instance, (6). There are indefinitely many nearby nomologically possible worlds in which no organisms tend to think that something is dangerous upon thinking that it is a tiger, and this holds independently of whether in those worlds tigers, if they exist, are dangerous. Similarly for many others. In fact, we do not even have to look at other worlds. To make the point, the actual world will do. For instance, (4) is false of ancient Greeks (it may even be false of some of our contemporaries): they believed that stars are holes in the celestial spheres that the cosmic fire shows through (cf. Fodor 1987, p. 88-9).

That such generalizations are not lawlike in this sense should hardly come as a surprise. For, intuitively, their truth depends on what beliefs* people have actually in common or what common content*-specific inferences* they are disposed to engage, and to a very large extent this is not a matter expressible in terms of nomological necessities.[13] We may put the problem slightly differently as follows. Among such generalizations, many do not seem to be necessary for a subject to hold a belief* with a particular content*. For instance, for a subject to have tiger-thoughts*, it does not seem necessary that the generalization (6) hold for her, even though (6) may hold for everybody else.[14] But this does not seem to be true of, at least (2)-(3).

There seems to be a difference between (2)-(3) and (4)-(6). That, say, (2) should hold seems to be necessary for someone to have any bachelor-thought* at all. It seems that this is another way of saying that (2), if anything, is a lawlike generalization. Anyone for whom (2) is false ipso facto cannot have bachelor-thoughts*, or so it seems. In other words, in all the possible worlds in which bachelor-thoughts* are defined, anyone who comes to think* that #x is a bachelor# tends to think* that #x is unmarried#.

So, given that generalizations like (4)-(6) are not lawlike, hence must be restricted to doxastically* homogeneous populations, there is also the following problem. Since generalizations like (4)-(6) seem to be merely empirical generalizations that simply record the statistics of which particular belief*-types occur most often, how is NFA supposed to come up with identity conditions for beliefs* with merely possible particular contents* (and, not just logically possible, but "really" possible)? And there are infinitely many of them. This is a serious problem precisely because NFA wants to give the individuation conditions for any possible or actual belief* on the basis of generalizations most of which are bound to be not lawlike and not analytic*. Where are we to find them?

It is ironic that if NFA can be rescued only by such generalizations, then NFA-style psychologies are essentially committed to casting their laws for more or less doxastically homogeneous populations. It is ironic because it was precisely due to the fact that an intentional psychology, according to Stich, would miss generalizations covering doxastically different people, exotics, children, higher animals, etc., that his Syntactic Theory of Mind, which is basically a NFA-style theory, was supposed to be superior. But now we see that, Stich's advertisement to the contrary, his own STM is plagued by exactly parallel problems that he claimed plague content-based psychologies (vagueness, observer relativity of predicates like `believes that Fa', holism, etc.). So even if NFA were capable of type-identifying particular symbol types, it would not fare better than SA that Stich claims is hopeless.[15]

To recap, the question to ask is this: can we find enough generalizations that are both lawlike and interpersonally applicable to secure a unique role for each possible symbol type? It is important to note that these two requirements pull in opposite directions.[16] To the extent to which you can furnish interpersonally applicable generalizations and secure a unique role for each symbol type, to that extent you go against the requirement that the generalizations be lawlike. And to the extent to which you can give lawlike generalizations and secure unique roles, to that extent you violate the condition that they be interpersonally applicable. I do not think that there is an optimum point in the continuum between these extremes such that you can meet both of these requirements and secure a unique functional role for each possible symbol type. This is in fact more or less acknowledged by leading functional role semanticists like Block.[17] Hence, the destructive holism to which they are said to be committed.

I believe that what I have said so far makes a very strong case against the ability of NFA to individuate symbol types. I will, therefore, take all these as stated and conclude that Stich's question I quoted at the start cannot be answered on the basis of PA and NFA. Hence, assuming that Stich's question has an adequate answer, it must be some form or version of SA.

Stich calls the Narrow Functional Account of typing symbol tokens "syntactic" typing, presumably meaning just non-semantic and non-physical. And this sort of typing, on his view, is what the Computational Theory of Mind (CTM) is committed to. He then claims that CTM is all a scientific psychology needs; hence, contra Fodor, no need to appeal to semantic/intentional properties of syntactically structured brain symbols in stating the laws of psychology. He accuses Fodor of wanting to have it both ways.[18] We are now in a position to see how it is possible to have it both ways.

Let us suppose that CRTM is true. Any scientific computational psychology needs to postulate states in terms of which it can explain (and predict) behavior (construed broadly -- bodily, verbal and mental behavior). This calls for covering laws or generalizations that subsume those states under an appropriate description. This means that these states, under the relevant description, are projectible, i.e. natural kinds from the perspective of the theory. As such, they must have identity conditions. CRTM characterizes these states as symbol tokens realized in the heads of cognitive organisms. Qua symbols they have both syntactic and semantic properties. OK then, how are we to type them to suit the psychological laws covering their tokens? We have seen that they cannot be typed, in the required sense, by their physical and narrow functional properties: PA and NFA are hopeless. Our only other option, the Semantic Account, is in fact mandatory if psychological processes are to be computational. In other words, if Stich's question has an answer it must be some version of SA. It must be on the basis of their semantic properties that we type symbol tokens across systems. This is a conclusion that we've known all along anyway, or so it seems to me.

I therefore conclude that CTM itself is essentially committed to semantic type individuation of Mentalese symbol tokens across systems. And it is across systems that a scientific psychology casts its laws. Hence, the necessity for an intentional psychology whose laws advert to the semantic properties of representations. If mental representations can be typed interpersonally only on the basis of their semantic properties, CTM cannot be an alternative to replace intentional psychology. Hence the answer to Fodor's Eponymous Question.[19]

Aydede, Murat

Barsalou, Laurence W.

Block, Ned

Devitt, Michael

Field, Hartry

Fodor, Jerry A.

Jacob, Pierre

Loar, Brian F.

Rey, Georges

Schiffer, Stephen

Stich, Stephen P.

[1] See, among many others, Stich (1983, 1991), Field (1978), Schiffer (1987), Fodor (1980, 1989), Devitt (1991), Jacob (1997) who take issue with the EQ one way or other. Fodor's early writings (e.g. 1978, 1980) are full of pronouncements to the effect that a rationalist psychology could reasonably set for itself the goal of discovering the syntactic principles governing intentional organisms' Mentalese, much in the same way a Chomskian program set out to discover the principles of Universal Grammar.

[2] CRTM may not itself be such a theory. In fact, CRTM is usually given as a theory about the foundational assumptions of scientific psychological theories. On such a reading, it is not the CRTM itself that will state the laws of psychology; rather, the claim is that whatever psychological theories there are (or, will be) to state the laws of mind, they will conform to the general framework provided by CRTM in that they will postulate a LOT. Although this distinction is important for certain purposes, I will generally ignore it in what follows, except when it matters.

[3] It should be noted that the problems that emerge in typing tokens across organisms also emerge in typing tokens across time-slices of a single organism. In what follows, I will focus my discussion to interpersonal individuation. My discussion will also be neutral on whether the organisms across whom the individuation of tokens are carried out speak the same or different languages.

[4] I will therefore be assuming that Stich's own idiosyncratic analysis of what the Semantic Account (SA) amounts to is mistaken. Nothing should hang on this for what follows, since his promotion of a purely functionalist semantic-free psychology (what he misleadingly calls "syntactic psychology" or "STM-style" theories) presupposes the possibility of such a psychology, against which I will be arguing here. If, however, Stich turns out to be right that SA is unfit for serious psychology and I am also right about the fate of semantic-free accounts of symbol individuation, then psychology as we presently understand it is impossible! It is therefore practically mandatory for me to assume that Stich is wrong in his analysis of SA, which is not implausible on independent grounds anyway. It is important to understand that the two aims Stich had in his (1983) book are independent of each other. One aim was to establish that semantic typing is unsuitable for serious scientific purposes because of observer-relativity involved in the attribution of semantic properties and their holism. The other was to establish the possibility of a semantic-free "syntactic" psychology; in fact, he also argued that the practice of contemporary computational psychology was actually semantic-free. Although it is not essential for me, I am here assuming that his attempt to establish the former aim is bankrupt. But my real target in this paper is his second claim: I will be arguing that there can be no semantic-free psychology.

[5] The reason I say `quasi-physical' is that physical properties in this sense can be abstract, higher-order properties. For instance, take a complete specification of the shape of a symbol token. This seems like an abstract property: shapes of letters, for instance, can be realized in a variety of physical media: just think of the letter 'A' inscribed in sand, wax, etc. If so, some physical properties of symbols can be multiply realized without being functionally defined or even definable. This is why I call them quasi-physical. Nothing will hang on this however.

[6] A cluster theorist, however, would claim that the type identity of a symbol is given by a substantial majority of the generalizations that cover its tokens.

[7] Devitt (1990) claims that even allowing for proximal stimuli and behavior suffices to make the individuation semantic, even though a narrow one at that. I think, he no longer holds this view.

[8] I should note that Stich calls the NF individuation of lexically specific symbol tokens "syntactic identification." I think this is misleading and confusing. However, as long as we keep in mind that what he means by this is simply (non-semantic) NF individuation, nothing harmful should follow. For more discussion of this and related issues, see Devitt's insightful discussion in his (1990). See my (1997a) for the proper understanding of the notion of a combinatorial syntax involved in LOTH.

[9] I would like to raise, en passant, the question of whether we really need such content-specific laws as Fodor and Stich assume, or their "syntactic" counterparts. It is not clear that we do, at least as far as thought processes are concerned. See Devitt (1991) for more discussion of this. I will assume, however, a positive answer to that question here. It seems that we need to advert to content-specific propositional attitudes at least to explain particular behaviors of agents interacting with their environments to satisfy their particular desires. This is, at any rate, the core case of common-sense explanation of behavior.

[10] See Devitt (1996) and Rey (in preparation) for attempts to develop and defend the a/s distinction stripped of such claims. I think their efforts to develop a weaker notion of analyticity are promising, and we may indeed be in need of such a notion at any rate. On the other hand, in considering what generalizations NFA could appeal to, appeal to analyticity as such should, of course, drop out since this would introduce semantic considerations into NFA, which is supposed to be completely semantic-free. All NFA has at its disposal are causal/nomic relations among symbols.

[11] See Block (1986, 1993) for exactly the same point vis-à-vis the maximal role the traditional notion of analyticity could play within his own brand of conceptual role semantics.

[12] This characterization is way too sloppy. But I am not sure whether any extensive attempt to make it more precise and defensible is worth the effort in our context. The point that must be kept in mind is that these generalizations must go beyond mere counterfactual-supporting ones in at least the following way: they must be applicable across people and times and their applicability in this way should not merely be the result of the accidental identity of the causal/functional organization or structure of each individual under their scope. In other words, it must be in virtue of the nomic nature of the properties connected in the generalizations that they get projected by them.

[13] There is a certain sense in which the prototype theory of concepts can be seen to support the claim that people have outstandingly robust set of contingent beliefs surrounding a particular concepts. But, see Barsalou (1987) for conclusive evidence that prototypes are not robustly shared even intrapersonally.

[14] It may be thought that a cluster theorist may accommodate this fact: what is required is not whether all the generalizations specified in the theory for a particular content* hold in the case of each subject, but rather that a substantial number of them do, while no particular generalization is necessary. This is also Stich's line on type individuating syntactic objects. However, this cannot be quite true. A cluster content functionalist must choose the generalizations in the cluster from among those that are lawlike. But it usually so happens that each lawlike generalization about, say, bachelor-thoughts* is also necessary for having bachelor-thoughts*.

[15] See my (forthcoming) for an extensive elaboration of this kind of argument against Stich. There I argue that, contrary to Stich's advertisement and announcement in his (1983) book, his STM is not superior in any way to content-based psychologies (Representational Theories of Mind -- RTM) since it has exactly the parallel problems that he claims plague RTM. More accurately, I argue for a conditional claim: if Stich is right in his claim that RTM has the problems he enumerates (like holism, vagueness, etc. involved in semantic typing), then Stich's own STM has exactly the parallel problems. I also argue that an STM-style theorist is committed to intentional vocabulary to type its basic posits at some stage of theory construction at any rate.

[16] For a similar point regarding his own content functionalism, see Loar (1982).

[17] In addition to other references I gave to Block before, see also and especially his (1991).

[18] Stich (1983). See also his (1991). Devitt (1991) joins Stich in accusing Fodor of trying to have it both ways but only with respect to processes governing thoughts without I/Os.

[19] There are, to be sure, problems with any version of SA, as is well known. Suppose that SA is broad as in Fodor. Then we have problems with Frege cases as well as Twin-Earth cases. Adverting to underlying vehicle types, as Fodor does, is of no help since we lack interpersonally applicable individuation criterion for those precisely because a narrow functional individuation is out of question if I am right. See my (1998) for a critique of Fodor on this issue. A narrow SA would be equally problematic if it relies on narrow functional roles of vehicles as their narrow semantic content. On the other hand, a Fodor-style notion of narrow content as mapping from context to broad content (Fodor 1987) can perhaps handle at best Twin-Earth cases, but not Frege cases (see my 1997b). But being problematic is one thing, being wrong is another: I think, a SA that works could after all be salvaged in the face of apparent difficulties.