Behaviorism

Behaviorism developed as a reaction to dualism, and begins by denying everything that the dualist holds dear: There are no (non-physical) minds. But there are (thinking) people. So each person is the very same thing as his/her body, and exists only so long as the body continues to function. No live body, no person.

These are simply denials of dualist doctrine. They relieve the Behaviorist of the No Interaction problem, since there no souls to interact with. But these denials don't amount to a theory of mind. The guts of Behaviorism lies in its claim that certain patterns of bodily motion are definitive of mental states: Bodies can think - mental states and processes are (dispositions or tendencies to exhibit) patterns of stimuli and response.

The Problem of Other Minds vanishes since, according to Behaviorism, we can literally see others think. To see why the latter is so and to appreciate the positive portion of Behaviorism, let us look at the last part.

Behaviorism is a form of materialism about mental processes: every mental process is some physical process or other. Your mental processes are certain physical processes that take place in your body. But that's not all that Behaviorism says. It says something much more specific about how to say which physical processes are mental processes. And what more it says is at the heart of Behaviorism.

The key notion is that of a disposition, or dispositional property. When we say that sugar is soluble in water, we attribute to sugar a certain disposition, roughly:

(Of course, that's not chemically correct, but I hope you get the idea.) This form of expression is so common that it's been labeled a "subjunctive conditional" by linguists and grammarians. It's the ideal linguistic tool for expressing dispositional properties. The most important logical characteristic of such properties is that an object can have them even when they are not being shown off. If you keep your favorite sugar cube in a fluid tight safe until the world ends, it will still be true of it, at every moment of its existence, that it's soluble - because if it were in water, then it would dissolve; but you never give it the chance to realize its potential. In the same way, the fragility of some glassware is a disposition that we may hope is never exhibited:

But just because it's never dropped, the glass does not stop being fragile. Some dispositional properties are so complex that entire branches of engineering are devoted to their study; for example, there's airworthiness, which you insist that the plane be certified to have before you get on board. Aeronautical engineering is focused on that extremely complex characteristic, which we can describe very roughly and oversimply, with no offense intended to the engineers, as

How does this pertain to mental states and Behaviorism? Let's consider a simple example. Suppose that we decide to investigate the mental process we call "desire" or "wanting". To make the example more definite and to avoid the naughty bits, let's talk about chocolate desire. Our experimental subject is my associate, Mr. Waldo.

What would we do to find out if he wanted some chocolate? That's an easy one: we offer him some and see if he takes it. If he does, he's got the desire, if he doesn't he doesn't. What Behaviorism says is that we should take this sort of pattern of acceptance and refusal as our definition of chocolate desire. So, to a first rough approximation, we have:

Behaviorists like to describe this by saying that there is a certain pattern of stimuli and responses that defines chocolate desire. Here, offering Mr. Waldo the chocolate is the stimulus and his acceptance is the response. Notice that this pattern-description can accurately describe Mr. Waldo even at a time when he is not being offered any chocolate. He might have the tendency or disposition described by the definition, even though he's not been sparked into action by the appropriate stimulus. So even in a chocolate-free environment, this Behaviorist definition says that it is possible to desire chocolate.

Notice something important about this definition: it says nothing about what goes on inside Mr. Waldo; it stops, so to speak, at the skin. So far as the Behaviorist definition is concerned, it makes no difference how the pattern is produced; all that matters is the skin-level pattern of stimuli and response. Therefore, so far as Behaviorism is concerned, Mr. Waldo may as well have a black box inside his head, a box that can't be opened by anyone. We can treat Mr. Waldo as a device that gives a certain sort of output for a specified input, and we don't have to concern ourselves with the details of internal processing.

All of this generalizes, to give the general form of Behaviorism's definitions of mental states and processes:

In the example we are using now, we are considering just one input and one output. But in general, there will be many inputs and outputs that characterize the relevant pattern - the pattern will typically be quite complex. So we should talk about inputs I1, ... , In, and outputs O1, ... , On, and we'll also need to talk about a variety of circumstances and environments, as well. That's nice for psychologists, because it gives them something to do, namely, figure out the complex patterns. This was no small reason for Behaviorism's popularity among psychologists. It's reassuring to be guided by a doctrine that says you'll have a job for the foreseeable future. Figuring out all the relevant input-output statements would take a long time.

Behaviorism and the Problem of Pattern Poverty

Like many simple views, Behaviorism is guilty of over-simplifying in a rather serious way. To appreciate Behaviorism's main problem, let us take a look at the Behaviorist approach to defining another mental activity. It is a process that Artificial Intelligence (AI) research is very concerned with: conversation. There are many AI researchers who would love to know how to make a machine that would be able to understand English. I will describe a machine that would behave just as if it could understand English, but which is really unable to understand anything. It is a machine that exhibits the behavioral patterns of a conversing human being, but which undergoes no mental activity at all. If we focus on the kind of understanding involved in conversation, we can define it Behavioristically in this way:

Notice two features of this definition that it shares with our first example. First, it does not imply that P is able to understand verbal stimuli only if he is actually listening or talking. S/he can have the ability even when s/he is alone and silent. Second, this definition entails that exhibiting a certain behavioral pattern is sufficient for understanding. So anything that exhibits the pattern is therefore a thinking thing, no matter how the pattern of behavior is produced. Also note that there is no restriction on stimulus subject matter: we can talk with P about anything at all. Of course, we have to allow that among his/her sensible responses might be "I don't know" or "It sounds like nonsense to me", but then those are sensible responses in some contexts. It might even be OK for P to utter nonsense, if we asked him, "Can you make some meaningless sounds?"

Now I'll describe Ned Block's design for a possible machine that gives a perfect simulation of understanding, which meets this Behaviorist definition, but which is clearly not a thinking thing.

Some preliminaries: the conversations of concern will take place over a computer terminal. Plenty of conversations take place in this way. (Anyway, for a few thousand dollars, we could equip our machine with voice synthesis and recognition, and let the conversation take place over the telephone. With a Disneyland budget, we could build a robot.) The conversation is to take a finite period of time; you get to pick the duration in advance. It can be any period of time, but there must be a limit. Let's suppose it's one hour. That's usually more than enough for us to tell if we're being understood under normal circumstances, but if you're more cautious, pick a larger number and let me know what it is. Call something a "typable string" if a person can type it into a terminal during an hour. There are lots of typable strings, but the number is finite. Many typable strings consist of nonsense: random assortments of keystrokes. Call something a "sensible string" if it is a typable string that consists of meaningful sentences. The set of sensible strings will be a relatively small subset of the set of typable strings. Simply to cut down on the number of strings we have to deal with, let's restrict them to a particular conversational style. There are many ways to do this. For convenience, we're going to model the set on the conversational style of Mr. Waldo. That subset will be smaller still, and it will consist of all the possible one-hour long conversations that one might have with Mr. Waldo. Call it the "set of all Waldo-strings". We get a large grant for Project Waldo. Assemble a team of programmers and have them spend a year with Mr. Waldo, following him around to workplace, supermarket, malls, bowling alleys, his kitchen at home, etc. Using their feel for his personality, the programmers put all the Waldo-strings into the memory of a big computer. Now, here's how the software works.

You type in a sentence:

The machine searches its memory for all those Waldo-strings that begin with A. There will be many of them. It selects one, randomly. It then prints out the second sentence in the string as its response:

You now respond:

The machine now searches its memory for all Waldo-strings that start with ABC, selects one at random, and prints out its fourth sentence as its response:

Then you respond with E, and the machine searches its memory for Waldo-strings beginning with ABCDE, selects one at random, and prints out its sixth sentence as output. And so it goes for the Unthinking Waldo Simulator. The algorithm that the computer uses to produce outputs for given inputs is a very simple one. It is clearly engaged in no understanding of the inputs: it simply matches input sequences against sequences stored in memory. It functions as a 'sieve for sentences', sorting sentence by their syntactical 'shapes' (written or spoken) rather than their meanings. The software is so simple, it would make a reasonable assignment for a first course in computer programming. Because of the ingenuity that the programmers have exercised, the patterns of inputs and outputs is just like that of a thinking human being. But despite this pattern of behavior, the machine has as much ability to understand English as does a sieve, that is, none at all. As much as we might admire the ingenuity of the programmers, the machine itself cannot think, any more than a TV's thoughts are the programs it transmits.

Unfortunately for Behaviorism, this machine meets the definition of conversational ability - for its behavior matches that of a thinking thing, namely, Mr. Waldo. But the machine is not a thinking thing.

The great virtue of this example is that it does more than merely demonstrate the falsity of Behaviorism. It also yields a diagnosis of the deficiency with which Behaviorism is afflicted. The reason that the Unthinking Waldo Simulator is not a thinking thing is that the sort of internal processing that goes on in it is not of the sort necessary for understanding. This example therefore reminds us of the importance of internal processing. Behaviorism says that internal processing doesn't count. How you get from input to output is, according to Behaviorism, irrelevant to mental processing. But, as we can now see, the 'how' is not only relevant but essential to defining mental activity. Internal processing counts. Behaviorism suffers from pattern poverty precisely because it leaves internal processing out.

Note that only some sorts of internal processing will count as parts of mental processes. Patterns of inputs and outputs - even when they constitute perfect simulation - aren't enough to distinguish the mental sorts of internal processing from the non-mental sorts. We need additional conceptual resources to distinguish mental from non-mental processes.

And that is the central insight of Machine Functionalism, the computer model of the mind. Machine Functionalism tries to overcome the pattern poverty of Behaviorism by adding the third element of internal processing, and by saying exactly what conceptual resources we need to describe such processing.

Additional Sources

2000 David F. Austin

Go to next section: Machine Functionalism