Information and Content

Jonathan Cohen

Mental states differ from most other entities in the world in having semantic or intentional properties: they have meanings, they are about other things, they have satisfaction- or truth-conditions, they have representational content. Mental states are not the only entities that have intentional properties - so do linguistic expressions, some paintings, and so on; but many follow [Grice, 1957] in supposing that we could understand the intentional properties of these other entities as derived from the intentional properties of mental states (viz., the mental states of their producers). Of course, accepting this supposition leaves us with a puzzle about how the non-derivative bearers of intentional properties (mental states) could have these properties. In particular, intentional properties seem to some to be especially difficult to reconcile with a robust commitment to ontological naturalism - the view that the natural properties, events, and individuals are the only properties, events, and individuals that exist. Fodor puts this intuition nicely in this oft-quoted passage:

I suppose that sooner or later the physicists will complete the catalogue they've been compiling of the ultimate and irreducible properties of things. When they do, the likes of spin, charm, and charge will perhaps appear upon their list. But aboutness surely won't; intentionality simply doesn't go that deep.... If aboutness is real, it must be really something else ([Fodor, 1987], 97).

Some philosophers have reacted to this clash by giving up one of the two views generating the tension. For example, [Churchland, 1981] opts for intentional irrealism in order to save ontological naturalism, while [McDowell, 1994] abandons naturalism (in the sense under discussion) in favor of a kind of intentional dualism as a way of preserving intentional realism. (Terminological caution: unfortunately, McDowell reserves the term `naturalism' for his own view, and refers to the sort of naturalism under discussion here - which he rejects - as `bald naturalism'.) However, others hope to find a way of reconciling their naturalistic ontology with intentional realism. In particular, many propose to locate intentionality within the natural order by naturalizing the intentional - by offering an account of intentional features in naturalistically respectable terms.

Many philosophers pursuing this project think that an appeal to information might satisfy their needs - they believe that this notion is both naturalistically acceptable and adequate to the analysis of the intentional. In this essay I shall review several attempts to naturalize the intentional in terms of information. In §1 I'll lay down some conditions of adequacy on informational theories of content that will be useful in evaluating the theories presented later. Next, in §2, I'll consider Dretske's influential early formulation of an informational theory. I'll move on in §§3-4 to discuss some elaborations of the view that incorporate notions of epistemic optimality and teleology. Finally, in §5 I'll discuss another informational theory, due to Fodor, that turns on his notion of asymmetric dependence.

1  Adequacy Conditions

It will be useful to structure the discussion around a number of adequacy conditions that an acceptable naturalistic theory of content must meet:
To vindicate the naturalism in whose service it is deployed, an informational theory of content must show how intentional properties can be characterized in naturalistic terms. While it is not obvious what terms should count as acceptably naturalistic, proposed naturalistic accounts of intentional properties should not presuppose that intentional properties are compatible with naturalism (on pain of begging the very question such theories are intended to answer). Consequently, the necessary and sufficient conditions for having intentional features required by a naturalistic theory of content must not themselves employ intentional idioms, nor should they employ any other notions whose compatibility with naturalism is in doubt. That is to say, the naturalism constraint is intended to force analyses of intentional features to break out of the circle of intentional features, and therefore proposed analyses of these features in terms of others of these features - satisfaction, truth, content, believing that p, etc. - will count for present purposes as unacceptably non-naturalistic. On the other hand, the insistence on naturalism is not intended to preclude abstract entities (e.g., laws, properties, numbers) from appearing in the analysis of content.
Since at least [Frege, 1892], philosophers of mind and language have made much of the extremely fine grain of individuation for content. In particular, there can be distinct contents that are equivalent in any of a number of senses (e.g., logical, analytic, metaphysical, or nomic equivalence). Many follow Frege in arguing for this conclusion from the observation that it is possible rationally to believe that p and disbelieve that q, even when p and q are equivalent in one of these ways; for example, even assuming that it is metaphysically necessary that the Morning Star is identical to the Evening Star, it is possible rationally to believe that the Morning Star is wet and simultaneously disbelieve that the Evening Star is wet (you might be in such a state if you didn't know that the Morning Star is identical to the Evening Star). A theory of content, therefore, must make it possible to distinguish between contents that are logically, analytically, metaphysically, or nomically equivalent.
An acceptable theory of intentional content must make it possible that intentional contents can represent the world erroneously - that they can represent the world as being a certain way even when, in fact, the world is not that way (the problem of formulating an account that makes room for misrepresentation is sometimes called `the problem of misrepresentation' or `the problem of error').

2  Dretske and the Flow of Information

One of the earliest systematic attempts to understand intentional content in terms of information occurs in Fred Dretske's seminal Knowledge and the Flow of Information [Dretske, 1981] (Dretske's was not, however, the first such proposal; antecedents include [Stampe, 1975], [Stampe, 1977], and some suggestive remarks about natural meaning from [Peirce, 1931] and [Grice, 1957].) (Dretske's account is also discussed in the chapter on knowledge in the present volume.)

Dretske wants to understand the intentional content of a signal (e.g., a mental state) in terms of the information that that signal carries under certain circumstances. He understands information in terms of objective conditional probabilities between events (I shall ignore problems concerning the interpretation of these probabilities raised in [Loewer, 1983] and [Loewer, 1987]): he writes (65) that a signal r carries the information that p just in case the conditional probability of p, given r, (and k, the knowledge of the receiver of r), is 1 (but given k alone, less than 1).

But we cannot straightforwardly identify information carried with intentional content because a signal will usually carry too much information. For example, an acoustic signal carrying the information that the doorbell is ringing will typically also carry the information that the doorbell's button is being pressed. In contrast, it seems that intentional content is more constrained: I can have a belief that the doorbell is ringing without having the belief that the doorbell's button is being pressed. Therefore, to give an account of intentional content, Dretske needs to rule out the sort of informational nesting found in the example described. To do this, he stipulates that the information that p is nested in the information that q just in case q carries the information that p (71), and then claims that a signal S has the fact that p as its semantic content iff:

  1. S carries the information that p, and
  2. S carries no other information, q, such that the information that p is nested (nomically or analytically) in q (185).

Dretske emphasizes that his notion of information transmission - hence also his notion of intentional content - presupposes a counterfactual-supporting connection between the signal and the information it carries. As a result, a signal correlated with p will fail to carry the information that p if the correlation is merely accidental or statistical: my thermometer carries information about the temperature of my room and not yours, even if the two rooms are the same temperature, because the state of my thermometer supports counterfactuals about the temperature of my room but not about the temperature of your room (that is to say, it is a true generalization that if the temperature of my room were different, the state of my thermometer would be different; in contrast, it is not generally true that if the temperature of your room were different, the state of my thermometer would be different).

The view of intentional content set out so far is plausibly thought of as meeting the naturalism requirement (putting aside worries about its interpretation of probabilities raised by Loewer - see above), but it faces challenges concerning the desiderata of grain and misrepresentation.

First consider the problem of grain. While Dretske's requirement of counterfactual support allows him to set aside merely correlated events in determining the content of a signal, it will not allow him to choose between properties whose covariation is necessary. (I assume, following Dretske, that our underlying theory of event individuation allows for a distinction between instantiations of properties that necessarily covary.) Suppose that a signal carries the information that p, but that it is (nomically, metaphysically, analytically, or logically) necessary that p covaries with a distinct property q; in any of these cases, it will be nomically necessary that p covaries with q. (A special case comes from Quine's famous `gavagai' puzzle from [Quine, 1964]. Quine argued that if a field linguist encountered the term `gavagai' in an unfamiliar language and noted that natives assented to the use of this term when and only when rabbits were present, there would be no fact of the matter which of many incompatible English translations of the term is correct. Live possibilities for the translation, according to Quine, include the following: `rabbit', `undetached rabbit part', `instantaneous temporal stage of a rabbit', `instance of the universal rabbithood', and `part of the scattered mereological sum of all rabbits'. This provides a special case of the problem under discussion because it is necessary that the properties picked out by these expressions covary. See the discussion of the `gavagai' puzzle in the context of informational theories of content in [Gates, 1996].) In this case, the signal will also carry the information that q. Presumably the intentional content that p can be distinct from the intentional content that q; so which is the intentional content of the signal? Dretske is prepared to admit that a signal cannot have one of these contents without having the other (264, note 2).

This admission has struck many as counterintuitive; however, the problem is even more serious than Dretske's admission would suggest. In fact, Dretske's account has the result that (not both, but) neither of the two pieces of information considered can be the content of any signal. In the case described, the information that p is nomically nested in the information that q, and the information that q is nomically nested in the information that p. But on Dretske's account, the intentional content of a signal cannot be a fact that is nomically nested in some other piece of information carried by that signal, so no signal can have either the content p or the content q. (It is left open that a signal could have the disjunctive content (p q)). This problem is obviously quite general, and so is a serious objection against Dretske's account.

Second, it seems that the account spelled out so far cannot accommodate the possibility of misrepresentation. This is because, according to that account, a signal S with the intentional content p must carry the information that p, and this requires that the conditional probability of p given S is 1 - i.e., p must be true. (Cf. the discussion of the so-called ``disjunction problem'' in [Fodor, 1990d]: if both p and q can cause Ss, how can a theory of content make it the case that q-caused Ss have the erroneous content p rather than the always veridical disjunctive content (pq)?) Thus, on the theory we have considered so far, intentional contents cannot misrepresent the world.

Dretske is aware of the problem of misrepresentation, and attempts to answer it by proposing that only some of the tokenings of a signal carry the information that determines that signal's content. In particular, he proposes that there is a learning period for a signal, during which that signal carries the information that p, and that the signal's intentional content is given only in terms of the information it carries in its learning period. After the learning period, when the intentional content of the signal has already been fixed as p, tokenings of that signal can fail to carry the information that p, and so can be erroneous:

In the learning situation special care is taken to see that incoming signals have an intensity, a strength, sufficient unto delivering the required piece of information to the learning subject.... But once we have meaning, once the subject has articulated a structure that is selectively sensitive to [the information that p]..., instances of this structure, tokens of this type, can be triggered by signals that lack the appropriate piece of information (194-195).

Dretske's answer to the problem of misrepresentation raises a number of problems of its own. First, it has seemed to many implausible that there is anything like a principled distinction between the learning period and the non-learning period for most signals. A second concern is that, even if there is a learning period for signals, this period must be characterized non-intentionally if the naturalism constraint is to be respected. This requirement precludes understanding the learning period for a signal simply as the period leading up to that signal's having the intentional content p, and it is not obvious that there is an alternative naturalistically acceptable understanding in the offing. Third, relying on the learning period to explain misrepresentation leaves Dretske without an account of how unlearned (innate) signals could misrepresent. Fourth, as [Loewer, 1997] notes, the account in terms of a learning period is implausible for many signals; for instance, it seems possible that a child could learn that the linguistic symbol `aardvark' has aardvarks (not pictures of aardvarks) as its content, even if all tokens of `aardvark' in the learning period carry information about pictures of aardvarks rather than aardvarks.

3  Epistemic Optimality

Many are convinced by consideration of these difficulties that content can't be reconstructed in terms of information alone. However, many believe these problems can be solved by a hybrid theory that appeals to both an informational factor and some other (not strictly informational) factor. The thought is that, so long as both the informational and the non-informational factors can be given naturalistic explications, they can work together to provide a naturalistic account of content that escapes the vulnerabilities of more strict informational theories.

One family of theories of this sort appeals to a notion of epistemic optimality to take up the slack left by information. On these views, a signal S has the content p iff there are epistemically optimal conditions Cp for p such that if Cp obtained, then S would nomologically covary with p.

Proponents of such accounts, including [Stampe, 1975], [Stampe, 1977], [Dretske, 1983a], and [Fodor, 1990b], hope that their appeal to epistemic optimality might resolve the problem of misrepresentation; this thought is motivated by the (reasonable) suggestion that misrepresentation occurs when cognitive systems attempt to represent the world while operating in epistemically sub-optimal conditions. Thus, for example, tokens of the linguistic symbol `aardvark' that are caused by armadillos seen on dark nights have as their content the property aardvark - rather than armadillo on a dark night or aardvark or armadillo on a dark night - because the symbol `aardvark' would covary only with instances of aardvark in epistemically optimal conditions (epistemically optimal conditions for aardvark presumably involve the lighting being up, the subject's attentively looking in the right direction, and so on).

It has also been suggested that appeals to optimality could solve the problem of grain. For even if p and q covary, we could say that a symbol has p rather than q as its content if S would nomologically covary with p but not q in epistemically optimal conditions for p. However, it is unclear that this appeal to epistemic optimality resolves the problem of grain. For one thing, this solution will fail if the optimal conditions for p (Cp) are identical to the optimal conditions for q (Cq); for in this case, S would, once again, nomologically covary with both p and q in Cp (and Cq, of course), and so would not determinately have the content p. A variant of this worry arises when `p iff q' is necessary. For, here again, S would covary with both p and q in all possible conditions, a fortiori in Cp (assuming condition Cp is possible; if not, then it would not be true that S nomologically covaries with p in Cp, so S could not have the content that p).

Moreover, the naturalistic credentials of epistemic optimality theories are questionable as well. This can be seen in two ways. First, it is plausible that the epistemic optimality conditions for a content p cannot be stated without adverting to the content p itself, since what is epistemically optimal seems to depend on what content we're hoping to reconstruct. For example, the optimality conditions for the belief that there's an aardvark in the room preclude looking through a microscope, but the optimality conditions for the belief that there's a paramecium in the room require looking through a microscope. But, of course, epistemic optimality versions of informational theories explain the content of signals in terms of the optimality conditions for that content. Consequently, the understanding of a signal's having the content p provided by an epistemic optimality theory must be stated in terms of the content p, contrary to the naturalism requirement. Second, insofar as belief fixation is widely thought to be a holistic enterprise, it is equally plausible that the epistemic optimality conditions for a content p cannot be stated without adverting to contents other than p. For example, my tokenings of the linguistic symbol `there are aardvarks in the room' won't nomologically covary with the presence of aardvarks in the room if I believe that aardvarks are not macroscopically observable. But if so, then an epistemic optimality theory's unpacking of what it is for a signal to have the content that p must advert to states that must be characterized by their contents. And once again, this seems a clear violation of the naturalism constraint.

4  Teleology

An alternative elaboration of the informational approach to content, appealing to considerations of teleology, rather than epistemic optimality, is advocated by several philosophers (see [Fodor, 1984], [Fodor, 1990b], [Millikan, 1984], [Millikan, 1986], [Millikan, 1989], [Dretske, 1988], [Papineau, 1993]). Here, too, the hope is that the theory's non-informational factor - its appeal to teleology - will resolve the problems of misrepresentation and grain that plague stricter informational accounts. Roughly put, the thought is that misrepresentation occurs when a signal S covaries with the information that q even though S's teleological function is to carry the information that p (p q). Similarly for the problem of grain: we could say that a symbol means p rather than q (even if p is satisfied when and only when q is satisfied) if the teleological function of the symbol is to carry the information that p. (These formulations assume - controversially - that teleological functions are assigned to individual signals, rather than the whole cognitive systems sustaining the signal-world covariations taken as basic by informational theories. Unfortunately, I cannot examine the alternative construal here for reasons of space.)

Proponents of teleological accounts need to explain the notion of teleological function naturalistically. Of course, they cannot appeal to the well-understood example of the teleological function of artifact symbols in this context (e.g., the symbols on the face of a pressure gauge have the teleological function of representing pressure), since these instances of teleological function are presumably constituted in terms of content: the gauge has the function of measuring pressure because that's what its makers intended it to do (for similar reasons, appeals to God's intentions to fix teleological functions are off limits to would-be naturalists as well). Instead, these theorists typically propose understanding teleological function in terms of natural selection (see [Wright, 1973], [Millikan, 1984], [Neander, 1991]). On this view, a signal S of type S* has the teleological function of carrying the information that p (in an organism o) just in case earlier tokens of type S* were selected (in o's species) by natural selection because they carried the information that p - that is, S has the function of carrying the information that p in o just in case the carrying of the information that p by earlier tokens of type S* increased the fitness of o's ancestors.

To be sure, questions remain about the naturalistic bona fides of the account just sketched - for example, some object that this formulation buys its naturalism at the price of an implausibly robust conception of natural selection. Moreover, the success of this account depends on the possibility of a naturalistic explanation of how tokens are assigned to signal types. However, the most important objections against teleological accounts allege that they cannot accommodate the desiderata of grain and misrepresentation.

The problem of grain for teleological theories is almost invariably presented in connection with the frog's capacity to snap at flies. Consider an internal state S in the frog that covaries with the presence of flies and mediates his snapping behavior. It may be that the frog's environment E is such that all the local small moving black objects are flies, all the items of frog-food are flies, all the flies-or-bee-bees are flies, and so on. Then state S covaries not only with the presence of flies, but also with the presence of small moving black objects in E, frog-food in E, flies-or-bee-bees in E, and so on. Indeed, state S carries information about all of these; so which (if any) is the content of S? According to teleological theories, the information that p is the content of S iff the carrying of the information that p by other tokens of the same type increased the fitness of the frog's ancestors. But a token's carrying information about any of the candidates considered above would have an equal effect on the fitness of ancestral frogs: snapping at flies, small moving black objects in E, frog-food in E, and flies-or-bee-bees in E would all get exactly the same things into the frog's belly so long as all the small moving black objects (/frog-food/flies-or-bee-bees) in E are flies. As Fodor puts the point,

... it's equally OK with Darwin which way you describe the intentional objects of fly snaps, so long as it's reliable (say, nomologically necessary; anyhow, counterfactual supporting) that all the local flies-or-bee-bees are flies. The point is, of course, that if all the local flies-or-bee-bees are flies, then it is reliable that the frog that snaps at one does neither better nor worse selection-wise than the frog that snaps at the other ([Fodor, 1990d], 73).
This problem is even more pressing when the covariation between candidate contents is not merely nomologically necessary, but metaphysically necessary: it is extremely difficult to see how natural selection could favor a signal with the content aardvark there over one with the content undetached aardvark part there.

The problem of misrepresentation for teleological theories is a consequence of the problem of grain. As we have seen, the teleologist's appeal to natural selection leaves it open that S could mean fly-or-bee-bee rather than fly. But if so, then nothing could make it the case that bee-bee caused S-tokens are erroneous rather than veridical. More generally, if this is right, it is difficult to see what could make it the case that any content-assignment by a teleological theory is erroneous.

5  Asymmetric Dependence

Fodor has proposed another informational account as a way of giving a semantics for expressions in a language of thought (henceforth, Mentalese). (Fodor has presented significantly different versions of the theory over the years; see [Fodor, 1987], [Fodor, 1990a], [Fodor, 1990e], [Fodor, 1994]. For reasons of space, I confine myself here to the formulation in [Fodor, 1990e], Fodor's most complete statement of the view.) The intuition behind this view is that, while a Mentalese symbol can carry information about a number of distinct properties, only one of them is its content - namely, its content is that single property on which its carrying information about all the other candidate properties depends. For example, the intuition runs, while my mental state can carry information about both aardvark and armadillo on a dark night, its content is aardvark because it would not carry information about armadillo on a dark night (or anything else) unless it carried information about aardvark. This intuition is fleshed out in [Fodor, 1990e] in terms of a relation of asymmetric dependence between laws, which is itself specified in terms of a pair of subjunctive conditionals: a law L1 is said to depend asymmetrically on a law L2 just in case (i) if L2 did not hold, then L1 would not hold either, and (ii) if L1 did not hold, then L2 would still hold. (That said, Fodor sometimes suggests that the asymmetric dependence relations are metaphysically basic and not in need of cashing in terms of subjunctive conditionals; see [Fodor, 1990e], 93-95.)

Using this apparatus, Fodor writes (121) that a Mentalese expression S has property p as its content if:

  1. `ps cause Ss' is a law.
  2. Some Ss are actually caused by ps.
  3. For all q p, if qs qua qs actually cause Ss, then qs causing Ss is asymmetrically dependent on ps causing Ss.
This formulation deserves supplementation. (In what follows I shall use capital letters to indicate Mentalese expressions: AARDVARK is a Mentalese expression that, we hope, has the property aardvark as its content.) First, Fodor understands the laws appealed to in this account as ceteris paribus generalizations - generalizations that are true (and counterfactual-supporting) when other things are held equal. As a result, the proposal is not vulnerable to potential counterexamples involving the failure of aardvarks to cause AARDVARKs in dark rooms, for inattentive subjects, and so on: these would be brushed aside as cases where the relevant paria aren't cetera, and therefore as falling outside the intended scope of the laws. Second, Fodor insists that the subjunctive conditionals in terms of which asymmetric dependence is characterized should be read synchronically rather than diachronically (134, note 18). That is, it might be that aardvark-pictures played a diachronically ineliminable role in the ontogeny of my thought (say, because I was taught to token AARDVARK exclusively by a process involving exposure to aardvark-pictures); in this case, aardvarks wouldn't cause AARDVARKs in me were it not that aardvark-pictures had first caused AARDVARKs in me. Still, according to Fodor, the aardvark-picture to AARDVARK nomic link is asymmetrically dependent on the aardvark to AARDVARK nomic link because, he thinks, the synchronic dependence goes in the opposite direction from the diachronic dependence: once the capacity to token AARDVARK is in place, aardvark-pictures wouldn't cause AARDVARKs unless aardvarks caused AARDVARKs.

Asymmetric dependence is designed to resolve the problems of grain and misrepresentation. Take the problem of grain first: why is it that AARDVARK means aardvark rather than nocturnal termite-consuming African burrowing mammal? (For the purposes of illustration, assume it is nomically necessary that these properties covary.) Answer: even if every AARDVARK I token is caused by something that is an instance of both properties, instances of nocturnal termite-consuming African burrowing mammal would not cause AARDVARKs unless instances of aardvark caused AARDVARKs, while the reverse dependency does not hold. Take the problem of misrepresentation next: given that some AARDVARK tokens are caused by armadillos on dark nights rather than aardvarks, why do those tokens have the erroneous content aardvark rather than the veridical content armadillo on a dark night (or the veridical disjunctive content aardvark or armadillo on a dark night)? Once again, asymmetric dependence supplies an answer: armadillos on dark nights cause AARDVARKs only because the former are mistakenly identified as aardvarks, so the nomic link between armadillo on dark night and AARDVARK is asymmetrically dependent on the nomic link between aardvark and AARDVARK (similarly, the link between the disjunctive property and AARDVARKs is asymmetrically dependent on the link between aardvark and AARDVARKs).

Several objections have been leveled against Fodor's theory. Many of these take the form of counterexamples designed to show that Fodor's conditions do not assign the correct content to Mentalese expressions ([Cummins, 1989] chapter 5, [Godfrey-Smith, 1989], [Baker, 1991], [Boghossian, 1991], [Manfredi and Summerfield, 1992], [Adams and Aizawa, 1994]. Fodor has responded to many of these criticisms (see [Fodor, 1990e], the replies of [Loewer and Rey, 1991], and [Fodor, 1994]). Unfortunately, these controversies are often extremely difficult to assess because it is unclear how we should understand the relevant subjunctive conditionals. (The difficulty in assessing alleged counterexamples to Fodor's theory is even more severe if asymmetric dependence is taken as metaphysically basic - see above.)

A further complication is that Fodor intends his theory to provide only sufficient, and not necessary, conditions for a state's having content. As Fodor construes it, his task is only to show how Mentalese expressions could have content - not to show how Mentalese expressions do have content. By reducing his aspirations in this way, Fodor forestalls objections about whether his conditions are necessary for a state's having content, and in particular about whether the subjunctive conditionals he invokes are true:

It's enough if I can make good the claim that ``X'' would mean such and such if so and so were to be the case. It's not also incumbent upon me to argue that since ``X'' does mean such and such, so and so is the case ([Fodor, 1990e], 96).

Whether or not the points considered so far are damaging, the asymmetric dependence theory faces other important obstacles.

First, it is unclear that Fodor's account is acceptably naturalistic. As noted, Fodor understands the laws on which his account rests as true, counterfactual-supporting, ceteris paribus generalizations. However, it seems that the ceteris paribus conditions governing the aardvark-AARDVARK link cannot avoid adverting to content for two reasons (this mirrors an objection discussed in §3). For one thing, just which cetera must remain paria to sustain the token-world links the theory requires depends on what that symbol's content is: instances of aardvark can't cause AARDVARKs in me if I am looking through a microscope, but instances of paramecium can't cause PARAMECIUMs in me unless I am looking through a microscope. For another, just which cetera must remain paria to sustain the token-world links the theory requires depends on what other contents the subject believes: I won't reliably token AARDVARK in the presence of aardvarks if I believe that aardvarks are not macroscopically observable. For both these reasons, it seems that the ceteris paribus clauses appearing in the laws required by the theory ultimately must be cashed in terms of content, and consequently that the theory cannot be stated without adverting to content. If so, then the view does not abide by the naturalism constraint.

Finally, while the asymmetric dependence account is arguably successful in dispatching many instances of the grain problem, it remains helpless to mark distinctions in content between properties whose covariation is metaphysically necessary. Because it is metaphysically necessary that instances of aardvark covary with instances of undetached aardvark part, there is a counterfactual-supporting generalization linking instances of aardvark to AARDVARK tokens iff there is a counterfactual-supporting generalization linking instances of undetached aardvark part to AARDVARK tokens. Moreover, since the covariation between the two properties is metaphysically necessary, this will be so in every metaphysically possible world, so there can be no asymmetric dependence between the two laws. (Presumably Fodor would not respond by appealing to asymmetric dependencies in metaphysically impossible worlds, since those worlds will include many where the dependency goes in the opposite direction from that required by the theory.) Consequently, the theory predicts that AARDVARK (i) has the content aardvark if and only if it has the content undetached aardvark part , or (ii) has the disjunctive content aardvark or undetached aardvark part. (In [Fodor, 1994] (chapter 3), Fodor recognizes this problem, and suggests that a theory of content can settle on a determinate and non-disjunctive content for a Mentalese expression by appeal to the inferential relations of thoughts containing that expression. This solution is criticized in [Gates, 1996] and [Ray, 1997].)

6  Conclusion

Many contemporary philosophers believe that informational theories are the most promising proposals for reconciling naturalism with intentional realism. However, it remains to be shown that there is an informational theory of content that satisfies the constraints of §1. Of course, this does not mean that no informational theory can succeed. It does mean that, so far, appeals to information have not resolved the problem of naturalizing content.

7  Glossary of Key Technical Terms

8  Suggested further reading


[Adams and Aizawa, 1994]
Adams, F. and Aizawa, K. (1994). Fodorian semantics. In Stich, S. P. and Warfield, T. A., editors, Mental Representation: A Reader, pages 223-242. Blackwell, Oxford.

Adams and Aizawa argue that there is no tenable reading of Fodor's asymmetric dependence account by proposing counterexamples (cases where the theory assigns the wrong content, or assigns content where it should not). They also argue that Fodor gets the order of explanation wrong - that asymmetric dependence is a consequence of intentional relations, not the ground of those relations.

[Baker, 1991]
Baker, L. R. (1991). Has content been naturalized. In [Loewer and Rey, 1991], pages 17-32.

Baker argues that Fodor vacillates between different (and inconsistent) versions of his theory as he attempts to handle different problem cases, but that no version is adequate as a general solution.

[Boghossian, 1991]
Boghossian, P. A. (1991). Naturalizing content. In [Loewer and Rey, 1991], pages 65-86.

Boghossian argues that Fodor's asymmetric dependence view collapses onto an optimal conditions theory, and that it is therefore susceptible to the kinds of objections that Fodor himself raises against the latter view.

[Churchland, 1981]
Churchland, P. M. (1981). Eliminative materialism and the propositional attitudes. The Journal of Philosophy, 78:67-90.

Churchland argues that ``folk psychology'' - the commonsense predictive/explanatory understanding of the mental that centrally involves the attribution of intentional states - is a radically and systematically erroneous theory in need of replacement by a scientific psychology pitched at a neural level.

[Cummins, 1989]
Cummins, R. (1989). Meaning and Mental Representation. MIT Press, Cambridge, Massachusetts.

Cummins discusses a number of (informational and non-informational) approaches to content in this slim book, and presents his own ``interpretational'' semantics.

[Dretske, 1981]
Dretske, F. I. (1981). Knowledge and the Flow of Information. MIT Press, Cambridge, Massachusetts.

This is Dretske's first major defense of an informational theory of content (see §2). This book was instrumental in bringing informational approaches to the attention of mainstream philosophers of mind.

[Dretske, 1983a]
Dretske, F. (1983a). The epistemology of belief. Synthese, 55:3-19.

Dretske argues that if a system cannot process information to the effect that something is F, then it cannot represent something as F, a fortiori cannot believe that something is F. He also endorses an epistemic optimality account.

[Dretske, 1988]
Dretske, F. (1988). Explaining Behavior: Reasons in a World of Causes. MIT Press, Cambridge, Massachusetts.

Dretske aims to explain how reasons, citing psychological states, can figure in the explanation of actions, given the (in principle) availability of complete explanations in terms of physical causes. As part of this project, Dretske defends a historical/teleological information-theoretic account of the intentional.

[Fodor, 1975]
Fodor, J. A. (1975). The Language of Thought. Harvard University Press, Cambridge, Massachusetts.

Fodor argues that our best psychological theories presuppose the existence of a language of thought (Mentalese) - a language-like system of symbols with syntactic and semantic properties.

[Fodor, 1984]
Fodor, J. A. (1984). Semantics, Wisconsin style. Synthese, 59:231-250. Reprinted in [Fodor, 1990c], 31-49.

Fodor criticizes views associated with Dretske and Stampe for failing to solve the disjunction problem.

[Fodor, 1987]
Fodor, J. A. (1987). Psychosemantics: The Problem of Meaning in the Philosophy of Mind. MIT Press, Cambridge, Massachusetts.

This book contains Fodor's first published statement of his asymmetrical dependence theory.

[Fodor, 1990a]
Fodor, J. A. (1990a). Information and representation. In Hanson, P., editor, Information, Language, and Cognition, pages 175-190. University of British Columbia Press, Vancouver.

Fodor marks a distinction between ``labelling'' uses (e.g., `that is an aardvark') and ``representational'' uses (e.g., `an aardvark is a beautiful animal') of the same symbol: representational uses are often not intended to indicate that the cause of the symbol has the property expressed by the symbol. He insists that the asymmetric dependence account is intended only to apply to labelling uses of symbols.

[Fodor, 1990b]
Fodor, J. A. (1990b). Psychosemantics or: Where do truth conditions come from? In Lycan, W. G., editor, Mind and Cognition: A Reader, pages 312-338. Blackwell, Cambridge, Massachusetts. Widely circulated as a manuscript circa 1980.

Fodor defends an optimality/teleology informational account. He repudiates this view in [Fodor, 1984], [Fodor, 1990d], and [Fodor, 1987].

[Fodor, 1990c]
Fodor, J. A. (1990c). A Theory of Content and Other Essays. MIT Press, Cambridge, Massachusetts.

This collection contains, among a number of previously published essays, two previously unpublished papers on content that comprise the most worked-out statement of Fodor's asymmetric dependence view.

[Fodor, 1990d]
Fodor, J. A. (1990d). A theory of content, I: The problem. In [Fodor, 1990c], pages 51-87.

Here Fodor discusses the problems that an informational theory of content must solve, and argues that going versions of such theories are unsuccessful at solving them.

[Fodor, 1990e]
Fodor, J. A. (1990e). A theory of content, II: The theory. In [Fodor, 1990c], pages 89-136.

This is Fodor's most complete presentation of his asymmetric dependence theory. Fodor motivates the view, explains and discusses his approach in some detail, and responds to a large number of potential objections.

[Fodor, 1994]
Fodor, J. A. (1994). The Elm and the Expert: Mentalese and its Semantics. MIT Press, Cambridge, Massachusetts.

Fodor engages in a number of projects in this book, derived from his Nicod Lectures of 1994. Among these, he offers an answer to Quinean inscrutability problems in chapter 3 and some substantive revisions and extensions of his asymmetric dependence theory in two appendices.

[Frege, 1892]
Frege, G. (1892). On sense and meaning. In McGuinness, B., editor, Gottlob Frege: Collected Papers on Mathematics, Logic, and Philosophy, pages 157-177. Basil Blackwell, Oxford, England.

In the opening pages of this paper, Frege notes that co-referential expressions are not intersubstitable salva veritate in propositional attitude contexts. This point is at the heart of the grain constraint discussed in §1.

[Gates, 1996]
Gates, G. (1996). The price of information. Synthese, 107(3):325-347.

Gates argues that informational theories of content presuppose a solution to Quinean inscrutability, and that no such solution is available.

[Godfrey-Smith, 1989]
Godfrey-Smith, P. (1989). Misinformation. Canadian Journal of Philosophy, 19:533-550.

Godfrey-Smith claims that Fodor fails to solve the disjunction problem - that he has no way to ensure that the content of wild tokens will depend asymmetrically on that of veridical tokens, as required by the asymmetric dependence theory.

[Grice, 1957]
Grice, H. P. (1957). Meaning. The Philosophical Review, 66:337-388.

Grice distinguishes between ``natural meaning'' - the kind of meaning that we are speaking of when we say something like, `Those spots mean measles' - and ``non-natural meaning'' - the kind of meaning we speak of when we say `Those three rings on the bell mean that he made it safely to shore'. He proposes that the non-natural meaning of symbols can be understood in terms of the intentions of the producers of those symbols. Informational theories of content are proposed to explain (as a kind of natural meaning) the content of the intentions constituting non-natural meanings.

[Loewer, 1983]
Loewer, B. (1983). Information and belief. Behavioral and Brain Sciences, 6:75-76.

Among other complaints, Loewer's brief comment on [Dretske, 1983b] objects that the probabilities to which Dretske appeals cannot be understood in any of the standard ways.

[Loewer, 1987]
Loewer, B. (1987). From information to intentionality. Synthese, 70:287-317.

Loewer argues that extant informational theories are unsuccessful in providing naturalistically acceptable understandings of content.

[Loewer, 1997]
Loewer, B. (1997). A guide to naturalizing semantics. In Hale, B. and Wright, C., editors, A Companion to the Philosophy of Language, pages 108-126. Blackwell, Oxford.

Loewer extends the arguments of [Loewer, 1987] to the effect that extant informational theories of content cannot meet reasonable desiderata.

[Loewer and Rey, 1991]
Loewer, B. and Rey, G., editors (1991). Meaning in Mind: Fodor and His Critics. Blackwell, Oxford.

This book contains a short summary of Fodor's work by the editors, fourteen critical essays on various aspects of his ouvre (including several on his asymmetric dependence theory), Fodor's replies to all of these, and an extensive bibliography of Fodor's writings.

[Manfredi and Summerfield, 1992]
Manfredi, P. and Summerfield, D. (1992). Robustness without asymmetry: A flaw in Fodor's theory of content. Philosophical Studies, 66:261-283.

Manfredi and Summerfield offer another purported counterexample to Fodor's account. They conclude that asymmetric dependence cannot be a general solution to robustness.

[McDowell, 1994]
McDowell, J. (1994). Mind and World. Harvard University Press, Cambridge, Massachusetts.

This book is of interest in the present context principally because it rejects the naturalism constraint of §1. McDowell believes that the normative dimension of content cannot be captured naturalistically, and therefore is unsympathetic to the goal of naturalizing content (and other forms of ``bald naturalism'').

[Millikan, 1984]
Millikan, R. G. (1984). Language, Thought, and Other Biological Categories: New Foundations for Realism. MIT Press, Cambridge, Massachusetts.

Millikan presents a general naturalistic theory of teleological function (one that is intended to apply to body organs, behaviors, and mental representations, inter alia), and defends a teleological account of content.

[Millikan, 1986]
Millikan, R. G. (1986). Thought without laws: Cognitive science with content. The Philosophical Review, 95:47-80.

Millikan holds that references to the content of psychological states are ineliminable in both folk psychology and cognitive science. However, she thinks, psychology is naturalistic because content can be unpacked by appeal to the evolutionary history of systems and states.

[Millikan, 1989]
Millikan, R. G. (1989). Biosemantics. The Journal of Philosophy, 86:281-297. Reprinted in [Stich and Warfield, 1994].

Millikan clarifies, elaborates, and defends the teleological theory of content presented in [Millikan, 1984].

[Neander, 1991]
Neander, K. (1991). Functions as selected effects: The conceptual analyst's defense. Philosophy of Science, 58:168-184.

Neander defends an etiological theory of biological function - a theory on which the function of a trait is the effect for which it was selected by natural selection. This view gives one way of understanding the notion of function that figures centrally in teleological informational theories of content.

[Papineau, 1993]
Papineau, D. (1993). Philosophical Naturalism. Blackwell, Oxford.

This book is a very general defense of naturalism in metaphysics and epistemology. Part II spells out a teleological construal of intentional properties that is intended to respect Papineau's naturalist aims.

[Peirce, 1931]
Peirce, C. S. (1931). Collected Papers. Harvard University Press, Cambridge, Massachusetts.

Volume 2 of this work contains a very early formulation of an informational theory of content.

[Quine, 1964]
Quine, W. V. O. (1964). Word and Object. MIT Press, Cambridge, Massachusetts.

This is the source of the gavagai problem that, according to [Gates, 1996], plagues contemporary informational theories of content.

[Ray, 1997]
Ray, G. (1997). Fodor and the inscrutability problem. Mind and Language, 12(3/4):475-489.

Ray attempts to undermine Fodor's answer to the inscrutability problem in [Fodor, 1994]. He gives several examples to which, he claims, Fodor's answer is inapplicable, and concludes that Fodor's solution cannot secure the scrutability of reference in the general case.

[Stampe, 1975]
Stampe, D. (1975). Show and tell. In Freed, B., Marras, A., and Maynard, P., editors, Forms of Representation: Proceedings of the 1972 Philosophy Colloquium of the University of Western Ontario. North-Holland Publishing Company, Amsterdam.

This is one of the earliest formulations of an informational theory of content, aimed primarily at the understanding of linguistic content.

[Stampe, 1977]
Stampe, D. (1977). Toward a causal theory of linguistic representation. In French, P., Uehling, T., and Wettstein, H., editors, Midwest Studies in Philosophy, volume 2. University of Minnesota Press, Minneapolis.

Here Stampe develops and defends the view of [Stampe, 1975]. This paper is widely cited as a primary source of contemporary interest in informational approaches to content.

[Stich and Warfield, 1994]
Stich, S. P. and Warfield, T. A. (1994). Mental Representation: A Reader. Blackwell, Oxford.

Stich and Warfield have produced a fine collection of previously published papers on mental content. The papers cover a wide range of (informational and non-informational) approaches.

[Wright, 1973]
Wright, L. (1973). Functions. The Philosophical Review, 82:139-168.

Wright proposes that the analysis of biological functions must make reference to the causal background of the bearers of such functions. He argues that this sort of ``etiological'' account avoids difficulties that plague competing analyses and secures several important desiderata for the notion of function. Proponents of teleological accounts of content have often relied on accounts of function in this tradition.

Jonathan Cohen is an Assistant Professor of Philosophy at the University of California, San Diego. He is the author of several articles in philosophy of mind and philosophy of language. Much of his recent work has concerned the metaphysics of color.

File translated from TEX by TTH, version 3.01.
On 10 Oct 2001, 17:57.