Thinking computers must hallucinate too

David Gelernter, Yale University, Project Syndicate

Creating a computer that "thinks" is one goal of artificial intelligence research. In 1950, the great British mathematician Alan Turing guessed that "thinking computers" would arrive by 2000. But computers capable of simulating human thought processes are no closer today than they were then, because we put science before common sense.

Put simply, we failed to notice obvious and important facts about the nature of thought that exist right before our eyes. Philosophy is intended to show us facts that are so obvious that we might miss them. Most philosophers, it seems, have been asleep on the job for a long time.

Consider four familiar situations. In one, you are thinking about a problem -- say, what pension plan you should join -- logically and analytically. In the second, you are thinking casually, drawing on experience rather than analysis. (My computer is stuck. The last time this happened, I fixed it by doing such and such; I'll try that again.) In the third, you are looking out a window, letting your mind wander. Finally, you are asleep and dreaming.

The single most important fact about thought follows from an obvious observation: These four styles are connected. We can label them "analysis," "common sense," "free association," and "dreaming." But the key point is that they are four points on a single continuous spectrum, with analysis at one end and dreaming at the other.

Psychologists and computer scientists like to talk about analysis and common sense as if they were salt and steel, or apples and oranges. We would do better to think of them as red and yellow, separated not by some sharp boundary, but by a continuous range of red-oranges and orange-yellows.

Most people do their best analytic thinking when they are wide-awake. As your alertness declines, your analytic ability declines -- and you rely less on analysis, more on common sense. As your alertness declines further, your mind starts to wander. At some point you are looking out the window, free-associating. Falling asleep is also a gradual transition. (One of experimental psychology's most interesting findings is that people start to dream before they fall asleep, while they are still physiologically awake.) Free-association and dreaming are blue and violet, with a beautiful spectrum of violet-blues in between.

To understand color, it is not enough to understand red and yellow, or blue and violet. Understanding color requires understanding the entire spectrum. Likewise, to understand thought, it is not enough to understand "analysis" and "common sense," or "free association" and "dreaming." Understanding thought implies understanding the entire cognitive spectrum.

It is therefore nonsense to believe that we can simulate thought on a computer without first understanding and simulating all of its component hues. Dreaming, for example, is merely hallucinating while you sleep. So, to put the same point another way, no thinking computer is possible until we can build a computer that hallucinates.

To understand the mind, we can conceive of it as being equipped with an imaginary dial marked "mental focus" or "alertness" or "attention." (I do not know the physiological basis of "mental focus"; physiologists cannot tell me, because most of them also do not know what "mental focus" is. It is right in front of their noses, too close to notice.)

Different people have different "cognitive personalities." Some are more comfortable than others thinking analytically, with their dials at "high focus." Everyone's focus oscillates during the day. Your analytic abilities decline as your "wide awakedness" declines. You cannot fall asleep unless your focus setting dips below some threshold. (If we understood the physiological basis of mental focus, we would almost certainly have a better understanding about sleep disorders.)

Psychologists tried for generations to understand creativity, which is distinct from analytical thinking. Creativity evidently depends on inventing analogies. If you are good at analogies ("that fluttering oak leaf reminds me of a butterfly wing"; "that 17th-century clock is a computer with a round screen"), you have a basis for creative thinking, for approaching old problems in new ways. But how do we invent analogies?

The answer must have something to do with the thought- spectrum. Notice the similarity between falling asleep and thinking creatively -- no matter how hard you try, you can not force yourself to do either. You cannot make yourself fall asleep by concentrating. To fall asleep you must un-concentrate.

Various studies have reached the same conclusion about analogies. Creative thinkers are most productive not when they are concentrating but when they are un-concentrating, when they are thinking not about the problem to be solved but about something else. So, to think analytically you must set your "focus" to maximum; to think creatively, you must let your focus setting decline -- but not so far that you fall asleep.

To complete the picture, we notice that emotion becomes increasingly important as "focus" falls. Emotions then become "signatures" or "encodings" of complex thoughts. When two ideas seem unrelated, but make you feel the same way, you have a basis for connecting them -- and for inventing a new analogy. Viewed from this perspective, it appears self-evident that no "thinking computer" whose focus is permanently set on high can come close to simulating human thought.