One of the virtues of Fred Dretske's recent work has been the salutary openness with which he has described the motivations he discovers controlling his thought, and this candor has brought a submerged confusion close to the surface. Since this confusion is widely shared by philosophers and others working on the problem of content ascription, an analysis of its influence on Dretske will at the same time illuminate the difficulties it is creating for other writers.
I think the confusion is born of the misalliance of two background images we have, each of which is valuable in itself but liable to be subscribed to in exaggerated form.Endnote 2
First, there is the image of mining the past for the future. The purpose of brains is to "produce future", as the poet Valéry said: to create anticipations and expectations that will guide an organism's actions along the paths that avoid harm and capture good.Endnote 3 The raw materials for the production of reliable anticipations must come in via our sense organs, to be stored in our memories. The plasticity of our brains permits us to learn. As this image would have it, learning is a matter of extracting meaning from our interactions for use in the future. There is really no doubt that this is what our nervous systems are for, but how exactly this process of information-extraction and subsequent meaning-manufacture is accomplished remains a mystery. How, we ask, can we get the fruits of our transactions with the past to inform and guide our future acts? "Theories of meaning" or "theories of content ascription" or "psychosemantics" are all attempts either to answer this question or at least to erect and justify families of constraints on acceptable answers.
Second, there is the image of what it feels like to come to understand something: there you are, encountering something somewhat perplexing or indecipherable or at least as yet unknown--something that in one way or another creates the epistemic itch, when finally Aha! I've got it! Understanding dawns, and the item is transformed; it becomes useful, comprehended, within your control. Before time t the thing was not understood; after time t, it was understood--a clearly marked state transition that can often be accurately timed, even though it is, emphatically, a subjectively accessible, introspectively discovered transition.
Put these two good images together, and they tend to spawn the idea of do-it-yourself understanding, an alchemical process in which the individual, the agent (the inner I or inner eye or homunculus or self) transforms mere information into understood content.
Let me elaborate just a bit on these images, and the way they infuse our thinking. As we encounter things, our senses take in huge amounts of information, which, like ore, needs to be refined and converted into usable form; it is no use having the information that p, even if that is exactly the information you need, unless you have it in a form you can understand. So there must be a process (or perhaps a variety of processes) for turning raw or crude information into the useful materials from which "we" can construct our beliefs and plans. This refining/transforming process is largely if not entirely unconscious, and is often called interpretation. Its product is traditionally called understanding. According to this view, one moves from mere receipt of information via some sort of interpretive process to the desirable state of understanding that information. Once understanding has been achieved, the mind has something to work with, something in position to guide action, inform plans, persuade, warn, remind, illuminate.
There is something privileged--or perhaps proprietary would be a better term--about the state of understanding. The state of understanding doesn't do me any good unless it is my understanding. Just as you can't take my naps or eat my lunches, you can't understand my thoughts and perceptions. More precisely, you can, perhaps, eat my lunch or understand my thoughts, but this will do me no good. My goal is not just getting my lunch eaten, but getting my lunch eaten by me. Similarly, unless I am the one who understands the import of my sensory states, their information is unavailable to serve me. For instance, suppose you can tell, by observing a certain flush of my skin, that there is a dangerous amount of carbon monoxide in the vicinity; a state of mine contains information, and you extract that valuable information from my state and put it to use in guiding your own action; if only I could also extract that information, and get it into usable form, I could perhaps save my life!
Such is the way we are all apt to think. There is surely a lot that is right about this way of thinking, but it also engenders some extraordinarily persistent illusions and false hopes. In particular, it engenders the hopeless and ultimately ill-motivated quest for do-it-yourself understanding, an infatuation we can best unmask by examining some of its kin.
Consider the earnest homeowner and handyman who shuns plumbers and electricians and wants to do it all himself. Moreover, when it comes to doing it himself, no prefabricated units or kits (with "some assembly required") find favor with him. He insists on building everything from scratch; he cuts down the trees, saws out the boards, mines the ore from which to smelt the iron from which to make the steel from which to draw the wire from which to make the nails. This is a fanatic do-it-yourselfer. And also, of course, a fool. Why not take advantage of the huge efforts of partial design and partial construction whose products are so readily available, we might well ask. Do you really insist on fashioning your own lightbulbs?
Consider now a rather opposite sort of fool, the person who claims to be an artist, but whose oeuvres are "paint by numbers" paintings, and sculptures made by pouring plaster of paris into purchased rubber molds. This man boasts of building his own bicycle; it turns out that when it arrived in its crate, the wheels had to be put on. He "designed and built his own house" he says; what he means is that when he ordered his tract house, he specified color combination D (from the available options A-G), and put in all the lightbulbs himself.
Somewhere in between these extremes lies both sanity and the responsibility of authorship (if that really matters to you). That is, you can quite properly claim--if you feel the need so to claim--that you are the sole creator of the coffee table, even if you did purchase the woodscrews ready-made, and the oak boards came already planed to their proper thickness.
We can now anticipate the moral for our story about understanding: there can be no central workshop of the mind/brain in which the agent "all by himself" achieves understanding by fashioning raw materials (materials that are not already pre-understood to some degree by some agency other than himself) into genuinely meaningful or contentful objects or structures. The processes responsible for understanding must be distributed, in both time and space, so that we make a mistake if we pose the following tempting questions: where does the understanding happen? Has he understood it yet? (Has he understood it yet? Has he understood it yet?) The introspectively familiar phenomenon of coming to understand something does happen--and people do make things for themselves on occasion--but these phenomena must be viewed as only relatively salient turning points in processes in which responsibility or authorship or agency is also diffused.Endnote 4
Now we are ready to look at the quest for do-it-yourself understanding in Fred Dretske's recent work. In a recent paper (1985, p.31), he gives a particularly clear and telling example of the elusive difference that he is trying to capture: the marijuana-detecting dog whose tail wags because an event occurs in its brain "meaning" (carrying the information) that marijuana is present, but which doesn't wag because the event means what it means. Unlike the dog's tell-tale tail-wag, Fred insists, our bodily actions often happen because of what the states that cause them mean: "it is the stucture's having this meaning (its semantics, not just the structure that has this meaning (the syntax), which is relevant to explaining behavior." (personal correspondence, quoted in Dennett, 1987, p.307).
In a more recent paper, "The Causal Role of Content,"Endnote 5 he puts it even better: he has been "increasingly preoccupied with the question, not of what meaning (or content) is, but what meaning (or content) does." He wants to give meaning "some explanatory bite" (ms. p.5), and ideally, this would involve showing "the way meaning . . . can figure in the explanation of why" an event causes what it does.
What difficulty lies in the way of this goal? The arguments, from myself and others, to the effect that the meaning of a particular event always must be, in Dretske's terms, epiphenomenal with regard to that event's actual causal powers. In Content and Consciousness (1969) I argued that the intentional interpretation of neural events is always at best a "heuristic overlay", and in "Intentional Systems," (1971) I tried to show how physical-stance predictions of the effects of structures always had hegemony over predictions based on the powers attributable to those structures (in idealization) in virtue of their meaning--as discerned from the intentional stance. More recently (1981, 1983, 1987), I have spoken, as Dretske notes, of the impotence of meaning; the brain is first and foremost a syntactic engine, which can be fruitfully viewed as reliably mimicking a semantic engine, but in which meanings themselves never overrule, overpower, or so much as influence the brute mechanistic or syntactic flow of local causation in the nervous system. (A semantic engine, I claim, is a mechanistic impossibility--like a perpetual motion machine, but a useful idealization in setting the specs for actual mechanisms.) Others have made similar claims: Fodor has long insisted on the inescapability of what he calls the Formality Constraint, and Dretske attributes to Schiffer (1987) the view that meaning is an "excrescence" which can do nothing and explain nothing.
Dretske's response to these claims is ambivalent. On the one hand he makes it clear that he would truly love to defend a doctrine of brains as real semantic engines, with real meanings locally throwing their weight around and making the most direct imaginable difference--but he knows better. He is convinced by the arguments that show, in LePore and Loewer's (198x) terms, that the historical facts on which the meaning of a structure supervenes are screened off from the explanation of the structure's causal powers and behavior. (In a similar vein, I have spoken of "inert historical facts.") Dretske's own presentation of the arguments for this conclusion are the clearest yet, and they draw, tellingly, on an analogy to the difference between value and perceived value, genuine currency versus counterfeit currency. In terms of our first image, Dretske acknowledges that there can be no guarantee that the precious ore of information actually mined from the past will be transformed into the appropriate future-guiding meaning-artifact (or "structure"). Conversely (and strangely, more importantly in the eyes of some), there is no guarantee that the "right" future-guiding structure was made out of the right stuff--that it was made from ore that was properly mined from the past.
Two kinds of all too familiar puzzle cases exhibit these dreaded dissociations. In one type, Tom has the information that the person standing before him is the murderer (let us say), but hasn't twigged. In some formulations, he believes of this individual that he is the murderer, but does not recognize or realize that he has this belief. That is, thanks to the proper sort of historical ancestry of his current state, the precious ore of information is in his possession, but for one reason or another it has not been transformed into a structure that he can put to the appropriate use on this occasion. The relevant understanding has not happened; or more pointedly, he doesn't understand the information he has. In the other, more science-fictional type, Tom (or Twin-Tom or Cosmic-Coincidence-simulacrum-of-Tom) is in just the right state so far as the future is concerned--given the structures present in his workshop he is well-poised to deal with the murderer or whatever--but these structures have bogus credentials. It is only as if he believed of the murderer that he was the murderer (and recognized this); it is only as if his counterfeit understanding were genuine. There may or may not have been understanding involved in the process that set up poor Tom in this way, but it was not his understanding. There are thus two ways such a structure can be disqualified according to this view: it isn't made of the right materials at all--cosmic coincidence cases in which the relevant information never was brought into the workshop--or it wasn't made by the agent. The policemen who trained Dretske's marijuana-sniffing dog to wag its tail when a certain aroma was present understood what they were designing, but since the dog itself is none the wiser, this doesn't count.
So Dretske wants to develop an account of content ascription that will ensure that only genuine, non-counterfeit contents get ascribed, and that moreover has the consequence that these contents "make a difference" in the only way that contents as contents could make a difference: by being understood. He wants meaning to have some explanatory bite, but since he endorses, and indeed strengthens, the arguments to the effect that the meaning of a particular event cannot directly and ipso facto account for the effects that flow from that event, he has a problem. The well-known responses to these arguments all involve finding some more indirect way in which there can be a regular, reliable correspondence between the meaning of a (neural) structure and its effects on behavior. For instance, Fodor's language of thought hypothesis is essentially an attempt to describe a system that satisfies what Haugeland has called the Formalists' Motto:
If you take care of the syntax, the semantics will take care of itself.(Artificial Intelligence: the Very Idea, 1985, p.106.)
This motto lies at the heart of what Haugeland calls GOFAI (Good Old Fashioned AI) and I call "High Church Computationalism" (Dennett, 1986). There are other, more noncommittal, ways in which materialists can postulate a reliable correspondence between semantics and mechanism, but Dretske calls all these theories "pre-established harmony" theories, and finds them all intolerable. He still yearns to establish a more direct and satisfying relation between meaning and mechanism than the rest of us have described.
I think it is wonderful that he chooses to illuminate this "corner materialists have painted themselves into" by some old-fashioned labels: "epiphenomenalism" and "pre-established harmony." These dreaded relics of prescientific philosophy of mind, these desperate and doomed escape routes from Cartesian dualism, are sure to strike terror in the heart of complacent materialists, and Dretske tells us exactly why he cannot abide them: any view of meaning according to which there is merely a pre-established harmony between the causal facts and the facts of meaning may permit us to predict an agent's behavior, and control an agent's behavior, but will not permit us to explain an agent's behavior--"and that," he says quite reasonably, "is what it takes to vindicate belief-desire psychology or our ordinary view about the causal efficacy of thought--that we stopped, for example, because we thought the light was red."Endnote 6
Our first image, of mining the past for the future, identifies two components or ingredients in meaning, reminiscent of Aristotle's matter and form: the informational ore is the matter or raw material, while the finished product has a behavior-directing competence or power which is directly a function of its structure or form. There is no meaning (in the relevant sense) without both contributions--something Dretske and I have both always insisted upon--but there is a temptation to think of meaning as somehow residing in the informational component, as potential meaning, perhaps, that just needs to be identified and refined in the course of manufacture. But since Dretske has reluctantly accepted the conclusion that there can be no direct mechanical extraction of meaning, he realizes that a straightforward vindication of this intuition is not in the cards. What he offers us instead is an attempt to salvage, if not locally potent meanings, then the next best thing: "the fact that A means M, though it fails to explain why B occurred, may help explain a closely related fact, the fact that events of type A, when they occur, cause events of type B . . . And this fact, especially when we are trying to explain the behavior of a system, is a fact eminently worth explaining." (ms. p.11)
What we need, in short, is not just a brute pre-established harmony, but an explanation of why and how the harmony is pre-established. Moreover (if Dretske has his druthers) this explanation will make an ineliminable appeal to the meanings of the elements thus linked. Now apparently he thinks that we pre-established harmony theorists have failed to offer such an explanation, for he is forthright in his opposition: "I don't think this works. Or, if it does work, it does so at a cost that I'm not prepared (unless forced) to pay." (ms. p.9) This is a mistake, for in the end, he does not offer us an alternative to pre-established harmony, but a version of it, a truncated version, in fact, of the version I have offered. I am pleased to see the convergence of our positions, which are now really very close. The main point of disagreement, as we shall see, stems from Dretske's quixotic quest for do-it-yourself-understanding.
There are exactly five ways in which such a correspondence--a "pre-established harmony" between the meanings of structures and their causal powers--could (in principle) come into existence. Dretske encounters them all, but fails to recognize them for what they are.
First, there are the Three Cousins:
Then there is the Philosopher's Fantasy:
4. the correspondence is the result of a Cosmic Coincidence.
Finally, there is the Theologian's Hope:
5. the correspondence is created and maintained by God.
Eager though he is to wed meaning and causation together, Dretske rightly dismisses this fifth possibility with only a passing allusion, for the obvious reason that it would be, quite literally, a deus ex machina. It is interesting to me that philosophers who would be embarrassed to spend more than a passing moment dismissing this fifth alternative are nevertheless irresistibly drawn to extended discussions of the implications of the fourth, Cosmic Coincidence, which is actually a more fantastic and negligible "possibility in principle."
Notice that there really cannot be a sixth route to pre-established harmony. If such a harmony is not just a single, large ("Cosmic") coincidence (4), or a miracle (5), it must be the product, somehow, of lots of tiny, well-exploited coincidences. This is because, as we have already agreed, meanings cannot directly cause things to happen, so they cannot directly cause themselves to correspond to any causal regularities in the world. So it will have to be via an indirect process of fortuitous coincidences that are duly "appreciated" or "recognized" or "valued" or "selected" by something--either something blind and mechanical, such as natural selection or operant conditioning or "neural Darwinism" (Edelman, 1988)(1 and 2), or something foresightful and intelligent, such as an engineer (3). Any such process is a design process, and must consist, at bottom, of such generate-and-test cycles, where the generation of diversity to be tested is somewhat random or coincidental (see "Why the Law of Effect Will Not Go Away", Dennett, 1974, 1978).
Dretske ends up Endnote 7 endorsing only the second way--learning--as a way of giving meanings explanatory bite, but the argument by which he proceeds would seem to legitimate all Three Cousins for the same reasons. It is in his unwillingness to accept this implication of his own argument that we will find the traces of his misbegotten desire for do-it-yourself understanding. He begins his attempt to give meaning explanatory bite by distinguishing between explaining B's happening (A caused it) and explaining A's causing B. He then tells a story about how a neural assembly can come to have a meaning, via species evolution by natural selection. The upshot is that such a story can explain why it is that A's, meaning what they do, cause B's. So far as I can see, this account follows just the path I laid out in Content and Consciousness, in my discussion of "the evolution of appropriate structures" (pp.47-63). Dretske describes an organism with a need to develop an avoidance mechanism against the highly toxic condition F, while I described different strains of organism which were differently wired up to avoid, approach, or ignore a particular stimulus condition that happens to be "more often than not"--an important proviso--injurious (p.49).The result of both thought experiments is the same: the lucky ones who happen to be wired up to avoid the toxic, ceteris paribus, are the ones who survive to reproduce, their coincidentally happy correspondence being selected for by natural selection.
Dretske goes on to suggest that the same result could also be achieved by redesign during an individual organism's lifetime via a conditioning process, or as I put it, via a process of "intra-cerebral evolution." (Content and Consciousness, pp. 56-63). In his Tepoztlan paper, he does not suggest that there is any important difference between the processes of species evolution and intra-cerebral evolution, but in his postscript, he insists that there is:
Natural selection gives us something quite different: reflex, instinct, tropisms, fixed-action-patterns, and other forms of involuntary behavior--behavior that is (typically) not explained in terms of the actors' beliefs and desires (if any). These genetically determined patterns of behavior often involve (as triggers for response) internal indictators (information-carying elements), but, unlike, belief, it isn't their content that explains the way they affect output. That is determined by the genes.
What exactly is the contrast?
In order to get meaning itself (and not just the structures that have meaning) to play an important role in the explanation of an individual's behavior (as beliefs and desires do) one has to look at the meaning that was instrumental in shaping the behavior that is being explained. This occurs only during individual learning. Only then is the meaning of the structure type (the fact that it indicates so-and-so about the animal's surroundings) responsible for its recruitment as a control element in the production of appropriate action.
The only difference I can discern, however, is that the "structure type" in the case of natural selection is a type that is identified in the genotype, while the structure type in the case of intra-cerebral evolution is a type that is identified only in the phenotype. In both cases "meaning was instrumental in shaping the behavior"--that is, in shaping the behavior-type, and in neither case was meaning instrumental in shaping any particular, individual token of a behavioral type.
Note, in any case, that while Dretske now endorses intra-cerebral evolution or conditioning as a process that yields meanings with explanatory bite, his doing so requires that he abandon, or at least soften, the hard line he has previously taken on these issues, in Knowledge and the Flow of Information (1981), and more recently in "Machines and the Mental" (1985) and "Misperception" (1986). In his book, he attempted to erect meaning on a foundation of information. That is, he developed an account of semantic information (the sort of meaning needed for psychology--a functional notion that applies only to designed channels) from a base of non-functional, non-semantic information channels, through which traveled items with "natural meaning"--items that informed (by definition) with perfect reliability; where natural meanings are concerned, no misrepresentation is possible. The task then was to turn the corner, somehow, from natural meaning to semantic or "natural functional meaning" (in which misrepresentation is possible), without, as he has more recently put it, "artificially inflating" the attributions of meaning to a structures one is so interpreting. Fred has tried to hold the line against inflation by insisting on what he now calls the "indicator relation." But the indicator relations he now endorses can only approximately carry information about the distal conditions one is tempted to say they are designed to inform the organism about.
The indicator relation, which he heralds as "a plausible, at least a possible, partial basis for meaning" need only be a rough-and-ready guide to the meaning of the chosen structure. At least Dretske ought to recognize this, for that is how evolution, at the species or neural level, works. Mother Nature is a stingy, opportunistic engineer who takes advantage of rough correspondences whenever they are good enough for the organism's purposes, given its budget. He correctly wants to boast that his account does find an explanatory role for meanings: it is because of what a structure happens to indicate that it is "selected for" or "reinforced in" a further causal role in the economy of control of the organism, but it will be selected for, or reinforced in, this role whenever it is "close enough for government work" as the engineers say.
Dretske shies away from embracing the First Cousin, then, in spite of his dalliance with it, but in any case the most he could get out of Cousins (1) and (2) would be an explanatory account that makes ineliminable appeal to quasi-indicator relations or approximate meanings. Mother Nature never holds out for high-priced indicator relations.
As my frequent allusions to engineering suggest, and as I have argued in The Intentional Stance (Ch.8), the design processes one encounters in (1) and (2) are not only fundamentally the same, but fundamentally the same as the artificial design process encountered in (3). It is striking that Dretske resists this conclusion, most emphatically in "Machines and the Mental" but also, with somewhat difference emphasis, in "The Causal Role of Content." If it is an engineer or computer scientist--rather than a learning history--who does the selecting (who esteeems a structure for its quasi-indicator relations and harnesses it in a particular functional role in a robot's control structure), this somehow gives the structure an illegitimate ancestry, in Dretske's eyes. But why?
Why should selection by an engineer disqualify a structure for one of Dretske's why-explanations? One might suppose that engineer-selection had an advantage over natural selection in this regard. One might say: whereas one must use scare-quotes when talking of natural selection's "appreciation" of the potential meaning of a structure, engineers sometimes really and truly respond to these potential meanings in the course of their conscious, deliberate, designing. But Dretske does not find this line of thought palatable, in spite of the fact that he often illustrates his point by the example of the wiring of a light switch, an example which explicitly appeals to just such different appreciations or intentions on the part of circuit designers. Is it just because he has a bad case of organophilia (otherwise known as silicophobia)? I think not. I suspect that he is distracted by an illusion: the illusion that somehow in ways (1) and (2)--but not in (3)--the organism itself (or even: its mind or soul) does the understanding--responds directly to the meaning.
After all, one might naturally but confusedly say, if some engineer is responsible for appreciating the meaning of a neural or computer structure, and is responsible for designing its further role in the light of that meaning, there is nothing left for the organism or robot to do--no task of understanding left over. The very same consideration would disqualify natural selection, for if it is the process of natural selection that has set up an innate correspondence between meaning and causation, there would likewise be no supervisory role for the organism (or its central, understanding homunculus) to play. (For more on this see my commentary of Ewert's "Prey-Catching in Toads" in BBS, 1988, "Eliminate the Middletoad!"). And this is, I suspect, exactly the natural confusion into which Dretske has fallen.
Both in his book, and more recently in his articles (and in personal correspondence, quoted in The Intentional Stance, p. 306), he has held the line against innate, unlearned meanings. "Beliefs and desires, reasons in general (the sort of thing covered by the intentional stance), are (or so I would like to argue) invoked to explain patterns of behavior that are acquired during the life history of the organism exhibiting the behavior (i.e., learned)."
Why has he been tempted to favor individual learning histories exclusively? Because, apparently, they seemed to give the organism itself a proper role in the acquisition and appreciation of the meanings in question. But once again, of course, if "mere conditioning" is responsible for the redesign of the individual organism's brain, this too looks like taking responsibility away from the inner understander that Dretske is so loath to lose. Consider, in this connection, the marijuana-sniffing dog. It was carefully conditioned, presumably, to wag its tail (the conditioned response) when stimulated by the odor of marijuana. If we attend to the process by which the harmony was achieved, we can no longer see any role for "the dog itself" to play. The theoretical beauty, after all, of theories of learning that broke learning down, in one way or another, into "mere" conditioning was that they promised to explain learning without improper recourse to inner, mysterious processes or states of understanding.
"But where does the understanding happen?" one might ask, having been treated to an account of the way in which an organism was caused, thanks to its design, to respond appropriately (first internally, and eventually externally) to events impinging on it. The temptation here is to insist that in the case of the dog, the understanding is all in the trainers, and that no amount of "mere" conditioning or other design-by-outsiders could ever succeed in getting the dog to achieve its own understanding.
Here we see the unfortunate influence of our mis-united images, which tend to create the following barely submerged conviction: Understanding isn't real understanding, understanding for which the the agent or organism is responsible, of which it is the proper author, unless it is entirely the product of the agent's own, do-it-yourself processes of interpretation. If the agent is "merely" the beneficiary of partially pre-understood, pre-interpreted, designed-by-other-agencies structures, then even if it appears as if the agent was understanding, there is really no agency there at all! Other, earlier creators are the proper authors of this design.
The appeal of the proprietary (it doesn't count unless it is my understanding!) turns Dretske's head against both natural selection and artificial intelligence. But its dire influence can also be seen, I believe, in Fodor's ever more mystical pronouncements about the psychosemantical relation, and the non-modularity of "belief-fixation"). In The Modularity of Mind (1983) Fodor argued heatedly for a vision of the mind in which the peripheral modules, both afferent and efferent, were blind, mechanical, encapsulated systems surrounding a central core, the arena of non-modular, global "belief fixation," a workshop in which the meanings of the products of the modules could be (somehow) interpreted by the mysterious central processes, of which Fodor emphatically insisted that neither he nor anyone else had any theories.
A lot is known about the transformations of representations which serve to get information into a form appropriate for central processing; practically nothing is known about what happens after the information gets there. The ghost has been chased further back into the machine, but it has not been exorcised. (p.127)
In effect Fodor is saying: "Here in the center is where understanding happens." Because of the Formality Constraint, Fodor is certain that the account to be given of the processes that achieve understanding will have to be, in the end, purely syntactically definable processes, and the only obvious candidate for such a Janus-faced process is inference, which can be viewed both semantically, as a meaning-extraction process, and syntactically, as a symbol-manipulation process. But aside from this dual-aspect appeal of inference as a potential gap-filler, there is little to recommend it. Such mechanistic accounts as there are of inference processes are hopelessly brittle and ineffective, as Fodor himself has insisted (in The Modularity of Mind). He has not yet offered his own model of how inferences are supposed to work to turn mere information into meaning, but in any event, for meaning-extraction to occur, there has to be meaning present (if hidden) in the products reaching the workshop via the senses--hence Fodor's attraction, in Psychosemantics and more recent work, to a purely denotational semantics that is somehow "information-based". Such a theory will ensure that the precious ore of information will make it into the home workshop, where the inner alchemist can turn it into useful products.
But this is an incoherent vision. No such account could work, in spite of its undeniable initial appeal. I have no formal proof of this, but an old thought experiment of mine shows why one might believe this:
Suppose you find yourself locked in a windowless room, with two walls covered with flashing lights, two walls covered with little buttons, and a note telling you that you are imprisoned in the control center of a giant robot on whose safety your own life now depends. Your task is simply to guide the robot through its somewhat perilous environment, learning to discriminate and cope with whatever comes along, finding "nourishment" and safe haven for the robot at night (so you can sleep) and avoiding dangers. All the information you need is conveyed by the flashing lights, and the robot's motor activity is controllable by pushing the buttons. To your dismay, however, you see that none of the lights or buttons are labeled. You can't tell whether the insistently flashing light in the upper left corner is warning danger, signaling a "full belly," informing you of the location of the sun, or requesting grease for a heel bearing. You don't know whether when you push a button and the light goes out, you've scratched an itch, occluded your view of something or destroyed an attacker. Clearly, if that is all you are given to go on, your task is impossible; if you succeeded in guiding your robot through the day it would be sheer luck. Yet in one sense (and a very familiar sense to cognitive psychologists) all the information you need is conveyed to you. For we needn't suppose the lights are mere repeaters of peripheral stimulation; their flashing can represent the products of perceptual analysis machinery as sophisticated as you wish, and similarly the output can be supposed to initiate devious actions guided by hierarchical sub-routine systems informed by multi-layered feedback. In short, the entire array of systems devised by the cognitive psychologists could be built into this robot, so that it conveyed to its control center highly mediated and refined information, and yet, though in one sense the information would be there, in another more important sense, it would not. Yet the task just described is in a sense just the brain's task; it has no windows out which it can look in order to correlate features of the world with its input. The problem of the control room could be solved for you, of course, if all the lights and buttons were correctly labeled (in a language you knew), but this can hardly be the brain's solution. The job of getting the input information interpreted correctly is thus not a matter of getting the information translated or transcribed into a particular internal code unless getting the information into that code is ipso facto getting it into functional position to govern the behavioral repertoire of the whole organism. (Dennett, 1978b, p.258)
There is no way that pure do-it-yourself interpretation could get started; the only hope of creating a system--a brain or a computer system--that can bootstrap its way to understandings it was not born with is to create it with a goodly stock of pre-established harmony built in. It is not clear that Dretske would diasagree with this conclusion, but then how are we to explain his earlier opposition to natural selection and his continuing opposition to engineering design as sources of Harmony? In recent correspondence with me, he has made it quite clear that he put so much stock in learning history that he was prepared to grant real meaning even to the structures in an artifact, so long as they were produced, in part, by individual learning by the artifact: "I think we could (logically) create an artifact that acquired original intentionality, but not one that (at the moment of creation as it were) had it." (personal correspondence, quoted in The Intentional Stance, p.305.)
Now here we must distinguish two claims, one plausible and important, and the other--Dretske's--obscurely motivated and, I think, strictly negligible. The plausible and important claim is that it is astronomically unfeasible to create, by the usual engineer's methods, the sorts of structures that are naturally and efficiently created by learning histories of the sort he champions. This suspicion strikes at the heart of a particular fantastic hope of some in AI, who would hand-craft the myriads of beliefs that would constitute the "world knowledge" of an adroit robot. Not only have some thought this was possible in principle; there is a multi-million dollar project in AI with that as its explicit goal: Douglas Lenat's CYC project, an effort which Lenat himself supposes will take person-centuries of programming to accomplish.(See my discussion in Daedalus, 1988). The majority opinion in AI, however, is that this is a hopeless approach, for reasons well-canvassed by David Waltz in "The Prospects for Building Truly Intelligent Machines," (same issue of Daedalus).
Dretske's point, in contrast, is philosophical, not practical. It is that even if the engineers could hand-craft all those structures, they wouldn't have any meaning until they had been somehow annealed in the fire of experience. He puts the astronomically unfeasible product of engineering design in the same category with the even more astronomically unlikely case of Cosmic Coincidence, in spite of the fact that in the former case, there would be explanations of the provenance of structures that made appeal to meanings. He imagines that a physical duplicate of himself might "materialize--miraculously or randomly--out of some stray collection of molecules" (ms p.19), and goes to some length to insist that this biological twin's motions would not be actions, with meanings behind them.
I move my arm in this way in order to frighten away a pesky fly. With such a purpose I am, let us say, shooing away a fly. That is my action. My biological twin, though he moves his arm in the same way (with the same result) does not shoo away a fly. He doesn't have wants or beliefs, the kind of purposes I have in moving my arm. He isn't, therefore, performing the same action. (p19)
Your intuitions may agree with his, or recoil, but in either case, they concern something negligible, for as he goes on to acknowledge, there is a loophole: this metaphysically shocking state of affairs is apt to be short-lived, since it will persist only "until the twin accumulates enough experience--until, that is, his internal processes acquire the requisite extrinsic relations--to give his control processes, the processes governing the movement of his hand, the same kind of explanation as mine." (p20).
How long, one wonders, should "acquiring the requisite extrinsic relations" take? I should think it would be instantaneous.Endnote 8 Signals from the bio-double's peripheral vision (or perhaps a faint blip from cutaneous sensors on the shoulder) happen to put the bio-double into what would be a bogus fly-out-there-sensing state--except that this time it is caused by a real fly. Moreover, the real fly's trajectory intimately determines the hand-eye coordination series that promptly leads to the (bogus or real?) "shooing" motions. How many flies must buzz around the head of a bio-double before he can start shooing them? If that isn't an angels-dancing-on-the-head-of-a-pin question, what would be?
This curious question, of how much traffic with the world is enough, somehow, to ensure that genuine meaning has been established, is simply the enlargement (via a Cosmic Coincidence) of the curious question that has bedeviled some evolutionary theorists: how much selection is required to endorse a tiny coincidence (a random mutation) as a genuine adaptation? (See my discussion in The Intentional Stance, pp.320-1 and footnotes on those pages, and in my response to commentary in BBS). But if nothing but arbitrary answers (e.g. 42 generations of selection) could "settle" the question for natural selection, only arbitrary answers (e.g., 42 flies must buzz) could settle the question for a learning history, for the processes have the same structure--they must begin with a fortuitious or coincidental coupling, thereupon favored--and they have the same power to design structures in indirect response to meaning.
Notice that these frivolous questions are analogous to the question of just how much do-it-yourself an agent must contribute before he is the real creator of something. The beneficiary of Cosmic Coincidence (or of interfering engineers) is like the sculptor who opportunistically avails himself of objets trouvés, taking advantage of unearned (unlearned) gifts. (As several punsters have noted, Dretske's view of content is that it must be acquired the old-fashioned way: you have to learn it.)
But where does this beneficiary keep the understanding required to appreciate these unearned gifts of meaning in the first place? The answer must be: there is no such original, central, untainted understanding; there are only the varieties of reliably meaning-tracking mechanical processes that, when reflected on themselves and their kin, tend to enhance the power of the indirect processes of meaning-extraction.
In The Intentional Stance, I looked at this issue from a slightly different perspective, using an example that might repay a second look in the present context.
The first thing a baby cuckoo does when it hatches is to look around the next for other eggs, its potential competitors for its adoptive partents' attention, and attempt to roll them over the edge. It surely has no inkling of the functional meaning of its activity, but that meaning is nevertheless there--for the organism and to the organism--unless we suppose by the latter phrase that the organims has to 'have access' to that meaning, has to be in a position to reflect on it, or avow it, for instance. The rationale of the cuckoo's chillingly purposive activity is not in question; what remains to be investigated is to what extent the rationale is the fledgling's rationale and to what extend it is free-floating--merely what Mother Nature had in mind. . . . For Dretske, however, this is an all-or-nothing question, and it is tied to his intuition that there must be unique and unequivocal (natural functional) meanings for mental states. Dretske seesm to be trying to do two things at one stroke: first, he wants to draw a princip[led (and all-or-nothing) distinction beteen free-floating and--shall we say?--'fully appreciated' rationales; and second, he wants to remove all interpretive slack in the specification of the 'actual' or 'real' meaning of any such appreciated meaning-states. After all, if we appeal to our introspective intuitions, that is just how it seems: not only is there something we mean by our thoughtsd--something utterly determinate even if sometimes publicly ineffable--but it is our recognition or appreciation of that meaning that explains what we thereupon do. There certainly is a vast difference between the extremes represented by the fledgling cuckoo and, say, the cool-headed and cold-blooder human murderer who 'knows just what he is doing, and why,' but Dretske wants to turn it into the wrong sort of difference. (pp.306-7)
The fledgling cuckoo, like Dretske's marijuana-sniffing dog and Ewert's prey-catching toads, doesn't seem to be "doing any of the understanding," doesn't seem, as I noted above, to "have access" to the meaning of its own states. The tempting idea is that there is not just a vast difference, but one big difference, a difference in kind, not degree, between such cases and the cases in which (as we notice "from the inside") we understand the meanings of our perceptual states and act upon them because of that understanding. But if we look closely, we will see that between the humble cuckoos and toads, and us, is a continuum of cases. The injury-feigning plover understands more about its broken-wing dance than one might skeptically have supposed at first, as revealed by its sensitivity to relevant variations. The deceptive acts of some primates (Whiten and Byrne, 1988) apparently reveals a still more acute responsivity to the (approximate) meanings of the circumstances in which they act. And our own imperfect "grasp" of the implications of our perceptual states, beliefs and intentions is the grist for hundreds of philosophical examples.
The human capacity to treat one's own states, reflexively, as the objects of further consideration--perception and thought--greatly enhances our powers of understanding.Endnote 9 Arguably, the sort of swift, sensitive self-redesign it permits is the basis for our undeniable cognitive superiority over all other organisms. But being the product of such a process of auto-engineering is neither a necessary nor a sufficient condition for being a type of structure whose meaning "has explanatory bite." As in other forms of design and manufacture, whether something works well or ill depends only indirectly and imperfectly on whether you made it yourself.
Dretske goes searching for a closer tie between meaning and structure because he is worried by the following argument:
If meaning supervenes, at least in part, on the extrinsic properties of an event--historical and relational facts that need not be mirrored in the event's current (= the time at which it has its effects) physical constitution or structure--then if A causes B, then the fact, if it is a fact, that A means M will not--indeed, cannot--figure in a causal explanation of B. It cannot because, in similar circumstances [my italics], an event lacking this meaning, but otherwise the same, will have exactly the same effects. So it isn't A's having meaning M that explains why B occurred.(p.10)
This would be a significant worry if "similar circumstances" were apt to be forthcoming with non-zero probability in cases where A did not mean M, but we can rest assured that we will virtually never encounter such an anomaly. That is to say, it is no accident that events with the meanings they have get to play the causal roles they play (and Dretske in fact gives a good account of this), but the other side of that coin is that the odds are astronomical against the occurrence of an event or structure that lacked the relevant meaning somehow arising to cause a bogus B-type event--at least more than once or twice. It is not just that A's cause B's, but that if A's cease to mean what they do, they will (shortly) cease to cause B's, thanks to the power of an adaptive (or learning) system to select against those of its structures that prove inappropriate. That is as tight a relationship as one can hope for, and it is tight enough to explain the admirable if imperfect human capacity for being responsive to content.
What Dretske misses is that it isn't only that given the past, the existence of these event types (meaning what they do) cause those event types (meaning what they do), but in the future should the sort of dissociation occur that leads Dretske to worry about "epiphenomenalism", it will (soon, not immediately or perfectly "locally") cause the causal link between A's and B's to lapse or be revised because it no longer serves the right behavior-guiding role. This is true, however, in all three cousins: the only difference is time scale. Engineers, as soon as they notice a glitch, will repair or debug or discard the system; natural selection takes longer, but works inexorably. Learning in the individual has the same effect. It is the fastest portable system of self-repair, but of course what gets repaired is not the self, and not the individual causes and effects, but only the event-type pairings.
If an artifact-Dretske (or robot) is well-designed, then even before it has begun to interact with the world in a genuine causal way, it will be true that the desired counterfactaual is true: A's (meaning that p) cause B's (meaning that q), but if A's stopped meaning that p, or B's stopped meaning that q, the causal link would soon be broken. Almost as soon as a change (or more realistically, a lapse) in meaning arises, a revision of causation will arise in response, thanks to the self-correcting capacities built in. I suppose the philosophers' limiting case is the robot that is so well-designed to meet the future that it goes through its whole life without ever having to learn (repair) anything. The fact that it would have learned if it had had to learn is all that matters.
We don't need any special history of experience to ensure that any harmonies we encounter are robust, and Dretske's own arguments show that a learning history, the "acquisition of extrinsic relations", could not impart any intrinsic power to those harmonies. All there can be are better or worse harmonies, designed into the system thanks to various processes of trial-and-error research and development, some of them in the distant past of our species, some in our early learning histories, and some, no doubt, recent spur-of-the-moment revisions in the design of current structures--the sorts of sudden and introspectively memorable moments we take as our paradigms of coming to understand. We do get to do it ourselves, sometimes, but when we do, we are doing what we are designed to do, using parts and procedures that we have not ourselves designed, and taking full advantage of all the understanding that has gone into our design up to now.
Akins, K., 1988, "On Piranhas, Narcissism and Mental Representation," PhD dissertation, Univ. of Michigan.
Dennett, D. 1969, Content and Consciousness, Routledge & Kegan Paul
1971, "Intentional Systems," J.Phil.
1975, "Why the Law of Effect Will Not Go Away," Journal of the Theory of Social Behavior, 5, pp 169-87
1978. Brainstorms, MIT Press/A Bradford Book
1978b, "Current Issues in the Philosophy of Mind," American Phil. Quarterly, 1978, pp 249-59.
1981, "The Kinds of Intentional Psychology," in R. Healey,ed., Reduction, Time and Reality, Cambridge Univ. Press
1983, "Intentional System in Cognitive Ethology: the Panglossian Paradigm Defended," Behavioral and Brain Sciences, 6, pp.343-90
1986, "The Logical Geography of Computational Approaches: a View from the East Pole," in M. Harnish and M. Brand, eds., Representation of Knowledge and Belief, Univ of Arizona Press.
1987, The Intentional Stance, MIT Press/A Bradford Book
1988, "When Philosophers Encounter Artificial Intelligence," Daedalus, winter (issue on Artififical Intelligence), pp. 283-95.
1988, "Précis of The Intentional Stance" Behavioral and Brain Sciences, 11, pp. 493-544.
Dretske, F., 1981, Knowledge and the Flow of Information, MIT Press/A Bradford Book
1985, "Machines and the Mental," Proceedings and Addreses of the American Philosophical Association
1986, "Misrepresentation," in R. Bogdan, ed., Belief, Oxford Univ. Press
1988, "The Causal Role of Content," SOFIA conference, Tepoztlan, Mexico, August 8.
Fodor, J. 1983, The Modularity of Mind, MIT Press/A Bradford Book.
1987, Psychosemantics, MIT Press/A Bradford Book.
Edelman, G. M., 1987, Neural Darwinism: the Theory of Neuronal Group Selection, NY: Basic Books.
Haugeland, J.,1985, Artificial Intelligence: the Very Idea, MIT Press/A Bradford Book
Schiffer, S., 1987, Remnants of Meaning, MIT Press/A Bradford Book
van Gulick, 1988, "Consciousness, Intrinsic Intentionality, and Self-Understanding Machines," in A. J. Marcel and E.Bisiach, eds., Consciousness in Contemporary Science, OUP
Waltz, D., 1988, "The Prospects for Building Truly Intelligent Machines," Daedalus, (winter) issue on artificial intelligence, pp.191-212.
Whiten, A. and Byrne, R. W., eds., 1988, Machiavellian Intelligence, Oxford: Clarendon Press.
1.This paper incorporates portions of an earlier paper, "Ways of Establishing Harmony," in E. Villanueva, ed., 1990, Information, Semantics & Epistemology, Oxford; Blackwell, and forthcoming in B. McLaughlin, ed., [volume on Dretske], Oxford; Blackwell.
2.This confusion was suggested to me by Kathleen Akins, in "On Piranhas, Narcissism and Mental Representation," Univ. of Michigan Phd dissertation, 1988, where she draws a similar distinction, but uses the distinction for quite different philosophical purposes."
3. See the discussion of the verbs of "making a difference", such as "avoid", "prevent", "foster" etc., in my Elbow Room, 1984. See also my "Producing Future by Telling Stories," forthcoming in Kenneth M. Ford and Zenon Pylyshyn, eds., The Robot's Dilemma Revisited: the Frame Problem in Artificial Intelligence, Ablex.
4. A related point about the fallacy of "Cartesian materialism," the view that the brain has a privileged functional central point, is argued in Dennett and Kinsbourne, unpublished, "Temporal Anomalies of Consciousness."
5.Presented at the SOFIA conference on information-based semantics at Tepoztlan, Mexico, August, 1988. The present paper is an elaboration of my reply to Dretske at that conference, "Ways of Establishing Harmony." The proceedings of that conference are tentatively forthcoming in a volume from Blackwells [?] edited by Enrique Villaneuva.
6.It is interesting that here Dretske announces his allegiance to the same goal that motivates Fodor--the vindication of folk psychology taken neat--while dismissing Fodor's own solution as insufficiently responding to the folk psychological intuition that (the Formality Condition be damned) meanings make it happen.
7.Dretske's Tepoztlan paper appears to endorse natural selection as a legitimate way of establishing harmony--and I so interpreted his paper then--but in a note added later, he disavows that interpretation.
8.See, for instance, my discussion (in "Mechanism and Responsibility" 1973) of the parallel case of the bogus belief (that he has an older brother living in Cleveland) surgically inserted to Tom (in Brainstorms, pp. 251-3); the discussion of the Panamanian debut of the two-bitser (how long does it take for the new functions to be "real"?);and my reply to Goldman, in BBS, 1988, p.x.
9.In a valuable article that treats many of these issues, Robert van Gulick (1988) puts it this way: The personal-level experience of understanding is none the less not an illusion. I, the personal subject of experience, do u;nderstand. I can make all the necessary connections with experience, calling up representations to immediately connect one with another. The fact that my ability is the result of my being composed of an organized system of subpersonal components which produce my orderly flow of thoughts does not impugn my ability. What is illusory or mistaken is only the view that I am some distinct substantial self who produces these connectionsin virtue of a totally non-behavioural form of understanding. (p. 96)