Skinner and Dennett
Philosophy of Mind
Skinner's main target in Science and Human Behavior (New York: Free Press, 1953), and also elsewhere, e.g. in Beyond Freedom and Dignity (New York: Alfred A. Knopf, 1971), is what he calls "mentalism," namely the appeal to inner psychological phenomena in the explanation of human behavior. What is wrong with mental notions in the explanation of behavior? Many things, according to Skinner, among them these:
1. We cannot directly observe mental phenomena. As a result they are "inferential." On Skinner's view, this disqualifies them as scientific explanations of behavior. Skinner appeals to this point at several places in our reading. In a discussion of psychoanalytic theory, he writes: "any mental event which is unconscious is necessarily inferential, and the explanation [which makes use of it] is therefore not based upon independent observations of a valid cause" (30 ; bracketed page references are to the excerpt from Science and Human Behavior in Ned Block, ed., Readings in the Philosophy of Psychology, Cambridge: Harvard Univ. Press, 1980). Later he makes a similar criticism of commonsense psychological explanation, e.g. explaining why someone drinks water by saying he is thirsty. Skinner writes, of this "explanation," that: "if it means that he drinks because of a state of thirst, an inner causal event is invoked. If this state is purely inferential--if no dimensions are assigned to it which would make direct observation possible--it cannot serve as an explanation" (33 ). At one point, Skinner even seems to identify "inferential" with "fictional" (28 ).
Now, there is surely something valuable and important about this. Invoking phenomena which we cannot directly observe in explanations of things we can observe is always risky. I do not mean risky in the sense that we may turn out to be wrong; virtually any scientific claim is risky in that sense, including claims about things we can directly observe. The more serious risk is that we will make claims which are really not testable at all, which empirical evidence can never show to be mistaken because we can always fudge the theory a bit to explain why the evidence was to be expected after all. To use Karl Popper's term, the danger of "inferential" states is that theories making use of them may not be falsifiable. (Popper himself wrote of psychoanalysis: "those 'clinical observations' which analysts naively believe confirm their theory cannot do this any more than the daily confirmations which astrologers find in their practice." Popper, Conjectures and Refutations: The Growth of Scientific Knowledge (New York: Basic Books, 1962), pp. 37-38.) So we can read Skinner as making the important point that when we invoke theoretical entities or phenomena we need to do so in such a way that the theory making use of them makes predictions about observable phenomena which can be falsified.
But Skinner seems to take himself to have shown something much stronger than this, namely that a scientific theory should not make use of inferred entities or phenomena at all. And this seems much too strong a claim. If we restricted physics, or even archeology or paleontology, to making use only of things that can be directly observed, we would deprive ourselves of most of their most interesting results -- and also of a good deal of their predictive power. It often happens that the best theory which accounts for observed phenomena and makes predictions about unobserved but observable phenomena makes use of a good deal of theoretical apparatus for which our only evidence is inferential. An analogy may be helpful in seeing this point. Imagine typing things into the keyboard of a computer, observing the computer's responses, and trying to formulate hypotheses about how the machine will respond to various future stimuli. Conceivably we could do this without appealing to any hypotheses about how the machine is programmed, so that our theory simply took the form of correlations between inputs and outputs. But it seems quite clear that it will be far more useful to hypothesize about the machine's (internal, not directly observable) program, using hypotheses about the program together with information about inputs to formulate predictions about the machine's output. Now we may not be quite like computers, but presumably the principles which govern our behavior are at least as complex as those that govern a computer, so we may reasonably expect that formulating hypotheses about our own internal states and processes will turn out to be the most effective way of explaining and predicting our behavior. At the very least, it seems clear that it would be a mistake to rule out a priori any theory which made use of such hypotheses.
2. Mentalistic accounts are not genuinely explanatory. Skinner argues that many supposed explanations are really just made up on the spot and do not provide a genuine account of one's behavior. After giving a number of examples (e.g. one is confused because his mind is failing; one is disorganized because his ideas are confused), he writes: "in all this it is obvious that the mind and the ideas . . . are being invented on the spot to provide spurious explanations" (30 ).
Again, Skinner seems to be providing a useful warning: it would be a mistake to take such offhand remarks as having much explanatory power. (But, for a defense of the view that commonsense psychology does provide a fairly powerful explanatory account of a good deal of our behavior, see the writings of Jerry Fodor, e.g. his Psychosemantics (MIT Press, 1987), Chapter One.) On the other hand, most such remarks are not even supposed to be explanations of behavior; often they are just casual ways of describing it. The explanatory emptiness of much of our ordinary talk about mental events is not evidence that mentalistic notions can find no place in a genuinely scientific account of human behavior.
3. Mentalistic explanations are typically redundant. Skinner claims that mentalistic explanations really just restate the facts of behavior in more obscure language. He writes, for example, that: "A single set of facts is described by the two statements: 'He eats' and 'He is hungry.' . . . A single set of facts is described by the two statements: 'He plays well' and 'He has musical ability.'" (31 ). Here there seems to be at least a trace of the linguistic thesis of philosophical behaviorism as exemplified by Carnap and, at one time, Hempel. The idea seems to be that the mentalistic statements have the same meaning as the behaviors that count as evidence for them. But 'He eats' and 'He is hungry' don't mean quite the same thing (either could be true without the other being true), and in cases where mentalistic notions are doing more theoretical work it will be even clearer that there is no straightforward translation from mentalistic talk into behavioristic talk.
4. The "middle link" argument. Skinner suggests that, since the inner mental states which are supposed to explain behavior are themselves determined by external stimuli, they can safely be ignored: we can leave out the middleman and simply study the relations between stimuli and behavior. "Unless there is a weak spot in our causal chain so that the second link is not lawfully determined by the first, or the third by the second, then the first and third links must be lawfully related. If we must always go back beyond the second link for prediction and control, we may avoid many tiresome and exhausting digressions by examining the third link as a function of the first" (35 ).
At first sight, this looks very reasonable. If S determines M and M determines R, then S indirectly determines R: why not just consider the relationship between S and R, ignoring M? Dennett's computer analogy, which I mentioned above, is helpful here. It may be that the most effective way of explaining the relationship between S and R is by way of hypotheses about the nature of M. What is "hard-wired" in aside (this is comparable to human genetic makeup), how the machine is programmed is determined by inputs to the machine and, together with current inputs, determines the machine's output: but trying to predict output on the basis of input alone, without hypotheses about the machine's internal states and processes, is likely to be a disaster. It is worth mentioning Noam Chomsky's discussion of this point early in his review of Skinner's Verbal Behavior (in Language vol. 35 no. 1, 1959; reprinted in Ned Block, ed., Readings in the Philosophy of Psychology Volume 1 (Harvard, 1980)): "Anyone who sets himself the problem of analyzing the causation of behavior will . . . concern himself with . . . the record of inputs to the organism and the organism's present response, and will try to describe the function specifying the response in terms of the history of inputs. . . . The differences that arise between those who affirm and those who deny the importance of the specific 'contribution of the organism' to learning and performance concern the particular character and complexity of this function" (49).
5. Mentalistic explanations are homuncular. Skinner in a number of places objects to mentalistic explanations that they in effect invoke a little person or homunculus with all the same abilities that the ordinary person has. "The inner man is regarded as driving the body very much as the man at the steering wheel drives a car" (29 ). Explaining the behavior of a person by appealing to a little person inside the head, "driving" the body, clearly does not accomplish anything, since the actions of the homunculus are just as much in need of explanation as the actions of the person were originally. This is the criticism Dennett takes most seriously; Dennett's version is: "Since psychology's task is to account for the intelligence or rationality of men and animals, it cannot fulfill its task if anywhere along the line it presupposes intelligence or rationality" (Dennett, "Skinner Skinned," in Dennett, Brainstorms, Cambridge: MIT Press, 1978, p. 58).
All right. There's clearly something to this. But notice two things. (1) It clearly doesn't accomplish anything to "explain" someone's behavior by reference to a "homunculus" just as smart as the original person. But it doesn't follow that homunculi are useless. They may nevertheless accomplish something if they are dumber than the original person. We might be able to understand the capacities of a person in terms of the interactions of a number of agents each of which has simpler capacities than the original person; we might then explain each of these dumber agents in terms of a system of still dumber agents, and so on until at the very bottom level we have something so simple it can be understood in terms of neurons firing or something of the sort. This kind of explanation is familiar from computer science: a big complicated program may have a number of subroutines which can be thought of as agents dumber than the original program; these subroutines may themselves be decomposed into more basic routines, and so on, until at the bottom we reach circuits opening and closing. For the view that something like this is the best way to understand the human mind, see e.g. Marvin Minsky, The Society of Mind; see also some of Dennett's essays in Brainstorms, especially "Artificial Intelligence as Psychology and as Philosophy."
(2) The second thing to notice is that even if we must ultimately explain intelligence or rationality in terms that don't presuppose intelligence or rationality, it doesn't necessarily follow that intelligence and rationality should not be appealed to at all, or that they are ultimately unreal. Rather than showing that we aren't really rational, such an explanation might instead show what rationality consists in, might show what it is to be rational. This is Dennett's main point in "Skinner Skinned." Dennett argues that there is a crucial difference between explaining and explaining away (65). If our explanation of apparently rational behavior turns out to be extremely simple, we may want to say that the behavior was not really rational after all. But if the explanation is very complex and intricate, we may want to say not that the behavior is not rational, but that we now have a better understanding of what rationality consists in. (Compare: if we find out how a computer program solves problems in linear algebra, we don't say it's not really solving them, we just say we know how it does it. On the other hand, in cases like Weizenbaum's ELIZA program, the explanation of how the computer carries on a conversation is so simple that the right thing to say seems to be that the machine isn't really carrying on a conversation, it's just a trick.)