CAN SELECTION EXPLAIN CONTENT

CAN SELECTION EXPLAIN CONTENT?

Pierre Jacob

jacob@ehess.fr

Typical things exemplifying content properties are representations. Representations can be mental or non-mental. Content can be conceptual or non-conceptual.[1] Philosophers who embark on the project of naturalizing content - or intentionality - typically subscribe to some version or other of physicalism. So they assume that things with content (or semantic) properties also instantiate non-semantic (physical, chemical, biological, physiological, informational and perhaps syntactic) properties. Furthermore, they subscribe to some kind of reductionism: to naturalize content is to provide a non-semantic explanation of the fact that a representation has the content property that it has. Thus, the project of naturalizing intentionality assumes that the fact that a representation has content is not a surd or ultimate metaphysical fact. Only the instantiation of a non-semantic property can be a surd metaphysical fact. This is why a representation must derive its content from some of its non-semantic property or properties or why some non-semantic property (or set of non-semantic properties) must underlie the representation's semantic property. Equivalently, the fact that a representation has the content that it has must depend (or supervene) on some of the representation's non-semantic properties. Finally, it is fair to say, I think, that philosophers who embark on the project of naturalizing content are intentional realists: they assume that a representation has a determinate content.[2] Here, I won't try to justify physicalism, reductionism about content or intentional realism. I will merely accept them. And I will assume that the project of naturalizing intentionality makes sense.[3]

There are presently three broad approaches to this task: a purely informational approach, a purely teleological approach and a mixed informationally-based teleological approach. On the purely informational (or correlational) approach whose main representatives are Dretske (1981) and Fodor (1987, 1990, 1994, 1998), for x to mean e.g., DOG is for it to stand in an informational relation to instantiations of doghood. There are then two ways of capturing what it takes for a symbol to stand in an informational relation to an instantiated property. According to Dretske (1981), x will not mean DOG at t unless the conditional probability that doghood was being instantiated at t - 1 if x is being tokened at t is 1. According to Fodor (1994), for a symbol to mean DOG is for tokens of this symbol to nomically covary with instantiations of doghood. On the second purely teleological approach - henceforth pure teleosemantics - whose main representatives are Millikan (1984, 1993) and Papineau (1987, 1993), the content of a representational device is basically the property which in the course of history contributed most to the success or the proliferation of the device. Finally, on the mixed informationally-based teleosemantic theory advocated by Dretske (1988), the content of a device is the device's indicator-function - what the device is supposed to carry information about.

I should say at the outset that my sympathies lie with the last mixed approach.[4] The main attraction of a teleological component to a theory of content is, I think, that it offers a promising account of the possibility of misrepresentation in non-semantic terms. The length of a simple metal bar carries information about temperature because it nomically covaries with temperature. But it cannot represent the temperature. It cannot represent the temperature because it cannot misindicate the temperature. Unlike a metal bar, a mercury thermometer can represent the temperature. What is the difference? The difference is that, unlike a metal bar, the height of the column of mercury in a thermometer may misrepresent the temperature. The reason why a thermometer, unlike a metal bar, may misrepresent the temperature is that the former has, while the latter lacks, a function: its function is to indicate the temperature. Only a device with an indicator function can misrepresent and therefore represent. It may misrepresent the temperature by failing to indicate what it is its function to indicate. As many philosophers have pointed out, no representation without misrepresentation.

If explaining the possibility of misrepresentation is to account for the normativity of content, then teleosemantics does, I think, have the potential to account for the normativity of content in non-semantic terms. I also think, however, that the most explicit version of a pure teleosemantic theory of content, namely Millikan's (1984, 1993) theory, faces a pair of objections. My goal in this paper is to spell out Millikan's pure teleosemantic theory; then to present two objections; and finally to ask the question whether a teleosemantic framework can be saved from the objections. As will appear, there is, I think, something paradoxical about her views: pure teleosemantics is tailored to account for the non-conceptual content of simple sensory representational systems of organisms without language or conceptual resources. But it violates what I take to be a plausible requirement on the contents of such sensory mechanisms (if they have any). My thesis in a nutshell is that Millikan's pure teleosemantic theory confuses the explanation of the proliferation of a representational device with an account of the device's content.

Broadly speaking, at the heart of any teleosemantic conception of content is the view that a representational device owes its content to the device's proper function or teleofunction where proper function is taken to be a teleological notion.[5] A device's proper function is a normative, not a merely causal or dispositional, notion: it is what the device is supposed to do, not what it actually does or is disposed to do.[6] To say that the proper function of a mammal's heart is to pump blood is to say that it is supposed to pump blood whether or not it does pump blood, let alone efficiently. If diseased or malformed, a mammal's heart may be incapable of pumping blood. Nonetheless - so long as it is a heart -, pumping blood is what it is supposed to do.

Most teleosemanticists further think - I among them - that the best account of a device's teleofunction(s) or proper function(s) is the etiological account elaborated and developed by Wright (1973), Millikan (1984, 1993), Neander (1991), Godfrey-Smith (1993) and others, according to which Z is the proper function of trait X in system S iff Xs were selected for doing Z because X's doing Z contributed to the proliferation of S's ancestors.[7] The etiological proper function of a device, then, results from a selection process: a device's proper function is what the device was selected for.[8] The selection process may be intentional or non-intentional. Artifacts derive their proper functions from an intentional process. Naturally evolved traits of living things derive their proper functions from natural selection which is a non-intentional process. So do sensory and cognitive systems such as an animal's visual system or its belief-forming mechanisms. Sensory and/or cognitive devices derive their proper functions from either natural selection or learning. While a sensory or cognitive device may have a direct proper function, an individual's particular states - his or her visual experiences and his or her particular beliefs - will have derived (adapted or relational) proper functions.

Three ingredients of Millikan's approach make it the most explicit theory of proper functions. First, she starts with the historical notion of a reproductively established family. My genes and my father's genes are members of a first-order reproductively established family: the former are direct copies of the latter. My heart, however, is not a direct copy of my father's heart. But my heart was produced by my genes; my father's heart was produced by his genes. Since our genes form a first-order reproductively established family, our hearts form a higher-order reproductively established family. Secondly, she derives a device's proper function from the fact that the device belongs to some reproductively established family or other. For something Z a device X does to be its direct proper function, Z must be something that X's ancestors (which were members of X's reproductively established family) did which contributed to the historical proliferation of the family. This is like a reproductive condition on proper functions. Third, Millikan's account of proper function appeals to the notions of Normal explanation and Normal conditions for performance of a function. Although a pacemaker does contribute to circulating an individual's blood and it is the function of a pacemaker to circulate the blood of the individual to whose heart it is attached, still a pacemaker does not perform its function in accordance with a Normal explanation in Normal conditions since pacemakers did not contribute to the historical proliferation of hearts in a reproductively established family.[9] Of course, "Normal" is a teleological (hence a normative), not a statistical, notion. As Millikan fondly reminds us, a Normal condition for a sperm cell to perform its proper function is to be the first of a battalion of sperm cells to reach an ovum. This is not a condition frequently encountered in the lives of many sperm cells.

Finally, when she turns her attention to the determination of a device's content, Millikan assumes that there are three things to distinguish: the representation (or representational state), the device M which produces the representation and the device M* which uses (or consumes) the representation.[10] On her view, the content of the representation rests on a collaboration between on the one hand the Normal conditions involved in a Normal explanation of how the consumer device M* performs its proper function and on the other hand the proper function of the producer mechanism M. To go to the heart of the matter, on Millikan's view, content arises from a delicate interplay between the Normal conditions mentioned in a Normal explanation of the proper functions of both the consumer system and the producer system. But there is no doubt that Millikan's key idea here is that the proper function of the producer mechanism is to help the consumer device achieve its proper function (see e.g., Millikan, 1993: 93). Writing about Dretske's famous marine bacteria with a magnetosome, she writes:

What the magnetosome represents is only what its consumers require that it correspond to in order to perform their tasks. Ignore, then, how the representation [...] is normally produced. Concentrate instead on how the systems that react to the representation work, on what these systems need in order to do their job.

So given that the proper function of the producer mechanism M is to help the consumer mechanism M* achieve its proper function, content then arises from the primacy of the Normal conditions mentioned in a Normal explanation of the way the consumer achieves its proper function. Perhaps it could even be said that on the consumer-based teleosemantic theory, the proper function of the producer is part of the Normal conditions involved in the Normal explanation of the way the consumer performs its proper function. In any case, what Millikan calls a "most proximal (or proximate) Normal explanation" of the performance of the consumer's proper function may abstract away from the process whereby the producer device delivered the representation.

Millikan's theory can thus be legitimately called a consumer-based teleosemantic theory. Clearly, the consumer-based theory can neatly solve the problem of the possibility of misrepresentation. Nor is it, it seems, plagued by indeterminacy of any sort. When a frog sticks its tongue out, the content of the frog's visual system M is FLY. This is so because catching flies is the proper function of the frog's (consumer) snapping motor system M*: the property of being a fly is part of the Normal conditions mentioned in the Normal explanation of how the frog's snapping motor system contributed to the proliferation of creatures with such motor systems. The property of being a fly must have been instantiated in the history of ancestors of frogs with visual system M. On the consumer-based theory, it is that property, not the property of being a small black moving dot, which helps explain why creatures with (producer) visual system M and (consumer) snapping motor-system M* were selected for. Now, when a frog sticks its tongue out in response to a lead pellet, then its visual system M is misrepresenting a pellet as a fly. Similarly, when a male hoverfly zooms upon a mating target, the content of its visual system is something like MATE (or FEMALE). The instantiation of the property of being a female hoverfly in the history of hoverflies helps explain why male hoverflies with visual system M and zooming motor system M* were selected. When a male hoverfly zooms in the direction of a bird or a large distant plane - as it apparently does -, then its visual system misrepresents a bird or a large distant plane as a female conspecific.

III

One difficulty with Millikan's consumer-based teleosemantic theory, as noted by Neander (1995), is that of spelling out the Normal conditions to be mentioned in the Normal explanation of the way the consumer performs its proper function. Surely, being the property of being a fly is part of the Normal conditions involved in a Normal explanation of how the frog's snapping behavior historically contributed to the proliferation of frogs. But so is the property of being nutritious. And flies historically ingested by frogs were nutritious in virtue of being composed of proteins. So why should the property of being composed of proteins not be parts of the Normal conditions for a Normal explanation of how the frog's snapping behavior historically contributed to the proliferation of frogs? Besides, it was not just successfully snapping at flies that explains the proliferation of frogs; ingested flies must have been healthy, i.e., not infected by a bacterium.[11]

I do not think that Millikan is really entitled - as she suggests (1993: 127-28) - to appeal to what the producer mechanism can actually do in order to select among the set of necessary conditions the subset of Normal conditions. Of course, the frog's visual system cannot discriminate a healthy fly from an infected fly. Nor is it sensitive to the fact that flies are composed of proteins. But it cannot discriminate a fly from a lead pellet either. So it seems that Millikan is faced with a dilemma: either the state of the frog's visual system means FLY even though the frog's visual system cannot tell a fly from a lead pellet. But then the content of the frog's visual system is not really limited by the causal powers of the frog's visual system. Or the content of the frog's visual system does depend on its causal powers. But then the content of the frog's visual system should be SMALL BLACK MOVING DOT. What Dretske (1981) calls the problem of "channel conditions" - i.e., the problem of spelling out the difference between what a signal carries information about (the source) and the channel over which the information is transmitted from the source to the signal - is a difficult problem for informational semantics.[12] Its counterpart - the problem of Normal conditions - is a difficult problem for a pure teleosemantic theory. As far as I can see, the consumer-based teleosemantic theory cannot appeal to what the producer mechanism can really achieve without giving up the crux of the consumer-based account.

I will now state my main objection: the consumer-based teleosemantic theory generates implausible content ascriptions. In a nutshell, there is a conflict between the consumer-based teleosemantic assignment of content to a sensory mechanism and the idea that a sensory mechanism can only detect what it causally interacts with or nomically covaries with. Let us call this principle the Nomic Correlation Principle (NCP):

[NCP] Unless its tokenings are nomically correlated with instantiations of property F, sensory mechanism M cannot represent property F.[13]

The NCP should be reminiscent of causal and/or reliabilist theories of knowledge. Granted, my present topic is content, not knowledge. However, I do believe that something like the NCP must be involved in fixing the content of a sensory mechanism - if it has one. In other words, I assume that the content of a sensory mechanism depends somehow on the mechanism's discriminatory abilities.

Insofar as what we want to determine is the content of the perceptual state of a sensory mechanism, then in virtue of the NCP, it is implausible that the content of a state of the frog's visual system be FLY when the frog's visual system cannot discriminate a bug from a lead pellet. Similarly, it is implausible that the content of a state of the male hoverfly's visual system is FEMALE CONSPECIFIC when the hoverfly's visual system cannot discriminate a female conspecific from a bird or a distant jet plane. On this view too, although they are magnets (hence magnetotactic, not chemical devices), the magnetosomes of Dretske's famous marine bacteria in the Northern hemisphere represent the direction of toxic anaerobic (oxygen-free) water, not geomagnetic north, since it is the former, not the latter, property which helps explain the proliferation of Northern marine bacteria with a magnetosome in the Northern hemisphere. Obviously, the NCP preserves the fact that a magnetosome is a magnetotactic, not a chemical device, as I think it should.

Consider Pietroski's (1992) Gedanken example of fictitious creatures, the kimus, who live near a hill and are preys of fictitious predators, the snorfs. At some point in their evolution, kimus come to have a visual sensory mechanism which makes them sensitive to red. As it happens, kimus with this sensory mechanism are fond of redness. As a result, each morning, they move towards the top of the hill where redness is instantiated. As snorfs never get up the hill, by so climbing up the hill, kimus avoid snorfs. An area which is free of snorfs happens to coincide with a source of redness. Kimus are sensitive to red, but a kimu would, so the Gedanken experiment goes, never tell visually or otherwise a snorf from an elephant. When a kimu moves towards the top of the hill in the morning, what is the content of the state of its visual system? On the consumer-based theory, the answer is bound to be something like: NO SNORF THIS WAY or SNORF FREE THERE or NO PREDATOR THERE or NO DANGER THERE. If we want to provide a Normal explanation of how the kimu motor system M* performed its proper function historically, we certainly must mention the property of the hill which the kimu is currently climbing of being free of snorf. It is the fact that the hill instantiates this property, not the property of being close to a source of redness, which contributes to explaining why kimus with a red-detector proliferated. But according to the NCP and since kimus cannot recognize a snorf, the content of the state of their visual system cannot have SNORF as an ingredient. If kimus had language, they would have a word for red, not for snorfs. Kimus enjoy visual experiences of red. There seems to be no sense in which their visual system is processing information about snorfs.

In all of the above cases, we use the NCP in order to determine the content of the state of a creature's sensory mechanism M. M responds nomically to instantiations of property F. Instantiations of property F may be in some weaker or stronger sense correlated with instantiations of another property. Call it G. Instantiations of property G, not F, are involved in the Normal explanation of the performance of the creature's consumer - motor - system M*. So instantiations of G, not F, contribute to explaining the proliferation of creatures with M. The consumer-based teleosemantic theory seems committed to the view that the state of the sensory mechanism M represents G, not F, in contradistinction with the NCP. This, I claim, is undesirable given the fact that in all of the above cases, the creatures have no means of telling whether G is instantiated other than by responding to instantiations of F-ness. If there is a phenomenology to being in an M-state (for any of these creatures) - something I am definitely not committed to -, then certainly it is the phenomenology of responding to the presence of Fs, not of Gs.

Whether we put it in terms of sensory discriminatory abilities or in terms of phenomenology, the point is that, in virtue of the NCP, an M-state cannot represent G. It may be beneficial for S to do something when G is instantiated. But if S can only tell if and when G is instantiated by telling whether F is instantiated and not vice-versa in virtue of entering into an M-state, then what the M-state represents is F, not G. The point is that there is a tension between a pure teleosemantic theory of content and what I take to be a plausible requirement on the content of a sensory mechanism, i.e., as the NCP requires, that the latter must depend on what the mechanism nomically covaries with.[14]

Perhaps one way to amend the pure consumer-based theory would be to accept a pair of distinctions recently recommended by Rowlands (1997). Suppose we distinguish two proper functions of a sensory mechanism M, one of which Rowlands calls its algorithmic proper function and the other of which he calls its organismic proper function. We can, I think, illustrate Rowlands' distinction in terms of Neander's (1995) notion of nested function: some device M in the antelope may allow the antelope to survive at higher ground by increasing the antelope's uptake of oxygen. The former is M's algorithmic proper function. The latter is M's organismic proper function.[15] Presumably, M has its organismic proper function in virtue of having its algorithmic proper function, and not vice-versa. I take it that it should be a constraint on this dualistic theory of content - algorithmic and organismic contents - that there must be a story whereby the latter can be derived from the former. On the basis of this distinction, Rowlands wants to build up a distinction between two kinds of contents: the content attributable to the state of the sensory mechanism M and the content attributable to the whole organism. The former content of the mechanism would be underwritten by M's algorithmic proper function; the latter organismic content would be underwritten by M's organismic proper function.

The idea would be that we would have the best of both worlds: the algorithmic content would satisfy the NCP. And the organismic content could fit the consumer-based teleosemantic theory. On this view, RED should be the content of kimus' sensory mechanisms, while SNORFLESSNESS would be the organismic content of a whole kimu. SMALL BLACK MOVING DOT would be the content of the frog's visual state, while FLY should be the organismic content of the whole frog. Well, should it or could it?

Consider the rattlesnake. The rattelsnake is a predator of mice and little rodents. It has two sensory systems: a warmth detector and a motion detector. Let's assume that the output of both detectors serves as input to a supervising module M. Only when both detectors are jointly stimulated does M in turn instruct the rattlesnake's motor system to engage into hunting behavior. In the environment of ancestors of current rattelsnakes, often enough ancestors of current mice were the distal source of the joint instantiations of warmth and movement. A Normal explanation of how the rattlesnake's hunting motor system performed its proper function will advert to the instantiation of the property of being a mouse. Armed with Rowlands' distinction, we might want to distinguish the content of an M-state from the organismic content attributable to the whole rattelsnake R: the former would represent warmth-and-movement; the latter would represent mice.

But I now want to argue, Rowlands's distinction cannot really dissolve the tension between the NCP and the consumer-based teleosemantic theory. As Rowlands recognizes, rattelsnakes cannot discriminate mice from rats, moles, voles and frogs. So MOUSE is too specific an organismic content to be attributable to a whole rattlesnake.[16] This is why on Rowlands' view, the content of the rattlesnake M-state is EATABILITY, not MOUSE.

But then first of all, it is as mysterious on this view as on Millikan's how the whole organism can derive the organismic content EATABILITY from the fact that the mechanism represents warmth and movement. The mechanism delivers a sensory representation of warmth and movement. If rattlesnakes represent EATABILITY, they must presumably derive it from a sensory representation of warmth and movement. Since rattelsnakes have no other sensory access to EATABILITY than by means of their sensory representation of warmth and movement, EATABILITY must be conceptually represented by rattelsnakes. Now, unless rattelsnakes have the appropriate inferential and conceptual resources to infer the concept EATABILITY from a sensory representation of warmth and movement, I do not see how they could represent EATABILITY. Since I don't think it's plausible that rattlesnakes have such conceptual and inferential resources, I don't think that they can represent EATABILITY.

Second of all, on this view, all states of a predator's sensory mechanism will share the very same content. They will share the same content whatever the differences among the mechanisms underlying different sensory systems and whatever the differences in the distal sources of stimulation. Whether the predator is a frog which occasionally gulps a fly by visually responding to a small black moving dot or whether it is a rattlesnake which occasionally swallows a mouse by combining the visual detection of movement and the thermal response to warmth.

Suppose property F is what a creature's sensory mechanism M can detect or respond to by nomically covarying with instantiations of F. Suppose in turn that instantiations of property F are correlated with instantiations of property G. If there is something useful (or beneficial) that creatures with mechanism M can do which will contribute to their proliferation when G is instantiated, then it is property G which helps explain the proliferation of creatures having M. Detection of F contributes to the creatures' proliferation insofar as F and G are correlated and detection of F allows the creatures to do something beneficial in the presence of G. But given their sensory limitations, it is only by means of their representation of F-instantiations that such creatures can tell when to act appropriately when G is instantiated. In those circumstances, what M-states represent - what they can represent - is that F, not that G, is instantiated. From the fact that the creature will benefit from engaging in a certain behavior when G is instantiated, I do not conclude that the content of an M-state is G.

As Pietroski (1992: 274) points out, instantiations of property G can explain why a sensory mechanism was selected and why creatures with such a mechanism proliferated even though there never was any direct causal interaction between instantiations of G and mechanism M. But, as the NCP stipulates, a mechanism cannot deliver sensory representations of a property unless the sensory mechanism can track the property, i.e., unless there are causal interactions (or a nomic correlation) between instantiations of the property and the sensory mechanism. In the envisaged circumstances, M does causally interact with instantiations of F, not G. So M-states can deliver sensory representations of F.

It certainly looks as though any theory of content based on selection processes will appeal to the adaptive advantages which are conferred onto a creature by its sensory mechanism in order to determine the content of the mechanism's states. So it may well seem as if a selection based telesosemantic theory is doomed to conflict with the NCP. However, this conclusion would, I think, be premature. I will now proceed to explain why I think so. In other words, I will briefly sketch my own attempt at reconciling teleosemantics with the NCP.

Let me remind you that M is a sensory mechanism in creature S. I will assume that M's being nomically correlated with F - as it must if M is to deliver sensory representations of F - is necessary though not sufficient for M-states to represent F. For M-states to represent, and not merely to indicate, F, they must be recruited by evolution as causes of output m* of S's motor system M*.[17] Being so recruited as causes of motor output m*, M-states acquire a (derived) proper function: the function to indicate F. This in essence is the gist of Dretske's (1988) mixed informationally-based teleosemantic theory. As I said, only a state with a function can misrepresent, hence represent, things. In cases of interest, evolution will hire an M-state as a cause of motor output m* if producing m* contributes to the proliferation of creatures with M in circumstances G. By so doing the M-state comes to represent F because it acquires the function of indicating F.[18]

On the consumer-based pure teleosemantic theory, what explains the proliferation of creatures with mechanism M just is what accounts for the content of an M-state. Explaining the former is accounting for the latter. This, I think, is a mistake. Explaining the proliferation of a mechanism is not thereby accounting for its content, if it has one.

I start with the notion of indication: M-states indicate F. They acquire the function to indicate and hence represent F by being hired as causes of m* (the output of S's motor system M*) because doing m* is adaptive when G is instantiated. Often enough instantiations of F coincide with instantiations of G. I agree with pure teleosemanticists that what explains the proliferation of creatures with an M-M* coordination is that it is adaptive to do m* when G is instantiated. But on my view, what was selected by evolution were not creatures with a sensory mechanism capable of representing Gs. On my view, what was selected by evolution were creatures S with a coordination between a sensory mechanism M representing Fs and a motor system M*. On the consumer-based pure teleosemantic theory, the target of selection is the representational power of mechanism M. On my view, the target of selection is the coordination between sensory system M and motor system M*. This coordination was selected in an environment in which instantiations of F were well enough correlated with instantiations of G. In such an environment, producing m* was adaptive for S when G was instantiated. On my view then, when S's motor neurons - states of S's motor system M* - are supplied information by M-states (representations of Fs), they do not need to know about the correlation between properties F and G. In a word, instantiations of property G, not F, helps explain the proliferation of creatures with mechanism M. But again explaining the proliferation of M is not fixing the content of M-states. Given the creatures' sensory limitations, it is only by means of their representation of F-instantiations that such creatures can tell when to act appropriately, i.e., when G is instantiated.

According to pure consumer-based teleosemantics, if acting when G is instantiated is adaptive for creature S - if instantiations of G are beneficial to S -, then S must have a way of representing instantiations of G. On my view, when S's motor neurons fire and produce output m*, they are instructed to do so by S's sensory neurons which in turn represent instantiations of F, not G. If instantiations of F and G are well correlated, then when S produces m*, S knows more often than not that G is instantiated. But S knows this indirectly - by means of its representation of F - not by harboring any representation of Gs. Although S must know indirectly when G is instantiated, M does not need to "know" this. So, I claim, we can have our cake and eat it. Teleology is after all compatible with the NCP. We can have teleology without turning M-states into representations of G.

References

Cummins, R. (1975) "Functional Analysis", in E. Sober (ed.)(1984) Conceptual Issues in Evolutionary Biology, Cambridge, Mass.: MIT Pess.

Cummins, R. (1983) The Nature of Psychological Explanation, Cambridge, Mass.: MIT Press.

Dennett, D. (1987) The Intentional Stance, Cambridge, Mass.: MIT Press.

Dretske, F. (1981) Knowledge and the Flow of Information, Cambridge, Mass.: MIT Press.

Dretske, F. (1988) Explaining Behavior, Cambridge, Mass.: MIT Press.

Evans, G. (1982) The Varieties of Reference, Oxford University Press.

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

Fodor, J.A. (1990) A Theory of Content and Other Essays, Cambridge, Mass.: MIT Press.

Fodor, J.A. (1994) The Elm and the Expert, Cambridge, Mass.: MIT Press.

Fodor, J.A. (1998) Concepts, Where Cognitive Science Went Wrong, Oxford: Oxford University Press.

Godfrey-Smith, P. (1993) "Functions: Consensus Without Unity", Pacific Philosophical Quarterly, 74, 3, 196-208.

Jacob, P. (1995) "Consciousness, Intentionality and Function. What Is the Right Order of Explanation?", Philosophy and Phenomenological Research, LV, 1, 195-200.

Jacob, P. (1997) What Minds Can Do, Cambridge: Cambridge University Press.

McDowell, J. (1994) Mind and World, Cambridge, Mass.: Harvard University Press.

Millikan, R.G. (1984) Language and Other Biological Objects, Cambridge, Mass.: MIT Press.

Millikan, R.G. (1993) White Queen Psychology and Other Essays for Alice, Cambridge, Mass.: MIT Press.

Neander, K. (1991) "The Teleological Notion of 'Function'", Australasian Journal of Philosophy, 69, 4, 454-68.

Neander, K. (1995) "Misrepresenting and Malfunctioning", Philosophical Studies, 79, 109-41.

Neander, K. (1996) "Dretske's Innate Modesty", Australasian Journal of Philosophy, 74, 2, 258-74.

Papineau, D. (1987) Reality and Representation, Oxford: Blackwell.

Papineau, D. (1993) Philosophical Naturalism, Oxford: Blackwell.

Pietroski, P.M. (1992) "Intentionality and Teleological Error", Pacific Philosophical Quarterly, 73, 267-82.

Rowlands, M. (1997) "Teleological Semantics", Mind, 106, 422, 279-303.

Salmon, W. (1984) Scientific Explanation and the Causal Structure of the World, Princeton: Princeton Unversity Press.

Searle, J.R. (1992) The Rediscovery of the Mind, Cambridge, Mass., MIT Press.

Sober, E. (1984) The Nature of Selection, Cambridge, Mass.: MIT Press.

Sperber, D. (1994) "The Modularity of Thought and the Epidemiology of Representations", in L.A. Hirschfeld & S.A. Gelman (eds.)(1994) Mapping the Mind, Domain Specificy in Cognition and Culture, Cambridge: Cambridge University Press.

Wright, L. (1973) "Function", in E. Sober (ed.) (1984) Conceptual Issues in Evolutionary Biology, Cambridge, Mass.: MIT Press.

[1] I know this claim is controversial and has been forcefully questioned by e.g., McDowell (1994). I guess Dennett's view that there is no real difference between perceptual experience and judgment is another way of rejecting the distinction. However, I side with Dretske (1981), Evans (1982) and others on this issue.

[2] Presumably Dennett does not. In recent discussion, Millikan has entertained the possibility that the content of such simple systems as Dretske's magnetosomes be indeterminate. This is not the view she has defended in her writings. See e.g., Millikan (1993).

[3] Discussion of the contents of simple sensory representational systems - which is the topic of this paper - must of course face a prior question: What justifies the ascription of content (or meaning) to such elementary systems in the first place? Only the explanatory pay-off in the explanation of behavior can, I think, justify the ascription of content. Intuitions cannot. I will not discuss this question here. I will merely note that one influential source of resistance to ascribing contents to simple sensory mechanisms is based on the thought that a creature who lacks the concept of representation cannot be ascribed representations with contents. I don't find the thought persuasive.

[5] Some philosophers have argued that ascriptions of functions are not factual, e.g. Dennett (1987) and Searle (1992) who agree on little else. I have tried to respond to their claims in Jacob (1995) and (1997, ch. 4).

[6] As argued by Millikan (1993), Neander (1991, 1995) and Godfrey-Smith (1993), there is room for two notions of function: the teleological notion and Cummins' (1975, 1983) non-teleological notion (which I call "analytic" in Jacob, 1997, ch. 4).

[7] As emphasized by Millikan (1993) and Neander (1995), many devices will have several "serial" or nested functions connected by a "by" relation.

[8] As Sober (1984) famously noted, the intensionality of being selected for is higher than that of being merely selected (e.g., as a by-product).

[9] What Millikan calls Normal conditions mentioned in a Normal explanation of an item's proper function is what within an informational framework Dretske (1981) calls "channel conditions". I won't discuss the question whether the notion of a Normal explanation is acceptable as a primitive notion within a naturalistic framework. I merely note that there is a non-epistemic notion of explanation available to a naturalist here. See e.g., Salmon (1984). I will, however, come back to the thorny problem of spelling out the Normal conditions involved in a Normal explanation.

[10] As will emerge shortly, the consumer system is a motor system. We can think of the contrast between the producer mechanism and the consumer mechanism as the collaboration between sensory neurons and motor neurons in an animal.

[11] Similarly, it was not just assaulting any female hoverfly which helps explain the proliferation of hoverflies. The females had to be healthy and fertile.

[13] Notice that the NCP only states a necessary, not a sufficient, condition on the content of a sensory mechanism. Given that I subscribe to an informationally-based teleosemantic framework, nomic covariation cannot be a sufficient condition on content: to have content, a device must have a function.

[14] There is an interesting analogy here between Fodor's (1994, 1998) atomistic purely informational view of content and Millikan's consumer-based teleosemantic view: both are strongly anti-verificationist. Millikan rejects the appeal to discriminatory abilities in determining the content of a sensory mechanism. Fodor rejects an inferential role theory of the content of concepts. In my opinion, calling a view "verificationist" does not ipso facto show it's wrong.

[15] It doesn't, I think, alter Rowlands' point if we recognize that the distinction between a device's algorthmic vs. organismic proper function might not pick a unique pair of functions.

[16] Similarly FLY can't be the content of the frog's visual system.

[17] When I say improperly that M-states have been recruited by evolution as causes of output m* of S's motor system M*, I mean of course to say that mechanism M has been recruited by evolution to deliver states which will be causes of m*. Selection does not act on particular states of a cognitive mechanism. It acts on the mechanism itself. Evolution confers direct proper functions on mechanisms and states derive their (derived) function from the proper function of the mechanism which produces them.

[18] So on this view, the device in the rattelsnake which fires when and only when its warmth detector and its movement detector are stimulated represents warmth and movement, not mice. Sperber's (1994) orgs which have two detectors, one for vibrations in the ground and one for noise represent vibrations and noise. They don't represent either elephants or hippos. Even though what explains the proliferation of orgs is that representing vibrations and noise allowed them to escape elephants at one time and hippos later on.