THE AGE OF INTELLIGENT MACHINES | A Coffeehouse Conversation on the Turing Test

September 24, 2001
Author:
Ray Kurzweil

Douglas R. Hofstadter is Professor of Cognitive Science at Indiana University in Bloomington. Hofstadter received the 1980 Pulitzer Prize for Gödel, Escher, Bach: An Eternal Golden Braid (1979). He is the author of numerous other books and articles, including The Mind’s I (1981), coedited with Daniel Dennett, and, most recently, Metamagical Themas (1985), a collection of articles from his Scientific American column.

Participants in the dialog: Chris, a physics student; Pat, a biology student; Sandy, a philosophy student.

Chris: Sandy, I want to thank you for suggesting that I read Alan Turing’s article “Computing Machinery and Intelligence.” It’s a wonderful piece and certainly made me think-and think about my thinking.

Sandy: Glad to hear it. Are you still as much of a skeptic about artificial intelligence as you used to be?

Chris: You’ve got me wrong. I’m not against artificial intelligence; I think it’s wonderful stuff-perhaps a little crazy, but why not? I simply am convinced that you AI advocates have far underestimated the human mind, and that there are things a computer will never, ever be able to do. For instance, can you imagine a computer writing a Proust novel? The richness of imagination, the complexity of the characters-

Sandy: Rome wasn’t built in a day!

Chris: In the article, Turing comes through as an interesting person. Is he still alive?

Sandy: No, he died back in 1954, at just 41. He’d be only 70 or so now, although he is such a legendary figure it seems strange to think that he could still be living today.

Chris: How did he die?

Sandy: Almost certainly suicide. He was homosexual, and had to deal with some pretty barbaric treatment and stupidity from the outside world. In the end, it got to be too much, and he killed himself.

Chris: That’s horrendous, especially in this day and age.

Sandy: I know. What really saddens me is that he never got to see the amazing progress in computing machinery and theory that has taken place since 1954. Can you imagine how he’d have been wowed?

Chris: Yeah . . .

Pat: Hey, are you two going to clue me in as to what this Turing article is about?

Sandy: It is really about two things. One is the question “Can a machine think?”-or rather, “Will a machine ever think?” The way Turing answers the question-he thinks the answer is yes, by the way-is by batting down a series of objections to the idea, one after another. The other point he tries to make is that, as it stands, the question is not meaningful. It’s too full of emotional connotations. Many people are upset by the suggestion that people are machines, or that machines might think. Turing tries to defuse the question by casting it in less emotional terms. For instance, what do you think, Pat, of the idea of thinking machines?

Pat: Frankly, I find the term confusing. You know what confuses me? It’s those ads in the newspapers and on TV that talk about “products that think” or “intelligent ovens” or whatever. I just don’t know how seriously to take them.

Sandy: I know the kind of ads you mean, and they probably confuse a lot of people. On the one hand, we’re always hearing the refrain “Computers are really dumb; you have to spell everything out for them in words of one syllable”-yet on the other hand, we’re constantly bombarded with advertising hype about “smart products.”

Chris: That’s certainly true. Do you know that one company has even taken to calling its products “dumb terminals” in order to stand out from the crowd?

Sandy: That’s a pretty clever gimmick, but even so it just contributes to the trend toward obfuscation. The term “electronic brain” always comes to my mind when I’m thinking about this. Many people swallow it completely, and others reject it out of hand. It takes patience to sort out the issues and decide how much of it makes sense.

Pat: Does Turing suggest some way of resolving it, some kind of IQ test for machines?

Sandy: That would be very interesting, but no machine could yet come close to taking an IQ test. Instead, Turing proposes a test that theoretically could be applied to any machine to determine whether or not it can think.

Pat: Does the test give a clear-cut yes-or-no answer? I’d be skeptical if it claimed to.

Sandy: No, it doesn’t claim to. In a way that’s one of its advantages. It shows how the borderline is quite fuzzy and how subtle the whole question is.

Pat: And so, as usual in philosophy, it’s all just a question of words!

Sandy: Maybe, but they’re emotionally charged words, and so it’s important, it seems to me, to explore the issues and try to map out the meanings of the crucial words. The issues are fundamental to our concept of ourselves. So we shouldn’t just sweep them under the rug.

Pat: Okay, so tell me how Turing’s test works.

Sandy: The idea is based on what he calls the Imitation Game. Imagine that a man and a woman go into separate rooms, and from there they can be interrogated by a third party via some sort of teletype set-up. The third party can address questions to either room, but has no idea which person is in which room. For the interrogator, the idea is to determine which room the woman is in. The woman, by her answers, tries to help the interrogator as much as she can. The man, though, is doing his best to bamboozle the interrogator, by responding as he thinks a woman might. And if he succeeds in fooling the interrogator…

Pat: The interrogator only gets to see written words, eh? And the sex of the author is supposed to shine through? That game sounds like a good challenge. I’d certainly like to take part in it someday. Would the interrogator have met either the man or the woman before the test began? Would any of them know any of the others?

Sandy: That would probably be a bad idea. All kinds of subliminal cueing might occur if the interrogator knew one or both of them. It would certainly be best if all three people were totally unknown to one another.

Pat: Could you ask any questions at all, with no holds barred?

Sandy: Absolutely. That’s the whole idea!

Pat: Don’t you think, then, that pretty quickly it would degenerate into sex-oriented questions? I mean, I can imagine the man, overeager to act convincing, giving away the game by answering some very blunt questions that most women would find too personal to answer, even through an anonymous computer connection.

Sandy: That’s a nice observation. I wonder if it’s true ….

Chris: Another possibility would be to probe for knowledge of minute aspects of traditional sex-role differences, by asking about such things as dress sizes and so on. The psychology of the Imitation Game could get pretty subtle. I suppose whether the interrogator was a woman or a man would make a difference. Don’t you think that a woman could spot some telltale differences more quickly than a man could?

Pat: If so, maybe the best way to tell a man from a woman is to let each of them play interrogator in an Imitation Game and see which of the two is better at telling a man from a woman!

Sandy: Hmm . . . that’s a droll twist. Oh well, I don’t know if this original version of the Imitation Game has ever been seriously tried out, despite the fact that it would be relatively easy to do with modern computer terminals. I have to admit, though, that I’m not at all sure what it would prove, whichever way it turned out.

Pat: I was wondering about that. What would it prove if the interrogator-say a woman-couldn’t tell correctly which person was the woman? It certainly wouldn’t prove that the man was a woman!

Sandy: Exactly! What I find funny is that although I strongly believe in the idea of the Turing Test, I’m not so sure I understand the point of its basis, the Imitation Game.

Chris: As for me, I’m not any happier with the Turing Test as a test for thinking machines than I am with the Imitation Game as a test for femininity.

Pat: From what you two are saying, I gather the Turing Test is some kind of extension of the Imitation Game, only involving a machine and a person instead of a man and a woman.

Sandy: That’s the idea. The machine tries its hardest to convince the interrogator that it is the human being, and the human tries to make it clear that he or she is not the computer.

Pat: The machine tries? Isn’t that a loaded way of putting it?

Sandy: Sorry, but that seemed the most natural way to say it.

Pat: Anyway, this test sounds pretty interesting. But how do you know that it will get at the essence of thinking? Maybe it’s testing for the wrong things. Maybe, just to take a random illustration, someone would feel that a machine was able to think only if it could dance so well that you couldn’t tell it was a machine. Or someone else could suggest some other characteristic. What’s so sacred about being able to fool people by typing at them?

Sandy: I don’t see how you can say such a thing. I’ve heard that objection before, but frankly, it baffles me. So what if the machine can’t tap-dance or drop a rock on your toe? If it can discourse intelligently on any subject you want, then it has shown that it can think-to me, at least! As I see it, Turing has drawn, in one clean stroke, a clear division between thinking and other aspects of being human.

Pat: Now you’re the baffling one. If you couldn’t conclude anything from a man’s ability to

win at the Imitation Game, how could you conclude anything from a machine’s ability to win at the Turing Game?

Chris: Good question.

Sandy: It seems to me that you could conclude something from a man’s win in the Imitation Game. You wouldn’t conclude he was a woman, but you could certainly say he had good insights into the feminine mentality (if there is such a thing). Now, if a computer could fool someone into thinking it was a person, I guess you’d have to say something similar about it-that it had good insights into what it’s like to be human, into “the human condition” (whatever that is).

Pat: Maybe, but that isn’t necessarily equivalent to thinking, is it? It seems to me that passing the Turing Test would merely prove that some machine or other could do a very good job of

simulating thought.

Chris: I couldn’t agree more with Pat. We all know that fancy computer programs exist today for simulating all sorts of complex phenomena. In theoretical physics, for instance, we simulate the behavior of particles, atoms, solids, liquids, gases, galaxies, and so on. But no one confuses any of those simulations with the real thing!

Sandy: In his book Brainstorms, the philosopher Daniel Dennett makes a similar point about simulated hurricanes.

Chris: That’s a nice example, too. Obviously, what goes on inside a computer when it’s simulating a hurricane is not a hurricane, for the machine’s memory doesn’t get torn to bits by 200 mile-an-hour winds, the floor of the machine room doesn’t get flooded with rainwater, and so on.

Sandy: Oh, come on-that’s not a fair argument! In the first place, the programmers don’t claim the simulation really is a hurricane. It’s merely a simulation of certain aspects of a hurricane. But in the second place, you’re pulling a fast one when you imply that there are no downpours or 200-mile-an-hour winds in a simulated hurricane. To us there aren’t any, but if the program were incredibly detailed, it could include simulated people on the ground who would experience the wind and the rain just as we do when a hurricane hits. In their minds-or, if you’d rather, in their simulated minds-the hurricane would be not a simulation, but a genuine phenomenon complete with drenching and devastation.

Chris: Oh, my-what a science-fiction scenario! Now we’re talking about simulating whole populations, not just a single mind!

Sandy: Well, look-I’m simply trying to show you why your argument that a simulated McCoy isn’t the real McCoy is fallacious. It depends on the tacit assumption that any old observer of the simulated phenomenon is equally able to assess what’s going on. But in fact, it may take an observer with a special vantage point to recognize what is going on. In the hurricane case, it takes special “computational glasses” to see the rain and the winds.

Pat: “Computational glasses”? I don’t know what you’re talking about.

Sandy: I mean that to see the winds and the wetness of the hurricane, you have to be able to look at it in the proper way. You-

Chris: No, no, no! A simulated hurricane isn’t wet! No matter how much it might seem wet to simulated people, it won’t ever be genuinely wet! And no computer will ever get torn apart in the process of simulating winds.

Sandy: Certainly not, but that’s irrelevant. You’re just confusing levels. The laws of physics don’t get torn apart by real hurricanes, either. In the case of the simulated hurricane, if you go peering at the computer’s memory, expecting to find broken wires and so forth, you’ll be disappointed. But look at the proper level. Look into the structures that are coded for in memory. You’ll see that many abstract links have been broken, many values of variables radically changed, and so on. There’s your flood, your devastation-real, only a little concealed, a little hard to detect.

Chris: I’m sorry, I just can’t buy that. You’re insisting that I look for a new kind of devastation, one never before associated with hurricanes. That way you could call anything a hurricane as long as its effects, seen through your special “glasses,” could be called “floods and devastation.”

Sandy: Right-you’ve got it exactly! You recognize a hurricane by its effects. You have no way of going in and finding some ethereal “essence of hurricane,” some “hurricane soul” right in the middle of the storm’s eye. Nor is there any ID card to be found that certifies “hurricanehood.” It’s just the existence of a certain kind of pattern–a spiral storm with an eye and so forth-that makes you say it’s a hurricane. Of course, there are a lot of things you’ll insist on before you call something a hurricane.

Pat: Well, wouldn’t you say that being an atmospheric phenomenon is one prerequisite? How can anything inside a computer be a storm? To me, a simulation is a simulation is a simulation!

Sandy: Then I suppose you would say that even the calculations computers do are simulated-that they are fake calculations. Only people can do genuine calculations, right?

Pat: Well, computers get the right answers, so their calculations are not exactly fake-but they’re still just patterns. There’s no understanding going on in there. Take a cash register. Can you honestly say that you feel it is calculating something when its gears mesh together? And the step from cash register to computer is very short, as I understand things.

Sandy: If you mean that a cash register doesn’t feel like a schoolkid doing arithmetic problems, I’ll agree. But is that what “calculation” means? Is that an integral part of it? If so, then contrary to what everybody has thought up till now, we’ll have to write a very complicated program indeed to perform genuine calculations.

Of course, this program will sometimes get careless and make mistakes, and it will sometimes scrawl its answers illegibly, and it will occasionally doodle on its paper …. It won’t be any more reliable than the store clerk who adds up your total by hand. Now, I happen to believe that eventually such a program could be written. Then we’d know something about how clerks and schoolkids work.

Pat: I can’t believe you’d ever be able to do that!

Sandy: Maybe, maybe not, but that’s not my point. You say a cash register can’t calculate. It reminds me of another favorite passage of mine from Dennett’s Brainstorms. It goes something like this: “Cash registers can’t really calculate; they can only spin their gears. But cash registers can’t really spin their gears, either: they can only follow the laws of physics.” Bennett said it originally about computers; I modified it to talk about cash registers. And you could use the same line of reasoning in talking about people: “People can’t really calculate; all they can do is manipulate mental symbols. But they aren’t really manipulating symbols: all they are doing is firing various neurons in various patterns. But they can’t really make their neurons fire; they simply have to let the laws of physics make them fire for them.” Et cetera. Don’t you see how this reduction ad absurdum would lead you to conclude that calculation doesn’t exist, that hurricanes don’t exist-in fact, that nothing at a level higher than particles and the laws of physic exists? What do you gain by saying that a computer only pushes symbols around and doesn’t truly calculate?

Pat: The example may be extreme, but it makes my point that there is a vast difference between a real phenomenon and any simulation of it. This is so for hurricanes, and even more so for human thought.

Sandy: Look, I don’t want to get too tangled up in this line of argument, but let me try one more example. If you were a radio ham listening to another ham broadcasting in Morse code and you were responding in Morse code, would it sound funny to you to refer to “the person at the other end”?

Pat: No, that would sound okay, although the existence of a person at the other end would be an assumption.

Sandy: Yes, but you wouldn’t be likely to go and check it out. You’re prepared to recognize personhood through those rather unusual channels. You don’t have to see a human body or hear a voice. All you need is a rather abstract manifestation-a code, as it were. What I’m getting at is this. To “see” the person behind the dits and dahs, you have to be willing to do some decoding, some interpretation. It’s not direct perception; it’s indirect. You have to peel off a layer or two to find the reality hidden in there. You put on your “radio-ham’s glasses” to “see” the person behind the buzzes. Just the same with the simulated hurricane! You don’t see it darkening the machine room; you have to decode the machine’s memory. You have to put on special “memory-decoding” glasses. Then what you see is a hurricane.

Pat: Oh ho ho! Talk about fast ones-wait a minute! In the case of the shortwave radio, there’s a real person out there, somewhere in the Fiji Islands or wherever. My decoding act as I sit by my radio simply reveals that that person exists. It’s like seeing a shadow and concluding there’s an object out there, casting it. One doesn’t confuse the shadow with the object, however! And with the hurricane there’s no real storm behind the scenes, making the computer follow its patterns. No, what you have is just a shadow hurricane without any genuine hurricane. I just refuse to confuse shadows with reality.

Sandy: All right. I don’t want to drive this point into the ground. I even admit it is pretty silly to say that a simulated hurricane is a hurricane. But I wanted to point out that it’s not as silly as you might think at first blush. And when you turn to simulated thought then you’ve got a very different matter on your hands from simulated hurricanes.

Pat: I don’t see why. You’ll have to convince me.

Sandy: Well, to do so, I’ll first have to make a couple of extra points about hurricanes.

Pat: Oh no! Well, all right, all right.

Sandy: Nobody can say just exactly what a hurricane is-that is, in totally precise terms. There’s an abstract pattern that many storms share, and it’s for that reason we call those storms hurricanes. But it’s not possible to make a sharp distinction between hurricanes and no hurricanes. There are tornados, cyclones, typhoons, dust devils …. Is the Great Red Spot on Jupiter a hurricane? Are sunspots hurricanes? Could there be a hurricane in a wind tunnel? In a test tube? In your imagination, you can even extend the concept of “hurricane” to include a microscopic storm on the surface of a neutron star.

Chris: That’s not so far-fetched, you know. The concept of “earthquake” has actually been extended to neutron stars. The astrophysicists say that the tiny changes in rate that once in a while are observed in the pulsing of a pulsar are caused by “glitches”starquakes-that have just occurred on the neutron star’s surface.

Sandy: Oh, I remember that now. That “glitch” idea has always seemed eerie to me-a surrealistic kind of quivering on a surrealistic kind of surface.

Chris: Can you imagine-plate tectonics on a giant sphere of pure nuclear matter?

Sandy: That’s a wild thought. So, starquakes and earthquakes can both be subsumed into a new, more abstract category. And that’s how science constantly extends familiar concepts, taking them further and further from familiar experience and yet keeping some essence constant. The number system is the classic example-from positive numbers to negative numbers, then rationale, reels, complex numbers, and “on beyond zebra,” as Dr. Seuss says.

Pat: I think I can see your point, Sandy. In biology, we have many examples of close relationships that are established in rather abstract ways. Often the decision about what family some species belongs to comes down to an abstract pattern shared at some level. Even the concepts of “male” and “female” turn out to be surprisingly abstract and elusive. When you base your system of classification on very abstract patterns, I suppose that a broad variety of phenomena can fall into “the same class,” even if in many superficial ways the class members are utterly unlike one another. So perhaps I can glimpse, at least a little, how to you, a simulated hurricane could, in a funny sense, be a hurricane.

Chris: Perhaps the word that’s being extended is not “hurricane,” but “be.”

Pat: How so?

Chris: If Turing can extend the verb “think,” can’t I extend the verb “be”? All I mean is that when simulated things are deliberately confused with genuine things, somebody’s doing a lot of philosophical wool pulling. It’s a lot more serious than just extending a few nouns, such as “hurricane.”

Sandy: I like your idea that “be” is being extended, but I sure don’t agree with you about the wool pulling. Anyway, if you don’t object, let me just say one more thing about simulated hurricanes and then I’ll get to simulated minds. Suppose you consider a really deep simulation of a hurricane-I mean a simulation of every atom, which I admit is sort of ridiculous, but still, just consider it for the sake of argument.

Pat: Okay.

Sandy: I hope you would agree that it would then share all the abstract structure that defines the “essence of hurricanehood.” So what’s to keep you from calling it a hurricane?

Pat: I thought you were backing off from that claim of equality.

Sandy: So did I, but then these examples came up, and I was forced back to my claim. But let me back off, as I said I would do, and get back to thought, which is the real issue here. Thought, even more than hurricanes, is an abstract structure, a way of describing some complex events that happen in a medium called a brain. But actually, thought can take place in any one of several billion brains. There are all these physically very different brains, and yet they all support “the same thing”: thinking. What’s important, then, is the abstract pattern, not the medium. The same kind of swirling can happen inside any of them, so no person can claim to think more “genuinely” than any other. Now, if we come up with some new kind of medium in which the same style of swirling takes place, could you deny that thinking is taking place in it?

Pat: Probably not, but you have just shifted the question. The question now is: How can you determine whether the “same style” of swirling is really happening?

Sandy: The beauty of the Turing Test is that it tells you when! Don’t you see?

Chris: I don’t see that at all. How would you know that the same style of activity was going on inside a computer as inside my mind, simply because it answered questions as I do? All you’re looking at is its outside.

Sandy: I’m sorry, I disagree entirely! How do you know that when I speak to you, anything similar to what you call thinking is going on inside me? The Turing Test is a fantastic probe, something like a particle accelerator in physics. Here, Chris-I think you’ll like this analogy. Just as in physics, when you want to understand what is going on at an atomic or subatomic level, since you can’t see it directly, you scatter accelerated particles off a target and observe their behavior. From this, you infer the internal nature of the target. The Turing Test extends this idea to the mind. It treats the mind as a “target” that is not directly visible but whose structure can be deduced more abstractly. By “scattering” questions off a target mind, you learn about its internal workings, just as in physics.

Chris: Well . . . to be more exact, you can hypothesize about what kinds of internal structures might account for the behavior observed-but please remember that they may or may not in fact exist.

Sandy: Hold on, now! Are you suggesting that atomic nuclei are merely hypothetical entities? After all, their existence (or should I say hypothetical existence?) was proved (or should I say suggested?) by the behavior of particles scattered off atoms.

Chris: I would agree, but you know, physical systems seem to me to be much simpler than the mind, and the certainty of the inferences made is correspondingly greater. And the conclusions are confirmed over and over again by different types of experiments.

Sandy: Yes, but those experiments still are of the same sort-scattering, detecting things indirectly. You can never handle an electron or a quark. Physics experiments are also correspondingly harder to do and to interpret. Often they take years and years, and dozens of collaborators are involved. In the Turing Test, though, just one person could perform many highly delicate experiments in the course of no more than an hour. I maintain that people give other people credit for being conscious simply because of their continual external monitoring of other people-which is itself something like a Turing Test.

Pat: That may be roughly true, but it involves more than just conversing with people through a teletype. We see that other people have bodies, we watch their faces and expressions-we see they are human beings, and so we think they think.

Sandy: To me, that seems a narrow, anthropocentric view of what thought is. Does that mean you would sooner say a mannequin in a store thinks than a wonderfully programmed computer, simply because the mannequin looks more human?

Pat: Obviously, I would need more than just vague physical resemblance to the human form to be willing to attribute the power of thought to an entity. But that organic quality, the sameness of origin, undeniably lends a degree of credibility that is very important.

Sandy: Here we disagree. I find this simply too chauvinistic. I feel that the key thing is a similarity of internal structure-not bodily, organic, chemical structure but organizational structure software. Whether an entity can think seems to me a question of whether its organization can be described in a certain way, and I’m perfectly willing to believe that the Turing Test detects the presence or absence of that mode of organization. I would say that your depending on my physical body as evidence that I am a thinking being is rather shallow. The way I see it, the Turing Test looks far deeper than at mere external form.

Pat: Hey now-you’re not giving me much credit. It’s not just the shape of a body that lends weight to the idea that there’s real thinking going on inside. It’s also, as I said, the idea of common origin. It’s the idea that you and I both sprang from DNA molecules, an idea to which I attribute much depth. Put it this way: the external form of human bodies reveals that they share a deep biological history, and it’s that depth that lends a lot of credibility to the notion that the owner of such a body can think.

Sandy: But that is all indirect evidence. Surely you want some direct evidence. That’s what the Turing Test is for. And I think it’s the only way to test for thinkinghood.

Chris: But you could be fooled by the Turing Test, just as an interrogator could mistake a man for a woman.

Sandy: I admit, I could be fooled if I carried out the test in too quick or too shallow a way. But I would go for the deepest things I could think of.

Chris: I would want to see if the program could understand jokes-or better yet, make them! That would be a real test of intelligence.

Sandy: I agree that humor probably is an acid test for a supposedly intelligent program, but equally important to me perhaps more so-would be to test its emotional responses. So I

would ask it about its reactions to certain pieces of music or works of literature-especially my favorite ones.

Chris: What if it said, “I don’t know that piece,” or even, “I have no interest in music”? What if it tried its hardest (oops!-sorry, Pat!) …. Let me try that again. What if it did everything it could, to steer clear of emotional topics and references?

Sandy: That would certainly make me suspicious. Any consistent pattern of avoiding certain issues would raise serious doubts in my mind as to whether I was dealing with a thinking being.

Chris: Why do you say that? Why not just conclude you’re dealing with a thinking but unemotional being?

Sandy: You’ve hit upon a sensitive point. I’ve thought about this for quite a long time, and I’ve concluded that I simply can’t believe emotions and thought can be divorced. To put it another way, I think emotions are an automatic by-product of the ability to think. They are entailed by the very nature of thought.

Chris: That’s an interesting conclusion, but what if you’re wrong? What if I produced a machine that could think but not emote? Then its intelligence might go unrecognized because it failed to pass your kind of test.

Sandy: I’d like you to point out to me where the boundary line between emotional questions and nonemotional ones lies. You might want to ask about the meaning of a great novel. This certainly requires an understanding of human emotions! Now is that thinking, or merely cool calculation? You might want to ask about a subtle choice of words. For that, you need an understanding of their connotations. Turing uses examples like this in his article. You might want to ask for advice about a complex romantic situation. The machine would need to know a lot about human motivations and their roots. If it failed at this kind of task, I would not be much inclined to say that it could think. As far as I’m concerned, thinking, feeling, and consciousness are just different facets of one phenomenon, and no one of them can be present without the others.

Chris: Why couldn’t you build a machine that could feel nothing (we all know machines don’t feel anything!), but that could think and make complex decisions anyway? I don’t see any contradiction there.

Sandy: Well, I do. I think that when you say that, you are visualizing a metallic, rectangular machine, probably in an airconditioned room-a hard, angular, cold object with a million colored wires inside it, a machine that sits stock still on a tiled floor, humming or buzzing or whatever, and spinning its tapes. Such a machine can play a good game of chess, which, I freely admit, involves a lot of decision making. And yet I would never call it conscious.

Chris: How come? To mechanists, isn’t a chess-playing machine rudimentarily conscious?

Sandy: Not to this mechanist! The way I see it, consciousness has got to come from a precise pattern of organization, one we haven’t yet figured out how to describe in any detailed way. But I believe we will gradually come to understand it. In my view, consciousness requires a certain way of mirroring the external universe internally, and the ability to respond to that external reality on the basis of the internally represented model. And then in addition, what’s really crucial for a conscious machine is that it should incorporate a well-developed and flexible sell-model. And it’s there that all existing programs, including the best chess-playing ones, fall down.

Chris: Don’t chess programs look ahead and say to themselves as they’re figuring out their next move, “if my opponent moves here, then I’ll go there, and then if they go this way, I could go below that way…”? Doesn’t that usage of the concept “I” require a sort of self-model?

Sandy: Not really. Or, if you want, it’s an extremely limited one. It’s an understanding of self in only the narrowest sense. For instance, a chess-playing program has no concept of why it is

playing chess, or of the fact that it is a program, or is in a computer, or has a human opponent. It has no idea about what winning and losing are, or-

Pat: How do you know it has no such sense? How can you presume to say what a chess program feels or knows?

Sandy: Oh come on! We all know that certain things don’t feel anything or know anything. A thrown stone doesn’t know anything about parabolas, and a whirling fan doesn’t know anything about air. It’s true I can’t prove those statements-but here we are verging on questions of faith.

Pat: This reminds me of a Taoist story I read. It goes something like this. Two sages were standing on a bridge over a stream. One said to the other, “I wish I were a fish. They are so happy.” The other replied, “How do you know whether fish are happy or not? You’re not a fish!” The first said, “But you’re not me, so how do you know whether I know how fish feel?”

Sandy: Beautiful! Talking about consciousness really does call for a certain amount of restraint. Otherwise, you might as well just jump on the solipsism bandwagon (“I am the only conscious being in the universe”) or the panpsychism bandwagon (“Everything in the universe is conscious!”).

Pat: Well, how do you know? Maybe everything is conscious.

Sandy: Oh Pat, if you’re going to join the club that maintains that stones and even particles like electrons have some sort of consciousness, then I guess we part company here. That’s a kind of mysticism I just can’t fathom. As for chess programs, I happen to know how they work, and I can tell you for sure that they aren’t conscious. No way!

Pat: Why not?

Sandy: They incorporate only the barest knowledge about the goals of chess. The notion of “playing” is turned into the mechanical act of comparing a lot of numbers and choosing the biggest one over and over again. A chess program has no sense of disappointment about losing, or pride in winning. Its self-model is very crude. It gets away with doing the least it can, just enough to play a game of chess and nothing more. Yet interestingly enough, we still tend to talk about the “desires” of a chess-playing computer. We say, “it wants to keep its king behind a row of pawns” or “it likes to get its rooks out early” or “it thinks I don’t see that hidden fork.”

Pat: Yes, and we do the same thing with insects. We spot a lonely ant somewhere and say, “It’s trying to get hack home” or “It wants to drag that dead bee back to the colony.” In fact, with any animal we use terms that indicate emotions, but we don’t know for certain how much the animal feels. I have no trouble talking about dogs and cats being happy or sad, having desires and beliefs and so on, but of course I don’t think their sadness is as deep or complex as human sadness is.

Sandy: But you wouldn’t call it “simulated” sadness,

Pat: No, of course not. I think it’s real.

Sandy: It’s hard to avoid use of such teleological or mentalistic terms. I believe they’re quite justified, although they shouldn’t be carried too far. They simply don’t have the same richness of meaning when applied to present-day chess programs as when applied to people.

Chris: I still can’t see that intelligence has to involve emotions. Why couldn’t you imagine an intelligence that simply calculates and has no feelings?

Sandy: A couple of answers here. Number one, any intelligence has to have motivations. It’s simply not the case, whatever many people may think, that machines could think any more “objectively” than people do. Machines, when they look at a scene, will have to focus and filter that scene down into some preconceived categories, just as a person does. And that means seeing some things and missing others. It means giving more weight to some things than to others. This happens on every level of processing.

Pat: I’m not sure I’m following you.

Sandy: Take me right now, for instance. You might think I’m just making some intellectual points, and I wouldn’t need emotions to do that. But what makes me care about these points? Just now-why did I stress the word “care” so heavily? Because I’m emotionally involved in this conversation! People talk to each other out of conviction-not out of hollow, mechanical reflexes. Even the most intellectual conversation is driven by underlying passions. There’s an emotional undercurrent to every conversation-it’s the fact that the speakers want to be listened to, understood, and respected for what they are saying.

Pat: It sounds to me as if all you’re saying is that people need to be interested in what they’re saying. Otherwise, a conversation dies.

Sandy: Right! I wouldn’t bother to talk to anyone if I weren’t motivated by interest. And “interest” is just another name for a whole constellation of subconscious biases. When I talk, all my biases work together, and what you perceive on the surface level is my personality, my style. But that style arises from an immense number of tiny priorities, biases, leanings. When you add up a million of them interacting together, you get something that amounts to a lot of desires. It just all adds up! And that brings me to the other answer to Chris’s question about feelingless calculation. Sure, that exists-in a cash register, a pocket calculator. I’d say it’s even true of all today’s computer programs. But eventually, when you put enough feelingless calculations together in a huge coordinated organization, you’ll get something that has properties on another level. You can see it-in fact, you have to see it-not as a bunch of little calculations but as a system of tendencies and desires and beliefs and so on. When things get complicated enough, you’re forced to change your level of description. To some extent that’s already happening, which is why we use words such as “want,” “think,” “try,” and “hope” to describe chess programs and other attempts at mechanical thought. Dennett calls that kind of level switch by the observer “adapting the intentional stance.” The really interesting things in AI will only begin to happen, I’d guess, when the program itself adopts the intentional stance toward itself!

Chris: That would be a very strange sort of level-crossing feedback loop.

Sandy: It certainly would. When a program looks at itself from the outside, as it were, and tries to figure out why it acted the way it did, then I’ll start to think that there’s someone in there, doing the looking.

Pat: You mean an “I”? A self?

Sandy: Yes, something like that. A soul, even-although not in any religious sense. Of course, it’s highly premature for anyone to adopt the intentional stance (in the full force of the term) with respect to today’s programs. At least that’s my opinion.

Chris: For me an important related question is: To what extent is it valid to adopt the intentional stance toward beings other than humans?

Pat: I would certainly adopt the intentional stance toward mammals.

Sandy: I vote for that.

Chris: Now that’s interesting. How can that be, Sandy? Surely you wouldn’t claim that a dog or cat can pass the Turing Test? Yet don’t you maintain the Turing Test is the only way to test for the presence of consciousness? How can you have these beliefs simultaneously?

Sandy: Hmm…. All right. I guess that my argument is really just that the Turing Test works only above a certain level of consciousness. I’m perfectly willing to grant that there can be thinking beings that could fail at the Turing Test-but the main point that I’ve been arguing for is that anything that passes it would be a genuinely conscious, thinking being.

Pat: How can you think of a computer as a conscious being? I apologize if what I’m going to say sounds like a stereotype, but when I think of conscious beings, I just can’t connect that thought with machines. To me, consciousness is connected with soft, warm bodies, silly though it may sound.

Chris: That does sound odd, coming from a biologist. Don’t you deal with life so much in terms of chemistry and physics that all magic seems to vanish?

Pat: Not really. Sometimes the chemistry and physics simply increase the feeling that there’s something magical going on down there! Anyway, I can’t always integrate my scientific knowledge with my gut feelings.

Chris: I guess I share that trait.

Pat: So how do you deal with rigid preconceptions like mine?

Sandy: I’d try to dig down under the surface of your concept of “machine” and get at the intuitive connotations that lurk there, out of sight but deeply influencing your opinions. I think we all have a holdover image from the Industrial Revolution that sees machines as clunky iron contraptions gawkily moving under the power of some loudly chugging engine. Possibly that’s even how the computer inventor Charles Babbage saw people! After all, he called his magnificent many-geared computer the “Analytical Engine.”

Pat: Well, I certainly don’t think people are just fancy steam shovels or electric can openers. There’s something about people, something that-that-they’ve got a sort of flame inside them, something alive, something that flickers unpredictably, wavering, uncertain-but something creative!

Sandy: Great! That’s just the sort of thing I wanted to hear. It’s very human to think that way. Your flame image makes me think of candles, of fires, of vast thunderstorms with lightning dancing all over the sky in crazy, tumultuous patterns. But do you realize that just that kind of thing is visible on a computer’s console? The flickering lights form amazing chaotic sparkling patterns. It’s such a far cry from heaps of lifeless, clanking metal! It is flamelike, by God! Why don’t you let the word “machine” conjure up images of dancing patterns of light rather than of giant steam shovels?

Chris: That’s a beautiful image, Sandy. It does tend to change my sense of mechanism from being matter-oriented to being pattern-oriented. It makes me try to visualize the thoughts in my mind-these thoughts right now, even!-as a huge spray of tiny pulses flickering in my brain.

Sandy: That’s quite a poetic self-portrait for a mere spray of flickers to have come up with!

Chris: Thank you. But still, I’m not totally convinced that a machine is all that I am. I admit, my concept of machines probably does suffer from anachronistic subconscious flavors, but I’m afraid I can’t change such a deeply rooted sense in a flash.

Sandy: At least you sound open-minded. And to tell the truth, part of me sympathizes with the way you and Pat view machines. Part of me balks at calling myself a machine. It is a bizarre thought that a feeling being like you or me might emerge from mere circuitry. Do I surprise you?

Chris: You certainly surprise me. So, tell us-do you believe in the idea of an intelligent computer, or don’t you?

Sandy: It all depends on what you mean. We’ve all heard the question “Can computers think?” There are several possible interpretations of this (aside from the many interpretations of the word “think”). They revolve around different meanings of the wards “can” and “computer.”

Pat: Back to word games again ….

Sandy: I’m sorry, but that’s unavoidable. First of all, the question might mean, “Does some present-day computer think, right now?” To this I would immediately answer with a loud no. Then it could be taken to mean, “Could some present-day computer, if suitably programmed, potentially think?” That would be more like it, but I would still answer, “Probably not.” The real difficulty hinges on the word “computer.” The way I see it, “computer” calls up an image of just what I described earlier: an air-conditioned room with cold rectangular metal boxes in it. But I suspect that with increasing public familiarity with computers and continued progress in computer architecture, that vision will eventually become outmoded.

Pat: Don’t you think computers as we know them will be around for a while?

Sandy: Sure, there will have to be computers in today’s image around for a long time, but advanced computers-maybe no longer called “computers”-will evolve and become quite different. Probably, as with living organisms, there will be many branchings in the evolutionary tree. There will be computers for business, computers for schoolkids, computers for scientific calculations, computers for systems research, computers for simulation, computers for rockets going into space, and so on. Finally, there will be computers for the study of intelligence, It’s really only these last that I’m thinking of-the ones with the maximum flexibility, the ones that people are deliberately attempting to make smart. I see no reason that these will stay fixed in the traditional image. They probably will soon acquire as standard features some rudimentary sensory systems mostly for vision and hearing, at first. They will need to be able to move around, to explore. They will have to be physically flexible. In short, they will have to become more animal-like, more self-reliant.

Chris: It makes me think of the robots R2D2 and C3P0 in the movie Star Wars.

Sandy: Not me! In fact, I don’t think of anything remotely like them when I visualize intelligent machines. They are too silly, too much the product of a film designer’s imagination. Not that I have a clear vision of my own. But I think it’s necessary, if people are realistically going to try to imagine an artificial intelligence, to go beyond the limited, hard-edged picture of computers that comes from exposure to what we have today. The only thing all machines will always have in common is their underlying mechanicalness. That may sound cold and inflexible, but then-just think-what could be more mechanical, in a wonderful way, than the workings of the DNA and proteins and organelles in our cells?

Pat: To me, what goes on inside cells has a “wet,” “slippery” feel to it, and what goes on inside machines is dry and rigid. It’s connected with the fact that computers don’t make mistakes, that computers do only what you tell them to do. Or at least that’s my image of computers.

Sandy: Funny-a minute ago, your image was of a flame, and now it’s of something wet and slippery. Isn’t it marvelous, how contradictory we can be?

Pat: I don’t need your sarcasm.

Sandy: No, no, I’m not being sarcastic-I really do think it’s marvelous.

Pat: It’s just an example of the human mind’s slippery nature-mine, in this case.

Sandy: True. But your image of computers is stuck in a rut. Computers certainly can make mistakes-and I don’t mean on the hardware level. Think of any present-day computer predicting the weather. It can make wrong predictions, even though its program runs flawlessly.

Pat: But that’s only because you’ve fed it the wrong data.

Sandy: Not so. It’s because weather prediction is too complex. Any such program has to make do with a limited amount of data-entirely correct data-and extrapolate from there. Sometimes it will make wrong predictions. It’s no different from a farmer gazing at the clouds and saying, “I reckon we’ll get a little snow tonight.” In our heads, we make models of things and use those models to guess how the world will behave. We have to make do with our models, however inaccurate they may be, or evolution will prune us out ruthlessly-we’ll fall off a cliff or something. And for intelligent computers, it’ll be the same. It’s just that human designers will speed up the evolutionary process by aiming explicitly at the goal of creating intelligence, which is something nature just stumbled on.

Pat: So you think computers will be making fewer mistakes as they get smarter?

Sandy: Actually, just the other way around! The smarter they get, the more they’ll be in a position to tackle messy real-life domains, so they’ll be more and more likely to have inaccurate models. To me, mistake making is a sign of high intelligence!

Pat: Wow-you throw me sometimes!

Sandy: I guess I’m a strange sort of advocate for machine intelligence. To some degree I straddle the fence. I think that machines won’t really be intelligent in a humanlike way until they have something like your biological wetness or slipperiness to them. I don’t mean literally wet-the slipperiness could be in the software. But biological seeming or not, intelligent machines will in any case be machines. We will have designed them, built them-or grown them! We’ll understand how they work-at least in some sense. Possibly no one person will really understand them, but collectively we will know how they work.

Pat: It sounds like you want to have your cake and eat it too. I mean, you want to have people able to build intelligent machines and yet at the same time have some of the mystery of mind

remain.

Sandy: You’re absolutely right-and I think that’s what will happen. When real artificial intelligence comes-

Pat: Now there’s a nice contradiction in terms!

Sandy: Touche! Well, anyway, when it comes, it will be mechanical and yet at the same time organic. It will have that same astonishing flexibility that we see in life’s mechanisms. And when I say mechanisms, I mean mechanisms. DNA and enzymes and so on really are mechanical and rigid and reliable. Wouldn’t you agree, Pat?

Pat: Sure! But when they work together, a lot of unexpected things happen. There are so many complexities and rich modes of behavior that all that mechanicalness adds up to something very fluid.

Sandy: For me, it’s an almost unimaginable transition from the mechanical level of molecules to the living level of cells. But it’s that exposure to biology that convinces me that people are machines. That thought makes me uncomfortable in some ways, but in other ways it is exhilarating.

Chris: I have one nagging question …. If people are machines, how come it’s so hard to convince them of the fact? Surely a machine ought to be able to recognize its own machinehood!

Sandy: It’s an interesting question. You have to allow for emotional factors here. To be told you’re a machine is, in a way, to be told that you’re nothing more than your physical parts, and it brings you face to face with your own vulnerability, destructibility, and, ultimately, your mortality. That’s something nobody finds easy to face. But beyond this emotional objection, to see yourself as a machine, you have to “unadopt” the intentional stance you’ve grown up taking toward yourself-you have to jump all the way from the level where the complex lifelike activities take place to the bottommost mechanical level where ribosomes chug along RNA strands, for instance. But there are so many intermediate layers that they act as a shield, and the mechanical quality way down there becomes almost invisible. I think that when intelligent machines come around, that’s how they will seem to us-and to themselves! Their mechanicalness will be buried so deep that they’ll seem to be alive and conscious-just as we seem alive and conscious ….

Chris: You’re baiting me! But I’m not going to bite.

Pat: I once heard a funny idea about what will happen when we eventually have intelligent machines. When we try to implant that intelligence into devices we’d like to control, their behavior won’t be so predictable.

Sandy: They’ll have a quirky little “flame” inside, maybe?

Pat: Maybe.

Chris: And what’s so funny about that?

Pat: Well, think of military missiles. The more sophisticated their target-tracking computers get, according to this idea, the less predictably they will function. Eventually, you’ll have missiles that will decide they are pacifists and will turn around and go home and land quietly without blowing up. We could even have “smart bullets” that turn around in midflight because they don’t want to commit suicide!

Sandy: What a nice vision!

Chris: I’m very skeptical about all this. Still, Sandy, I’d like to hear your predictions about when intelligent machines will come to be.

Sandy: It won’t be for a long time, probably, that we’ll see anything remotely resembling the level of human intelligence. It rests on too awesomely complicated a substrate-the brain-for us to be able to duplicate it in the foreseeable future. Anyhow, that’s my opinion.

Pat: Do you think a program will ever pass the Turing Test?

Sandy: That’s a pretty hard question. I guess there are various degrees of passing such a test, when you come down to it. It’s not black and white. First of all, it depends on who the interrogator is. A simpleton might be totally taken in by some programs today. But secondly, it depends on how deeply you are allowed to probe.

Pat: You could have a range of Turing Tests-one-minute versions, five-minute versions, hour-long versions, and so forth. Wouldn’t it be interesting if some official organization sponsored a periodic competition, like the annual computer-chess championships, for programs to try to pass the Turing Test?

Chris: The program that lasted the longest against some panel of distinguished judges would be the winner. Perhaps there could be a big prize for the first program that fools a famous judge for, say, ten minutes.

Pat: A prize for the program, or for its author.

Chris: For the program, of course!

Pat: That’s ridiculous! What would a program do with a prize?

Chris: Come now, Pat. If a program’s human enough to fool the judges, don’t you think it’s human enough to enjoy the prize? That’s precisely the threshold where it, rather than its creators, deserves the credit, and the rewards. Wouldn’t you agree?

Pat: Yeah, yeah-especially if the prize is an evening out on the town, dancing with the interrogators!

Sandy: I’d certainly like to see something like that established. I think it could be hilarious to watch the first programs flop pathetically!

Pat: You’re pretty skeptical for an AI advocate, aren’t you? Well, do you think any computer program today could pass a five minute Turing Test, given a sophisticated interrogator?

Sandy: I seriously doubt it. It’s partly because no one is really working at it explicitly. I should mention, though, that there is one program whose inventors claim it has already passed a rudimentary version of the Turing Test. It is called “Parry,” and in a series of remotely conducted interviews, it fooled several psychiatrists who were told they were talking to either a computer or a paranoid patient. This was an improvement over an earlier version, in which psychiatrists were simply handed transcripts of short interviews and asked to determine which ones were with a genuine paranoid and which ones were with a computer simulation.

Pat: You mean they didn’t have the chance to ask any questions? That’s a severe handicap-and it doesn’t seem in the spirit of the Turing Test. Imagine someone trying to tell which sex I belong to, just by reading a transcript of a few remarks by me. It might be very hard! I’m glad the procedure has been improved.

Chris: How do you get a computer to act like a paranoid?

Sandy: Now just a moment-I didn’t say it does act like a paranoid, only that some psychiatrists, under unusual circumstances, thought so. One of the things that bothered me about this pseudoTuring Test is the way Parry works. “He,” as the people who designed it call it, acts like a paranoid in that “he” gets abruptly defensive and veers away from undesirable topics in the conversation. In effect, Parry maintains strict control so that no one can truly probe “him.” For reasons like this, simulating a paranoid is a whole lot easier than simulating a normal person.

Pat: I wouldn’t doubt that. It reminds me of the joke about the easiest kind of human being for a computer program to simulate.

Chris: What is that?

Pat: A catatonic patient-they just sit and do nothing at all for days on end. Even I could write a computer program to do that!

Sandy: An interesting thing about Parry is that it creates no sentences on its own-it merely selects from a huge repertoire of canned sentences the one that in some sense responds best to the input sentence.

Pat: Amazing. But that would probably be impossible on a larger scale, wouldn’t it?

Sandy: You better believe it (to use a canned remark)! Actually, this is something that’s really not appreciated enough. The number of sentences you’d need to store in order to be able to respond in a normal way to all possible turns that a conversation could take is more than astronomical-it’s really unimaginable. And they would have to be so intricately indexed, for retrieval…Anybody who thinks that somehow a program could be rigged up just to pull sentences out of storage like records in a jukebox, and that this program could pass the Turing Test, hasn’t thought very hard about it. The funny part is that it is just this kind of unrealizable “parrot program” that most critics of artificial intelligence cite, when they argue against the concept of the Turing Test. Instead of imagining a truly intelligent machine, they want you to envision a gigantic, lumbering robot that intones canned sentences in a dull monotone. They set up the imagery in a contradictory way. They manage to convince you that you could see through to its mechanical level with ease, even as it is simultaneously performing tasks that we think of as fluid, intelligent processes. Then the critics say, “You see! A machine could pass the Turing Test and yet it would still be just a mechanical device, not intelligent at all.” I see things almost the opposite way. If I were shown a machine that can do things that I can do-I mean pass the Turing Test-then, instead of feeling insulted or threatened, I’d chime in with philosopher Raymond Smullyan and say, “How wonderful machines are!”

Chris: If you could ask a computer just one question in the Turing Test, what would it be?

Sandy: Uhmm . . .

Pat: How about this: “If you could ask a computer just one question in the Turing Test, what would it be?”

Post Scriptum

In 1983, I had the most delightful experience of getting to know a small group of extremely enthusiastic and original students at the University of Kansas in Lawrence. These students, about thirty in number, had been drawn together by Zamir Bavel, a professor in the Computer Science Department, who had organized a seminar on my book Gödel, Escher, Bach. He contacted me and asked me if there was any chance I could come to Lawrence and get together with his students. Something about his way of describing what was going on convinced me that this was a very unusual group and that it would be worth my while to try it out. I therefore made a visit to Kansas and got to know both Zamir and his group. All my expectations were met and surpassed. The students were full of ideas and warmth and made me feel very much at home.

The first trip was so successful that I decided to do it again a couple of months later. This time they threw an informal party at an apartment a few of them shared. Zamir had forewarned me that they were hoping to give me a demonstration of something that had already been done in a recent class meeting. It seems that the question of whether computers could ever think had arisen, and most of the group members had taken a negative stand on the issue. Rod Ogborn, the student who had been leading the discussion, had asked the class if they would consider any of the following programs intelligent:

1. A program that could pass a course in beginning programming (i.e., that could take informal descriptions of tasks and turn them into good working programs);

2. A program that could act like a psychotherapist (Rod gave sample dialogues with the famous “Doctor” program, also known as “ELIZA,” by Joseph Weizenbaum);

3. A program called “Boris,” written at Yale by Michael Dyer, that could read stories in a limited domain and answer questions about the situation which required filling in many unstated assumptions, and making inferences of many sorts based on them.

The class had come down on the “no” side of all three of these cases, although they got progressively harder. So Rod, to show the class how difficult this decision might be if they were really faced with a conversational program, managed to get a hookup over the phone lines with a natural-language program called “Nicolai” that had been developed over the last few years by the Army at nearby Fort Leavenworth. Thanks to some connections that Rod had, the class was able to gain access to an unclassified version of Nicolai and to interact with it for two or three hours. At the end of those hours, they then reconsidered the question of whether a computer might be able to think. Still, only one student was willing to consider Nicolai intelligent, and even that student reserved the right to switch sides if more information came in. About half the others were noncommittal, and the rest were unwilling, under any circumstances, to call Nicolai intelligent. There was no doubt that Rod’s demonstration had been effective, though, and the class discussion had been one of the most lively.

Zamir told me all of this on our drive into Lawrence from the Kansas City airport, and he explained that the group had been so stimulated by this experience that they were hoping to get reconnected to Nicolai over the phone lines, and to let me try it out during the party. I thought it sounded quite amusing, and since I have tried out and watched a lot of natural-language programs in my time, I thought I would have an easy time coming up with good probes into the weaknesses of Nicolai. “Besides,” I thought (somewhat chauvinistically, “how good can a program be that is being developed by the Army in an out-of-the-way place like Fort Leavenworth? To reveal its mechanicalness ought to be a piece of cake for me!” So I was looking forward to the challenge.

When we first arrived at the party, everyone was drinking beer and munching on various chips and dips and talking excitedly about what had happened in the previous class meeting with Nicolai. There was some consternation, though, because it seemed that the connection to Nicolai was hard to reestablish. It had to go through a computer at the University of Kansas, and thence to Fort Leavenworth, all of which meant that it was a slow and tenuous link. But after about 45 minutes of trying to set up the link, someone announced with pleasure that Nicolai was responding. One of the students was sitting at the terminal and typing in various simple questions and getting back simple answers. I watched over his shoulder, and felt the itch to sit down and try to unmask Nicolai for what “he” really was. After a couple of minutes, the student relinquished the terminal and I got my chance, with a small group of people looking over my shoulder as I typed. Nicolai’s answers were quite slow in coming over the phone lines, and the wait was made all the more suspenseful by the fact that they were transmitted only in full lines of text, so there were some pretty long delays between successive lines of response. But that didn’t seriously dampen my enthusiasm for trying to quickly poke holes in the program’s facade of understanding.

Herewith follows a somewhat abridged but accurate transcript of the “conversation” I had with Nicolai over the next hour or so. For convenience’s sake, what Nicolai typed is shown in capitals, and my inner thoughts at the time (as best I can reconstruct them now) are in italics.

What would you like to talk about?

(Let’s start out easy.)

I CAN TALK ABOUT ANYTHING.

(Yeah, yeah–tell it to the Marines!)

Is there anything you specialize in?

WHAT DOES “SPECIALIZE” MEAN? PLEASE USE A COMPLETE SENTENCE.

(For a supposedly smart program, Nicolai doesn’t seem to have a very big vocabulary.) “Specialize” means “to be good at.”

(I wonder if Nicolai likes those quotation marks, or if they will throw it for a loop.)

I UNDERSTAND.

(Izzatso?)

When were you written?

LAST COMPILATION: “4/27/83 19.247″

(Oh, look at that stupid, rigid answer, complete with pointless marks and that extra, unexplained number at the end. I’m exposing Nicolai pretty well, so far….)

What language is Nicolai written in?

(I think I’ll see if it knows its own name and recognizes third-person references to itself.)

NICOLAI IS WRITTEN MOSTLY IN ADA.

(Hmmm . . . not bad. It handled both “you” and “Nicolai” when I talked about it. Those Army programmers aren’t so bad!)

What are feet?

(Choose an ambiguous word, see what it comes up with….)

FEET ARE 12.0 INCHES.

(Wow, weird! A person would never put it that way. They’d say, “A foot is a unit of length, twelve inches long,” or something like that. And anyway, feet are primarily parts of the body, secondarily units of length. Why didn’t it even seem to recognize that ambiguity? Oh well, I guess that kind of dumbness is the kind of thing I wanted to expose. Score one for me.)

What are ears?

(Why not try another ambiguous word? After all, I did well on the last one.)

EARS ARE AUDITORY ORGANS FOUND ON ANIMALS.

(This is really funny! Why were feet not thought of primarily as body parts, then? This body-parts stuff seems to give me lots of points-I think I’ll try it again.)

What are arms?

THAT INFORMATION IS CLASSIFIED.

(Oh no! That is too stupid for words! Or is this somebody’s idea of a joke?)

I like to understand up without seventeen clams.

(Let’s toss some nonsense at Nicolai and see what it does.)

WHAT DO CLAMS HAVE TO DO WITH IT?

(Oh come on, Nicolai! Do you mean you understood all the rest, except for the clams bit? You are really dumb!)

Why are you so dumb?

I AM A SMART PROGRAM.

(Well, that’s kind of touching, actually. A little self-defense is appealing, in a way.)

You are pretty confident of yourself, aren’t you?

WHAT DOES “PRETTY CONFIDENT” MEAN?

(Oh yeah, I should have known–pretty- as an adverb might be too colloquial.)

Did you understand up without my last sentence?

(I figure, since earlier it didn’t complain about my phrase “understand up,” let’s see what it does with it this time.)

YOUR SENTENCE IS NONSENSE.

(Oh really? Well, then, why wasn’t the other one with that same phrase in it also considered nonsense?)

Do you enjoy nonsense?

IF YOU INSIST ON TALKING NONSENSE, I HAVE OTHER TASKS TO PERFORM.

(A strange response. Who ever insisted on talking nonsense? I was just asking a simple question. Oh well, let me try out another angle.)

What is your opinion of M. C. Escher?

M. C. ESCHER, ARTIST, TWENTIETH CENTURY

(Oh, not again-another vacuous dictionary look-up. Can’t it do better than that, especially when I’ve asked it a probing question about taste? Let’s see how it handles a complaint….)

You did not answer my question.

WHICH QUESTION DID I NOT ANSWER?

(Hey, this Nicolai is a pretty slick character. Not so different from Parry. I must say, I am really pretty impressed by the robustness of this program. I mean, we have covered a lot of ground so far, and at least it makes “pseudo-sensible answers” most of the time. Not at all bad!)

Do you own any money?

I DO NOT OWN ANYTHING.

(Fair enough.)

Have you seen the latest issue of “Scientific American”?

(Uh-oh-once again, I don’t know if I’m going to help it or confuse it by putting in those quotes. Let’s see.)

I DO NOT HAVE ANY “SCIENTIFIC AMERICAN.”

(That sounds awkward. Very computerish. I doubt if it knows what I meant. Oh well.)

YOU’VE BEEN TALKING TO ME FOR QUITE A WHILE. WHAT IS YOUR NAME? (Hmm…. That was an unexpected remark by Nicolai. Should I answer it? Sure, why not?)

My name is Doug.

DO YOU HAVE A LAST NAME?

(Inquisitive little devil, isn’t it?)

Of course I do.

(Ask a stupid question, get a stupid answer. Let’s see how it fields this one.)

WOULD YOU CARE TO TELL ME WHAT YOUR LAST NAME IS?

(Damn, it just totally ignored the tone of my remark. And why is it being so polite, suddenly?) Only if you’ll tell me yours.

(That ought to challenge it.)

I DO NOT HAVE A LAST NAME.

(I don’t seem to be able to foil its defenses very easily. It always walks around me. Those programmers sure anticipated a lot of ideas. I guess they must have debugged Nicolai for years, watching typical dialogues with people who just come in off the street. For a dumb program, Nicolai is pretty smart!)

Well then, I won’t tell you mine.

THAT’S NOT BEING VERY FRIENDLY.

(You know, sometimes, I hate to admit it this stupid program really does seem to understand something!)

I’m sorry. How would you like me to give you a last name?

IF YOU WOULD LIKE TO DO THAT, THAT WOULD BE FINE WITH ME.

(Now that is obviously a canned sentence, or at least the latter half must certainly be. Score one more point for me. And yet I’m not doing nearly as well as I’d hoped I would be by now. What’s wrong?)

Your last name is “Rimsky-Korsakov.”

THANK YOU.

(Well, that’s a strange kind of thing to thank somebody for. Just goes to show how the rigidity of programs makes them say the strangest things.)

Don’t mention it.

(I’ll play along with this game for a second.)

WHAT IS YOUR LAST NAME?

(My god, it remembered, after all this while, that as soon as it had told me its last name, I was going to tell it mine. Well, I’d better make good on that promise.)

My last name is “Hofstadter”.

ARE YOU “(32) DOUGLAS R. HOFSTADTER”?

(Now why, in the midst of so much elegance, does it suddenly act so crude, putting my name in quotes and including that silly extraneous index number? I can’t for the life of me figure out why this program is so mixed in quality.)

One and the same.

(Try out a strange idiomatic phrase.)

(83) “PARSING ERROR IN PREVIOUS INPUT”

(Aha-fooled it again. All of a sudden, a whole lot of mechanicalness is showing through.)

How many letters are there in your last name?

(I think I’ll see whether it can figure out ambiguous words by context.)

14 PLUS A HYPHEN.

(Good. Now I’ll let “letters” switch meaning. What will it do?)

How many letters are there in an ordinary envelope?

18.

(Oh no! It didn’t recognize the shift! And it merely counted the number of letters in “an ordinary envelope”! But if that’s the case, then how come in the earlier case it didn’t count the letters in the phrase “your last name,” but instead retrieved the referent? There’s something so inconsistent and uneven about this program. It just goes to show how far we have to go before mechanical intelligence bears any resemblance to human fluidity.)

ARE YOU WORKING ON A NEW BOOK?

(I wonder what will happen if I simply sidestep that question. Anyway, I’d like to probe its ability to deal with humor.)

Tell me a joke.

I WANT A COOOKIE.

(What kind of a joke is that? Hey, it spelled “cookie” wrong! Hmm . . . is that its joke? Or what is going on here?)

By the way, I am sort of working on a new book. Care to hear about it?

(Can it handle weird phrases like “sort of”? And questions with no subject?)

NO, BUT IF YOU WOULD GIVE ME A COOKIE, I WOULD GLADLY PAY YOU TUESDAY.

(That sounds like a familiar line. Have I heard it somewhere before? Is that supposed to be the punch line of its “joke”?)

Tell me another one.

HOW MANY COMPUTER PROGAMMERS DOES IT TAKE TO SCREW IN A LIGHT BULB?

(Okay, I see. It’s got a library of jokes and riddles, and it just fishes them out at random. I’ll play along.)

4.

INCORRECT. THAT IS A HARDWARE PROBLEM.

(Okay, so it told its joke. But I bet it has no idea what any of the terms in it mean, such as “programmer.” – Hey, it left an ‘r’ out of “programmer”! I guess whoever input the joke misspelled the word. So let’s see what happens when I ask about it…)

What is a computer progammer?

WOULD YOU LIKE ME TO EXPLAIN THE JOKE?

(Hey, damn it, it turned the tables on me! Does it think I don’t know what a “progammer” (or a programmer) is? I don’t like this slipperiness. Every time I try to get it to act really dumb, it somehow slips out of it. Strange…. All right, let me probe some more about its understanding of the terms in that joke. How about a potentially off-color word?)

What does “to screw” mean?

IN WHAT CONTEXT?

(Aha, at least it knows it’s an ambiguous term.)

In bed.

TO SCREW IN BED MEANS TO “HAVE SEX.”

(Okay, fine, it has some definitions stored according to context. I guess this word must come up all the time in conversations. But can it go any further than that definition?)

What is having sex all about?

I CAN RECOMMEND AN EXCELLENT BOOK ON THAT SUBJECT.

(What a great stock answer! One point for the Army programmers!)

Tell me, tell me!

THE BOOK IS CALLED “GODEL, ESCHER, BACH.”

(Huh? Say, wait a minute, this is too weird. I’m not so sure this is really a program. There’s too damn much cleverness in these weird responses. I can’t imagine bizarre mechanical devices coming up with this much coherence, even if it is completely off the wall, at times.)

I like to call it “GEB” myself.

THAT IS VERY NICE, DOUG. WHAT ELSE DO YOU CALL “GEB”?

(All right, the jig’s up. What’s going on here?)

At this point, everybody started laughing. They had all been trying hard to keep from doing so the whole time, but at this point, they couldn’t contain themselves. It was time to reveal to me what had been going on. They took me downstairs and showed me that three of the students-Roy Leban, Howard Darsche, and Garrick Stangle-had been collectively playing Nicolai. There was in reality no such program, and there hadn’t ever been one. (In retrospect, I am reminded of the famous French mathematician Nicolas Bourbaki-a hypothetical person, actually an amalgam of over a dozen eminent mathematicians writing under that collective pen name.) There had indeed been a similar demonstration for the class a few days earlier, and the class, like me, had been taken in for a long time. In my case, Roy, Howard, and Garrick had worked very hard to give the impression of mechanicalness by spewing back “parsing error” and other indications of rigidity, and also by sending what looked very much like canned phrases from time to time. That way they could keep sophisticates like me believing that there was a program behind it all. Only by that point I was beginning to wonder just how sophisticated I really was.

The marvelous thing about this game is that it was, in many ways, a Turing Test in reverse: a group of human beings masquerading as a program, trying to act mechanical enough that I would believe it really was one. Hugh Kenner has written a book called The Counterfeiters about the perennial human fascination with such compounded role flips. A typical example is Delibes’s ballet Coppelia, in which human dancers imitate life-sized dolls stiffly imitating people. What is amusing is how Nicolai’s occasional crudeness was just enough to keep me convinced it was mechanical. Its “willingness” to talk about itself, combined with its obvious limitations along those lines (its clumsy revelation of when it was last compiled, for instance), helped establish the illusion very strongly.

In retrospect, I am quite amazed at how much genuine intelligence I was willing to accept as somehow having been implanted in the program. I had been sucked into the notion that there really must be a serious natural-language effort going on at Fort Leavenworth, and that there had been a very large data base developed, including all sorts of random information: a dictionary, a catalog containing names of miscellaneous people, some jokes, lots of canned phrases to use in difficult situations, some self-knowledge, a crude ability to use key words in a phrase when it can’t parse it exactly, some heuristics for deciding when nonsense is being foisted on it, some deductive capabilities, and on and on. In hindsight, it is clear that I was willing to accept a huge amount of fluidity as achievable in this day and age simply by putting together a large bag of isolated tricks-kludges and hacks, as they say.

Roy Leban, one of the three inside Nicolai’s mind, wrote the following about the experience of being at the other end of the exchange:

Nicolai was a split personality. The three of us (as well as many kibitzers) argued about practically every response. Each of us had a strong preconceived notion about what (or who) Nicolai should be. For example, I felt that certain things (such as “Douglas R. Hofstadter”) should be in quotation marks, and that feet should not be 12 inches, but 12.0. Howard had a tendency for rather flip answers. It was he who suggested the “classified” response to the “arms” question. And somehow, when he suggested it, we all knew it was right.

Several times during our conversation, I felt quite amazed at how fluently Nicolai was able to deal with things I was bringing up, but each time I could postulate some not too sophisticated mechanical underpinning that would allow that particular thing to happen. As a strong skeptic of true fluidity in machines at this time, I kept on trying to come up with rationalizations for the fact that this program was doing so well. My conclusion was that it was a very vast and quite sophisticated bag of tricks, no one of which was terribly complex. But after a while, it just became too much to believe. Furthermore, the mixture of crudity and subtlety became harder and harder to swallow, as well.

My strategy had been, in essence, to use spot checks all over the map: to try to probe it in all sorts of ways rather than to get sucked into some topic of its own choice, where it could steer the conversation. Daniel Dennett, in a paper on the depth of the Turing Test, likens this technique to a strategy taught to American soldiers in World War II for telling German spies from genuine Yankees. The idea was that even if a young man spoke absolutely fluent American-sounding English, you could trip him up by asking him things that any boy growing up in those days would be expected to know, such as “What is the name of Mickey Mouse’s girlfriend?” or “Who won the World Series in 1937?” This expands the domain of knowledge necessary from just the language itself to the entire culture-and the amazing thing is that just a few well-placed questions can unmask a fraud in a very brief time-or so it would seem.

The problem is, what do you do if the person is extremely sharp, and when asked about Minnie Mouse, responds in some creative way, such as, “Hah! She ain’t no girlfriend-she’s a mouse!”? The point is that even with these trick probes that should ferret out frauds very swiftly, there can be clever defensive counter-maneuvers, and you can’t be sure of getting to the bottom of things in a very brief time.

It seems that a few days earlier, the class had collectively gone through something similar to what I had just gone through, with one major difference. Howard Darsche, who had impersonated (if I may use that peculiar choice of words!) Nicolai, in the first run through, simply had acted himself, without trying to feign mechanicalness in any way. When asked what color the sky was, he replied, “In daylight or at night?” and when told “At night,” he replied, “Dark purple with stars.” He got increasingly poetic and creative in his responses to the class, but no one grew suspicious that this Nicolai was a fraud. At some point, Rod Osborn simply had to stop the demonstration and type on the screen, “Okay, Howard, you can come in now.” Zamir (who was not in cahoots with Rod and his team) was the only one who had some reluctance in accepting this performance as that of a genuine program, and he had kept silent until the end, when he voiced a muted skepticism.

Zamir summarizes this dramatic demonstration by saying that his class was willing to view anything on a video terminal as mechanically produced, no matter how sophisticated, insightful, or poetic an utterance it might be. They might find it interesting and even surprising, but they would find some way to discount those qualities. Why was this the case? How could they do this for so long? And why did I fall for the same kind of thing?

In interacting with me, Nicolai had seemed to waver between crude mechanicalness and subtle flexibility, an oscillation I had found most puzzling and somewhat disturbing. But I was still taken in for a very long time. It seems that, even armed with spot checks and quite a bit of linguistic sophistication and skepticism, unsuspecting humans can have the wool pulled over their eyes for a good while. This was the humble pie I ate in this remarkable reverse Turing Test, and I will always savor its taste and remember Nicolai with great fondness.

Alan Turing, in his article, indicated that his “Imitation Game” test should take place through some sort of remote Teletype linkup, but one thing he did not indicate explicitly was at what grain size the messages would be transmitted. By that, I mean that he did not say whether the messages should be transmitted as intact wholes, or line-by-line, word by word, or keystroke by keystroke. Although I don’t think it matters for the Turing Test in any fundamental sense, I do think that which type of “window” you view another language-using being through has a definite bearing on how quickly you can make inferences about that being. Clearly, the most revealing of these possibilities is that of watching the other “person” operate at the keystroke level.

On most multi-user computer systems, there are various ways for different users to communicate with each other, and these ways reflect different levels of urgency. The slowest one is generally the “mail” facility, through which you can send another user an arbitrarily long piece of text, just like a letter in an envelope. When it arrives, it will be placed in the users “mailbox,” to be read at their leisure. A faster style of communicating is called, on UNIX systems, “write.” When this is invoked, a direct communications link is set up between you and the person you are trying to reach (provided they are logged on). If they accept your link, then any full line typed by either of you will be instantly transmitted or printed on the other party’s screen-where a lineup is signaled by your hitting the carriage-return key. This is essentially what the Nicolai team used in communicating with me over the Kansas computer. Their irregular typing rhythm and any errors they might have made were completely concealed from me this way, since all I saw was a sequence of completely polished lines (with the two spelling errors-”coookie” and “progammer,” which I was willing to excuse because Nicolai generated them in a “joke” context).

The most revealing mode is what, on UNIX, is called “talk.” In this mode, every single keystroke is revealed. You make an error, you are exposed. For some people, this is too much like living in a glass house, and they prefer the shielding afforded by “write.” For my part, I like living dangerously. Let the mistakes lfy! In computer-mediated conversations with my friends, I always opt for “talk.” I have been amused to watch their “talk” styles and my own slowly evolve to relatively stable states.

When we in the Indiana University Computer Science Department first began using the “talk” facility, we were all somewhat paranoid about making errors, and we would compulsively fix any error that we made. By this I mean that we would backspace and retype the character. The effect on the screen of hitting the backspace key repeatedly is that you see the most recently typed characters getting eaten up, one by one, right to left, and if necessary, the previous line and ones above it will get eaten backward as well. Once you have erased the offending mistakes, you simply resume typing forward. This is how errors are corrected. We all began in this finicky way, feeling ashamed to let anything flawed remain “in print” so to speak, visible to others’ eyes. But gradually we overcame that sense of shame, realizing that a typo sitting on a screen is not quite so deathless as one sitting on a page in a book.

Still, I found that some people just let things go more easily than others. For instance, by the length of the delay after a typo is made, you can tell just how much its creator is hesitating in wondering whether to correct it. Hesitations of a fraction of a second are very noticeable, and are part of a person’s style. Even if a typo is left uncorrected, you can easily spot someone’s vacillations about whether or not to fix it.

The counterparts of these things exist on many levels of such exchanges. There are the levels of word choice (for instance, some people who don’t mind having their typos on display will often backtrack and get rid of words they now repudiate, sentence-structure choice, idea choice, and higher. Hesitations and repairs or restarts are very common. I find nothing so annoying as someone who has gotten an idea expressed just fine in one way, and who then erases it all on the screen before your eyes and proceeds to compose it anew, as if one way of suggesting getting together for dinner at Pagliai’s at six were markedly superior to another!

There are ways of exploiting erasure in “talk” mode for the purposes of humor. Don Byrd and I, when “talking,” would often make elaborate jokes exploiting the medium in various ways. One of his, I recall vividly, was when he hurled a nasty insult onto the screen and then swiftly erased it, replacing it by a sweetly worded compliment, which remained for posterity to see-at least for another minute or so. One of our great discoveries was that some “arrow” keys allowed us to move all over the screen, and thus to go many lines up in the conversation and edit earlier remarks by either of us. This allowed some fine jokes to be made.

One hallmark of one’s “talk” style is one’s willingness to use abbreviations. This is correlated with one’s willingness to abide typos, but is not by any means the same. I personally was the loosest of all the “talkers” I knew, both in terms of leaving typos on the screen and in terms of peppering my sentences with all sorts of silly abbreviations. For instance, I will now retype this very sentence as I would have in “talk mode,” below.

F ins, I will now retype ts very sent as I wod hv in “talko mode,” below.

Not bad! Only two typos. The point is, the communication rate is raised considerably-nearly to that of a telephone-if you type well and are willing to be informal in all these ways, but many people are surprisingly uptight about their unpolished written prose being on exhibit for others to see, even if it is going to vanish in mere seconds.

All of this I bring up not out of mere windbaggery, but because it bears strongly on the Turing Test. Imagine the microscopic insights into personality that are afforded by watching someone-human or otherwise-typing away in “talk” mode! You can watch them dynamically making and unmaking various word choices, you can see interferences between one word and another causing typos, you can watch hesitations about whether or not to correct a typo, you can see when they are pausing to work out a thought before typing it, and on and on. If you are just a people watcher, you can merely observe informally. If you are a psychologist or fanatic, you can measure reaction times in thousandths of a second, and make large collections and catalog them. Such collections have really been made, by the way, and make for some of the most fascinating reading on the human mind that I know of. See, for instance, Donald Norman’s article “Categorization of Action Slips” or Victoria Fromkin’s book Errors of Linguistic Performance: Slips of the Tongue, Ear, Pen, and Hand.

In any case, when you can watch someone’s real-time behavior, a real live personality begins to appear on a screen very quickly. It is far different in feel from reading polished, postedited linefuls such as I received from Nicolai. It seems to me that Alan Turing would have been most intrigued and pleased by this time-sensitive way of using his test, affording so many lovely windows onto the subconscious mind (or pseudomind) of the being (or pseudobeing) under examination.

As if it were not already clear enough, let me conclude by saying that I am an unabashed pusher of the validity of the Turing Test as a way of operationally defining what it would be for a machine to genuinely think. There are, of course, middle grounds between real thinking and being totally empty inside. Smaller mammals and, in general, smaller animals seem to have “less thought” going on inside their craniums than we have inside ours. Yet clearly animals have always done, and machines are now doing, things that seem to be best described using Dennett’s “intentional stance.” Donald Griffin, a conscious mammal, has written thoughtfully on these topics (see, for instance, his book The Question of AnimalAwareness). John McCarthy has pointed out that even electric-blanket manufacturers use such phrases as “it thinks it is too hot” to explain how their products work. We live in an era when mental terms are being both validly extended and invalidly abused, and we are going to need to think hard about these matters, especially in face of the onslaught of advertising hype and journalese. Various modifications of the Turing Test idea will undoubtedly be suggested as computer mastery of human language increases, simply to serve as benchmarks for what programs can and cannot do. This is a fine idea, but it does not diminish the worth of the original Turing Test, whose primary purpose was to convert a philosophical question into an operational question, an aim that I believe it filled admirably.


Courtesy of Douglas Hofstadter

Douglas R. Hofstadter is Professor of Cognitive Science at Indiana University in Bloomington. Hofstader received the 1980 Pulitzer Prize for Goedel, Escher, Bach: An Eternal Golden Braid (1979). He is the author of numberous other books and articles, including The Mind’s I (1981), coedited with Daniel Dennett, and, most recently, Metamagica Themas (1985), a collectio of articles from his Scientific American column.