In the mid-twentieth century, digital computers became much more widely available. Impressed as people were by their speed and accuracy, they were sometimes called “electronic brains.” Of course the makeup of computers (then and now) has very little in common with the makeup of biological brains. The more plausible question that came out of all this was whether computers were capable of doing some of the things that human brains were capable of doing.
For John McCarthy, what was most striking about the sort of commonsense behavior exhibited by people was how it was conditioned by knowledge. In deciding what to do in novel situations, people were able to use what they had been told and what they already knew about the world around them. What he had in mind were examples like the following:
A very wide range of intelligent behaviors seemed to be unified by the idea that they all depended in similar ways on a store of background knowledge. Once you know that lemons are yellow, for instance, this helps you do a number of things: paint one in a picture, spot one in a grocery store, make a list of yellow fruits, and others. McCarthy was the first to propose that to understand intelligent behavior we needed to study the knowledge that behavior depends on, and how this knowledge might be applied in deciding how to behave.
Before we get back to the Turing Test and to why McCarthy’s proposal still matters in the next chapter, we need to slow down and take a closer look at the idea of knowledge itself and what it means to say that intelligent behavior depends on it.
In understanding human behavior, there are many occasions where the simplest and best explanation for why we do what we do is in terms of stimulus and response. This is seen quite clearly in other animals, and we exhibit it too, including when we use language. We get our clothing caught on a nail unexpectedly and say “What the heck?” We bump into a person on a crowded train and say “Sorry.” We see a child leaning precariously over a rail and say “Watch out!” We hit our thumb with a hammer and say a number of colorful things.
But it would be a Big Puzzle mistake to think that the rest of our language use was just more of the same. Consider how language is used in the following story:
You are with a group of people talking about the movie 2001: A Space Odyssey. At one point your friend Tim says “The film is just gorgeous on Blu-ray.” You say “It’s hard to believe it first came out in 1968.”
So what do we have here? There is again a stimulus (Tim’s comment) and a response (your comment). But now it is much harder to account for how the stimulus elicits the response. What caused you to say the word “1968,” for instance?
There is clearly a gap here between stimulus and response, and the only plausible explanation about what fills this gap is some thinking on your part. You knew something, namely that the movie was released in that year, and you decided that saying so would advance the conversation in an interesting way. Maybe all you really wanted to do was to support Tim, or to connect with your friends, or to be seen as contributing your share of the conversation. That desire does not by itself explain why you said what you did. What accounts for it is applying what you knew about the movie.
Here is what the psychologist Sir Frederic Bartlett says about this process:
The important characteristics of [the] thinking process, as I am proposing to treat it, can now be stated: The process begins when evidence or information is available which is treated as possessing gaps, or as being incomplete. The gaps are then filled up, or that part of the information which is incomplete is completed. This is done by an extension or supplementation of the evidence, which remains in accord with the evidence (or claims to do so), but carries it further by utilizing other sources of information besides those which started the whole process going, and, in many instances, in addition to those which can be directly identified in the external surroundings. (Thinking, 1958, p. 75)
It is these “other sources of information” that make all the difference.
The back story: At some point you must have found out that 2001 was released in 1968. Maybe you saw the movie when it first came out, maybe you saw that date watching the movie later, maybe you read about the date somewhere, maybe you saw it on a poster, maybe you heard somebody talking about it. But however you found out, that stimulus caused a change in you. You might not have said anything at the time, but you started believing that 2001 was released in 1968. (We will have more to say about the difference between “knowing” and “believing” below.)
Later, when the movie is being discussed and Tim says his bit, thoughts occur to you very quickly. You recall watching the Blu-ray and how impressed you were, just like Tim. But you don’t just agree with Tim or say “Awesome!” You do not want to appear banal or uninformed, and you quickly form a thought you had never consciously entertained before, which is that 2001 looked better than other movies of the same period. But you don’t say that either. You happen to know something specific: the period was 1968. And this is what you bring to the table.
The moral of the story is this:
Knowledge, what you know and can bring to bear on what you are doing, is an essential component of human behavior.
This idea is obvious to many of us. But it does appear to be problematic for many others, mainly certain psychologists. They might say:
Why not simply say that there is one stimulus (reading about 2001, say) that trains you, and a later stimulus (Tim’s comment) that causes you to respond according to your training. Why all this mumbo-jumbo about knowledge and belief that we have no way of accounting for scientifically?
The trouble with the bare stimulus-response story is that it misses something very important. As the philosopher Zenon Pylyshyn has stressed in a related context, the actual stimulus you encountered (when reading about 2001) is neither necessary nor sufficient to account for the response you produced.
First of all, note that what you saw written on a page is only one among very many visual stimuli that will do the trick. The information can be presented in a way that looks nothing at all like what you saw: different paper, different colors, different fonts, different words, even a different language. And there is really nothing visual about it either. You might have heard someone talking about 2001. But this could have been quite different too: different volume, different pitch, different intonation, different words, different language.
Going further, maybe you did not even see or hear the word “1968.” Maybe you heard that 2001 was released the year before Neil Armstrong walked on the moon. Or maybe all you were told is that the Best Picture that year went to the movie Oliver!, which might be enough if you knew enough about Oliver!. In terms of what the sensory stimulus needs to look or sound like, the range is incredibly wide.
Furthermore, the words “The movie 2001 was released in 1968” themselves might not do the trick. If they come right after the words “Here is something that is definitely false”, you would not respond to Tim in the same way. Or maybe you see the words as an item in a long piece entitled “Common misconceptions about the movies of Stanley Kubrick.” Or maybe somebody tells you that 2001 was released in 1968, so that you hear those words, no more, no less, but you have good reason to believe that the person saying them is only guessing or is lying. Again your response to Tim would be different.
In the end, what matters about the stimulus, what causes it to do the job, is not what it looks or sounds like at all. There are many sorts of stimuli in context that will work, and many others that will not. What matters is whether or not they are enough to get you to believe something, namely, that 2001 was released in 1968.
Moreover, once you are in this state of belief, how you got there is no longer important. Your belief can affect any number of other actions you later do, linguistic ones and nonlinguistic ones, regardless of the original stimulus. If you decide to sort your Blu-ray movies by decade, or if you decide to have a moviefest of Great Movies From the Sixties, or if somebody asks you a direct question about Kubrick, all these actions will be affected by this one belief, however it came about.
Intelligent behavior is, in the words of Pylyshyn, cognitively penetrable: the decisions you make about what actions to perform is penetrated by what you believe, just as McCarthy first emphasized. If you come to believe that 2001 was released in 1972, for instance, the actions you choose will all change systematically.
Of course not all our actions are affected by what we believe in this way. Involuntary reflexes, for example, are not. Whether we raise our leg if someone taps us on the knee, or whether we blink as something approaches our eye, or even (to a certain extent) what we say when we hit our thumb with a hammer, these actions are selected very quickly and do not seem to depend on the beliefs we have. But sorting our movies by decade is one action that does.
So to summarize: As with all animals, we act in response to stimuli. In some cases, the mapping is direct: we sense something, and we react to it. But for very many others, the mapping is less direct: we sense something, but how we react depends on the beliefs we happen to have.
What exactly do we mean by knowledge? The basic idea is simple: knowing something means taking the world to be one way and not another. But let us go over this a bit more carefully.
First, observe that when we say something like “John knows that …,” we fill in the blank with a declarative sentence. So we might say that “John knows that 2001 was released in 1968” or that “John knows that Mary will come to the party.” This suggests, among other things, that knowledge is a relation between a knower, like John, and a proposition, that is, the idea expressed by a declarative sentence of English or some other language, like “Mary will come to the party.”
And what are propositions then? They are abstract entities, not unlike numbers, but ones that can be true or false, right or wrong. When we say that “John knows that P,” we can just as well say that “John knows that it is true that P.” Either way, to say that John knows something is to say that John has formed a judgment of some sort, and has come to realize that the world is one way and not another. In talking about this judgment, we use propositions to classify the two cases, for example, those where Mary will come to the party, and those where she won’t.
A similar story can be told about a sentence like “John hopes that Mary will come to the party.” or “John worries that Mary will come to the party.” The same proposition is involved, but the relationship John has to it is different. Verbs like “knows,” “hopes,” “regrets,” “fears,” and “doubts” all denote what philosophers call propositional attitudes, relationships between agents and propositions. In each case, what matters about the proposition is its truth condition, what it takes for it to be true: if John hopes that Mary will come to the party, then John is hoping that the world is a certain way, as classified by the proposition.
Of course, there are sentences involving knowledge that do not explicitly mention propositions. When we say “John knows who Mary is taking to the party,” or “John knows how to get there,” we can at least imagine the implicit propositions: “John knows that Mary is taking … to the party,” or “John knows that to get to the party, you go two blocks past Main Street, turn left …,” and so on. On the other hand, when we say that John has a deep understanding of someone or something, as in “John knows Bill well,” or a skill, as in “John knows how to skate backwards,” it is not so clear that any useful proposition is involved. (While this type of “procedural” knowledge is undoubtedly useful, we will not have much more to say about it, except briefly in the section “Learning behavior” in chapter 5.)
A related notion that we will be very concerned with, however, is the concept of belief. The sentence “John believes that P” is clearly related to “John knows that P.” We use the former when we do not wish to claim that John’s judgment about the world is necessarily accurate or held for appropriate reasons. We also use it when the judgment John has made is not based on evidence, but is perhaps more a matter of faith. We also use it when we feel that John might not be completely convinced. In fact, we have a full range of propositional attitudes, expressed in English by sentences like “John is absolutely certain that P,” “John is confident that P,” “John is of the opinion that P,” “John suspects that P,” and so on, that differ only in the level of conviction they attribute to John.
Clearly how a person uses what he or she believes depends on how certain the person is about what is believed. This is called the degree of belief. For now, it will suit our purposes to not distinguish between knowledge and belief, or to worry about degrees of belief. What matters is that they all share the same idea: John taking the world to be one way and not another. (We will return to degrees of belief briefly in the section “Knowledge representation and reasoning” in chapter 9.)
But this account of knowledge seems to suggest that it can apply only to entities that understand language the way we do. If P is a sentence of English, how can somebody or something know or believe that P is true if they cannot make sense of P? So are humans unique in their use of knowledge?
No. A dog lunging for a flying Frisbee is not simply reacting to what it sees. What it sees is a Frisbee at position A; where it lunges is at position B. It is not too much of a stretch to say that at the time of the jump, the dog anticipates that the Frisbee will be at position B. It could be wrong of course. Maybe there’s a serious crosswind, maybe the Frisbee is on a string, maybe the Frisbee is remote controlled. Nonetheless, the dog has something that goes beyond what it sees, smells, and hears that compels it to jump to a place where there is currently no Frisbee. Why not call this belief? The dog believes the Frisbee will be at position B. Maybe we want to say that this is very simple knowledge, or proto-knowledge (or only “knowledge” in quotation marks), but it seems on a continuum with the sort of knowledge discussed above, even though the dog does not understand English.
Here’s where a little philosophy will help clarify things. The philosopher Daniel Dennett suggests that when we look at complex systems (biological or otherwise), it’s often useful to take what he calls an “intentional stance.” By this he means describing the system as if it had beliefs, goals, desires, plans, intentions, and so on, the same way we talk about people. So we might say of the dog that it believes that the Frisbee will be at position B and wants to catch it when it arrives. We might say of a computer chess program that it worries that its board mobility is at risk and wants to get its knight out early to control the center of the board. We might also say of a thermostat that it believes that the room is warmer than it should be and intends to make it cooler.
In some cases (like the chess program), the intentional stance is useful and helps us decide how best to interact with the system. In other cases (like the thermostat), the intentional stance seems overblown and needlessly anthropomorphic, and the system is better dealt with in different terms. As for dogs, it is clearly useful to think about them in intentional terms, and we do it all the time, as in the following:
What does she want? Why is she scratching at the door? Oh I get it, she thinks her toy is in the other room and she’s trying to get at it! Open the door and show her it’s not there.
Dennett’s main point is that these are stances. They are not inherently right or wrong. There are no real facts to the matter. A stance is just a way of looking at a complex system, and it may or may not be useful.
But it does raise an interesting question: is all the talk of knowledge and belief in people just a stance? Do we really want to say that people behave as if they had beliefs, desires, intentions, the same way we might talk about computer systems, dogs, and thermostats?
We are not ready to answer this question yet. But we will be taking it up in detail in chapters 8 and 9. There we will discuss a type of system (called a knowledge-based system) that is designed to work with its beliefs in a very explicit way. For these systems, beliefs are not just a useful way of talking; they are, in a clear sense, ingredients that cause the system to behave the way that it does.
This is more like the way we talk about gas in a car. A car is obviously designed and built to run on gas (or at least it was, prior to the hybrid and electric ones). So gas-talk is not just some sort of stance that we can take or leave. When it comes to talking about how the car is made to work, there is really no choice but to talk about how it uses gas.
Beliefs for knowledge-based systems will turn out to be the same.
What is intelligent behavior anyway? If this is what we intend to study, don’t we need a definition? We can’t just say it’s how humans behave, since so much of what we do is not very intelligent at all. We find it greatly entertaining to watch ourselves at our very dumbest, like the clueless daredevils on America’s Funniest Home Videos or the squabbling couples on The Jerry Springer Show. There is so much buffoonery, recklessness, poor judgment, and sheer idiocy all around us, that behavior we would label as truly intelligent appears to be more the exception than the norm! As Eric Idle sings in the Monty Python movie The Meaning of Life:
And pray that there’s intelligent life somewhere out in space,
’Cause there’s bugger all down here on Earth!
Somewhat of an exaggeration, maybe, but the point remains.
So what do we have in mind for those cases—rare as they may be—when a person is actually behaving in an intelligent way? We do not want to limit ourselves to things like playing chess, discussing Heidegger, or solving calculus problems. These “intellectual” activities are fine, of course, but we are looking for something much more mundane. Roughly speaking, we want to say that people are behaving intelligently when they are making effective use of what they know to get what they want. The psychologist Nicholas Humphrey puts it this way: “An animal displays intelligence when he modifies his behavior on the basis of valid inference from evidence.”
Imagine a person, Henry, thinking aloud as follows:
Where are my car keys? I need them. I know they’re in my coat pocket or on the fridge. That’s where I always leave them. But I just felt in my coat pocket, and there’s nothing there.
What do we expect Henry to do next? The obvious answer is that he will think about his keys being on the fridge. This is what intelligent behavior is all about: using what you know to make decisions about what to do. Henry wants those keys, and we expect him to draw an inference to get to them.
But will he? Although this is the “right” thing for him to do, and Henry might well agree with us (for instance, watching a replay of his own behavior later), what he actually does could be quite different. He might get sidetracked and think about something completely unrelated, like going to the washroom, or eating pizza, or what other people are thinking about him. He might make a joke. He might just collapse to his knees, sobbing “I can’t do this anymore!” Even if he really wants those keys, what he ends up doing involves many other factors, like these:
So if we are truly interested in the intelligent behavior that people like Henry produce, do we need to concern ourselves with factors like these too?
The answer is: it depends on what we want to study. As it turns out, we are not so interested in accounting for the behavior that people like Henry produce; it is too complicated and too fiddly. What we are interested in is a certain idealization of that behavior, that is, behavior that people like Henry would recognize as being intelligent, even if they admit that they sometimes fall short of achieving it.
A similar distinction has come into play in the study of human language in what the linguist Noam Chomsky called the competence/performance distinction. The idea roughly is this: linguists want to understand human language, but by this they do not necessarily mean the actual linguistic output of human subjects, which is loaded with quirks of various sorts.
For example, consider a sentence that a person might write, like this one:
The hockey players celebrated there first win.
We might be tempted to ask: what is the grammatical role of the word “there” in this sentence? The answer, of course, is that it has no role; it’s a mistake, confusing the adverb “there” with the adjective “their.” From a linguistic point of view, this use of “there” is not really part of the language, so there is nothing grammatical to explain. And yet people do write this way! Similar considerations apply to the “umm,” “like,” and other hesitations that are part of everyday speech. We recognize that they are really not part of the language even though people say them all the time.
Linguists call the utterances speakers actually produce the performance of native speakers. However, they may prefer to study what the speakers of a language would recognize as genuine grammatical expressions of their language, which they call the competence of native speakers. While there is nothing wrong with wanting to study the performance of speakers, it is phenomenally complex. The advantage of looking at the competence of speakers is that we get to see certain abstractions and generalizations that might otherwise get lost in the vagaries of actual speech.
As a very simple example, consider the length of sentences. We recognize that whenever we have two declarative sentences, we can join them with an “and” to make a new one. But this recognition is only a competence judgment. It implies that there is no limit to the length of sentences, whereas there is quite clearly a limit to the length of a sentence that a person can produce in a lifetime.
When it comes to intelligent behavior, we can take a similar position: instead of trying to make sense of intelligent behavior in a way that accounts for the actual choices of real people (who may be under the influence of alcohol, or have short attention spans, flagging motivation, and so on), we focus on what people would recognize as proper, reasonable, intelligent behavior. This will give us something much more manageable to study.