Public perception of artificial intelligence (AI)

Sometimes a work of science fiction tells more about the time of its creation then it does the future it purports to predict. In the Stanley Kubrick/Arthur C. Clarke 1968 epic film 2001: A Space Odyssey, about the central character -- the HAL 9000 computer -- talks amiably, renders aesthetic judgments of drawings, recognizes the emotions in the crew, but also murders four of the five astronauts in a fit of paranoia and concern for the mission. At the time of the filming -- before anyone had a Ph.D. in computer science, before the PC and Macintosh, before duogenarians started buying Ferraris from the IPOs of their software companies -- the general public knew little about computers and had virtually no direct experience with them. As such, the film's compelling and carefully considered representation of HAL and his abilities embodied almost as much hope and fear as it did knowledge and analysis.

Shortly after we meet HAL, he plays chess against astronaut Frank Poole, and this scene tells us a great deal about computers and 1960s society's view toward them. First, it is significant what is not shown. Kubrick originally filmed the scene with Dave playing a new game, "Pentominoes," then being promoted by the Milton-Bradley Game Corporation. Kubrick rejected this because although Pentominoes might have gone on to popularity, film goers wouldn't quite know what the astronaut was doing (programming or controlling some aspect of the ship perhaps?). Even if they did recognize it as a game, viewers wouldn't know just how difficult it was and thus how impressive HAL's inevitable win would be. Kubrick chose chess in large part to show how "intelligent" HAL was; chess has long been held as a paradigm of the heights of human logic and reasoning. (It should be pointed out that Kubrick is an avid chess player and as a teenager hustled chess in the parks of Brooklyn during the 1950s. Moreover, in the novel 2001, HAL is programmed to lose 50% of the time – to keep things interesting for the astronauts.)

Next, consider the particular sequence of moves Kubrick shows. These are moves 13 through 15 from an obscure game between two German masters played in Hamburg in 1913 (the Ruy Lopez or Spanish opening). We can presume that Kubrick chose this set of moves because of the cleverness of the checkmate – clever enough that an astronaut might not see it, yet short and easy enough for chess-literate viewers to recognize and admire.

Recall, too, Frank's reaction to his loss -- or more specifically his lack of reaction. He clearly accepts defeat without anguish. His pause is brief and he doesn't even take time to confirm the mate carefully -- he knows HAL is correct in his announcement of the checkmate. But Frank's lack of reaction is very significant. Although it may seem a bit quaint in the late 1990s, at the time of the filming of 2001, in the public's mind there was a clear pro-human/anti-computer sentiment, at least as far as chess was concerned. Even in the human-machine tournaments in the '70s and early '80s, audiences rooted for the human and against the machine. Nowadays, few of us feel deeply threatened by a computer beating a world chess champion -- any more than we do at a motorcycle beating an Olympic sprinter. It is true, chess masters will be the most distraught (Kasparov has called his last tournament "species defining"), and true, it will garner great public interest. Nevertheless, among scientists working in the field it may be important, but for other reasons I'll explain below.

In short, in 2001 Kubrick and Clarke were absolutely right to predict that

- we would play games with computers for diversion (this was not obvious in the 1960s, incidentally, before Pong, Nintendo and Sega Genesis)
- computers would become superb chess players, surely able to beat an astronaut.

Let me digress and take the Kasparov/Deep Blue rematch as a good chance to dispel a myth that has grown up around 2001. The story has it that the name HAL was chosen because each letter is just one step ahead of IBM. However, this is pure coincidence, and in fact the first incarnations of the computer had a woman's voice and was called "Athena," goddess of wisdom. When someone pointed out the spurious association between HAL and IBM, Kubrick wanted to change the computer's name and refilm the scenes but was dissuaded because of production costs. (For the record, HAL comes from "Heuristically programmed ALgorithmic" computer.)

But back to HAL: Yes, HAL was brilliant and amiable but, at least in the public's mind, capable of evil. His triumph over Frank in chess presages the murder in outer space. HAL surely was "intelligent," and his prowess at chess helps to convince us of that.

But how do we "measure" or test his intelligence?

The chess Turing test

The question "Is a machine intelligent?" is a notoriously difficult one and thus in 1950 the computer science pioneer Alan Turing proposed his famous test: There are two keyboards in front of you, one connected to a computer, the other leads to a person. You type in questions on any topic you like; both the computer and the human type back responses that you read on the respective computer screens. If you cannot reliably determine which was the person and which the machine, then we say the machine has passed the Turing test. HAL, of course, passed with flying colors. No computer can pass such an unrestricted Turing test today – or will for quite some time. For this reason, we restrict the test, to give the computer a fighting chance.

One such restricted test is a chess Turing test: You play chess against an opponent whose moves are presented by means of a computer screen, and you try to determine whether your opponent is a computer or a human. Chess Turing tests have been conducted, but with a slight twist. You get to see only the recorded moves of a previous game and must state if either opponent -- or possibly both or possibly neither -- was a computer. This is a just a bit tougher than a true chess Turing test because now you're seeing a game played by others, one where the opponents were trying to win. Consider: If you were playing and your goal was to determine whether the opponent is a computer, you might be clever and lose the game in a particularly interesting way that might reveal the identity of your opponent.

In such modified chess Turing tests (based on recorded games), just as in the original Turing test, the better the computer, the harder it is to distinguish between human and machine opponents. In an informal experiment, Kasparov could occasionally, but not reliably, guess from recorded games whether opponents were human or machine. We now know that we can make computers excel on limited problems, such as chess, or controlling an oil refinery, or even complicated medical diagnoses, but we are very far indeed from creating a computer to pass an unrestricted Turing test. But more on that below.

Should we mimic humans to achieve AI?

Should we try to make artificial intelligence by duplicating how humans do it, or instead try to exploit the particular strengths of machines? Humans are slow but exquisitely good at pattern recognition and strategy; computers, on the other hand, are extremely fast and have superb memories but are annoyingly poor at pattern recognition and complex strategy. Kasparov can make roughly two moves per second; Deep Blue has special-purpose hardware that enables it to calculate nearly a quarter of a billion chess positions per second.

Here is an illustration of the difference, taken from chess: Controlled psychological experiments have shown that human chess masters are far more accurate than non-chess players at remembering chess board positions taken from real games, where the placement of pieces arose in strategic play and represented meaningful tactical positions. However, these masters were no better than non-chess players at memorizing random arrangements of pieces. Chess masters remember positions based on certain patterns, alignments and structure whereas, of course, computers have no difficulty remembering -- storing -- all the games or random arrangements ever made and need no "meaning" in the placements.

There are other differences too, of course. Humans have emotions --- they have pride at winning, shame at a bad loss, satisfaction when extracting revenge; not so computers (well, not yet). Computers don't get tired, and don't have "bad" days -- at least so long as the hardware doesn't break down!

Early chess systems sought to duplicate or mimic the methods of humans. But this proved to be far too difficult: What precisely suggests any particular move? Instead, successful chess programs capitalize on the particular strengths of computers: rapid and massive parallel search. This is quantified by how many moves ahead, or "plies," the computer can search in a given time. If I move, that's one ply; if you then also move, that's two plies, and so forth. Naturally, the deeper the computer can search, the stronger a player it is. The interesting thing about computer chess is the extremely good correlation between the average depth of search (measured in plies) and the strength in chess, quantified by the rating.

Humans have an uncanny ability to see or "sniff out" sequences of moves that are likely to pay off down the line -- an example of sophisticated pattern recognition -- for instance, as if to say "Gosh, I'm not sure why, but I think that attacking his king's knight's pawn with my queen's bishop looks promising... let me explore that line of attack..." As such, human grandmasters don't waste time on unpromising sequences but tend to look far down the promising avenues and see traps beyond the search "horizon" of a brute force machine search. This, in fact, is one of the strategies employed by grandmasters, including Kasparov, when playing against computers. Well it turns out that for the 1913 Hamburg game used in the scene in 2001, the earliest moves (which would have occurred before the film's scene) were indeed quite "trappy" in this way, and assuming HAL played them, we would say that he would pass the chess Turing test --- not surprising, since the game was actually from a human match.

Another difference between human and computer chess play involves learning or adaptability. Machines play the same way again and again, and given a particular setup will always play the same move. During a game, however, a human grandmaster might notice that his opponent is aggressive or conservative or risky or trappy, and change his own style of play accordingly. Humans even do this from game to game in a tournament against a single opponent. Garry Kasparov stated quite clearly that his tournament victory over Deep Blue in February 1996 came because he could analyze the first game and play to exploit Deep Blue's evident weaknesses in the next. Conversely, Kasparov changed his own style of play in the middle of games, and Deep Blue was not prepared to adapt accordingly.

The ability of computers to do extensive search has had ramifications on endgame play, where only a few pieces are left on the board, such as a king, bishop and knight versus a king and a pawn. Certain endgame arrangements were always thought to represent a draw -- no human had ever seen a way to win. Nevertheless, through deep -- very deep -- computer searches, some of these positions were proven to be a forced win. In one setup, a Connection Machine supercomputer found a forced checkmate in an astounding 249 moves! There is a scientific paper on the subject in the original German, which said playing against a computer programmed with these endgame sequences was "Schachspielen wie gegen Gott"-- chess playing as if against God. Garry Kasparov himself said it best: "Sometimes quantity becomes quality."

Once some of these endgames were shown to be wins, a few tournament players tried to memorize the necessary sequence of moves. This leads to a fascinating point: Whereas researchers began by trying to make computer chess systems imitate the style of humans, paradoxically it turns out that some humans play their endgames by imitating computers! Thus, at least in some aspects of endgame play, machines are clearly superior to humans.

It turns out that the Japanese board game "go" will not succumb to such brute force methods -- there are simply too many possible moves for even the most powerful supercomputer imagined to ever examine. Instead, "real" AI will be needed -- intelligence based on pattern recognition, "insight" and strategy. Indeed, for those of us who work in pattern recognition, machine learning, or various fields allied with artificial intelligence, it is the weaknesses of Deep Blue that are the most interesting. How should we program computers to recognize and understand the style of their opponent's play and adapt accordingly? How should we program the machine to distinguish the most promising lines of attack from the ones that are not likely to pay off? How do we program machines to make complex plans? Whereas there are some aspects of the Deep Blue system that employ crude versions of methods we know are important in human intelligence (in particular when scoring the quality of a position), their weaknesses are compensated by the brute force search through possible moves.

It must be emphasized, too, that even if such subtle and complicated techniques of pattern recognition, reasoning and so forth are ultimately achieved in chess, there would still remain an enormous gulf between their use in chess and in other general aspects of intelligence, for instance, in planning a story or recognizing a scene. For these, we may have to duplicate the human, at least at some level of abstraction.

A new era?

I am assuming that computers will ultimately triumph over humans in the domain of chess, if not at this Deep Blue/Kasparov tournament, then in the not-too-distant future. True, technologists tend to be optimists, and the predictions of a number of computer scientists vis-a-vis chess -- from Alan Turing to Marvin Minsky to Raymond Kurzweil -- have been overly optimistic. Furthermore, humans will continue to improve – surely Kasparov is improving, in part from his competition with computers. Nevertheless the trends are clear enough, and although I will not hazard a guess as to when it will occur, I am confident that someday a computer will reign supreme in chess.

It has been said that when computers become world champions, we will either:

- think more of computers
- think less of humans or
- think less of the game of chess.

My view is that we will think just a bit more of computers (at least for these and related problems) and still admire the game of chess. I think we will -- or at least we should -- think more of humans, not less. We will appreciate just how difficult problems like pattern recognition and planning and creativity are, and how poorly scientists and technologists have done in trying to reproduce these human behaviors.

The public should understand one of the central lessons of the last 40 years in AI research: that problems we thought were hard turned out to be fairly easy, and that problems we thought were easy have turned out to be profoundly difficult. Chess is far easier than innumerable tasks performed by an infant, such as understanding a simple story, recognizing objects and their relationships, understanding speech, and so forth. For these and nearly all realistic AI problems, the brute force methods in Deep Blue are hopelessly inadequate.

With the (presumed) forthcoming "solution" to the chess problem, I think we will come to the end of an era – the era of the "quick kill" where hardware and brute force solve interesting problems. Chess is the last problem in traditional AI that will garner great public attention and be solvable by "simple" brute force techniques. Natural language understanding, scene analysis, speech recognition, and much more will require much more work and work of a different sort, than that used for chess. We'll be at the end of one road that leads just part of the way through the forest.

Thus we are led to ask, after chess, "whither AI"?

I think there will be a sober reconsideration and broad acceptance of the magnitude of the AI problem, and a realization that the techniques that proved successful in chess will be of only limited use in the domains where humans currently dominate and which we view as essential for AI. In addition to continued research progress, what happens then will depend in part upon whether scientists and engineers can make a coherent case that we know enough about the foundations of AI and that we are in a position for larger scale projects. What might the research projects be like?

Doug Lenat's CYC project -- a several-decades-long knowledge engineering mission for entering common sense and general knowledge into machines -- may give the flavor of the future. Although it is too early to judge the eventual success or failure of his particular system, which is still years from coming fully on-line, we can note some of its attributes that presage things to come. First, his system addresses general intelligence, rather than just a specific domain such as chess. For example, his system could be used for interpreting handwriting (by providing knowledge of reasonable alternative readings of ambiguous words) or searching labeled images (by inferring related terms and concepts that are not in a thesaurus) or spotting inconsistencies in a database. Second, Lenat's program isn't concerned with proving arcane theorems that may be of limited use, but instead relies on lots of repetitive (and I would imagine boring) knowledge engineering -- the digital age's answer to sweat shops. Third, it acknowledges the magnitude of the problem -- it already embodies over a person-century of data entry during the last dozen years, and still has much to do.

The CYC project, as large as it may be, is still peanuts on the scale of big science in other disciplines. The physicists have their multibillion-dollar particle accelerators, the astronomers their space missions, and microbiologists their human genome project, but there is no equivalent in computer science and AI, at least not in the U.S. We can imagine enormous software projects for learning simple objects or animal shapes, which would be useful in searching the World-Wide Web, and work on integrating and mediating large numbers of experts on subproblems.

Conclusion

The problems addressed by AI are some of the most profound in all of science: How we know the world? What is the "essense" of an object or pattern? How do we remember the past? How do we create new ideas? For centuries, mankind had noticed hearts in slaughtered animals; nevertheless, the heart’s true function was a mystery until one could liken it to an artifact and conclude: a heart is like a pump. (Similarly, an eye is like a camera obscura, a nerve is like an electric wire....) In the same way, we have known for centuries the brain that is responsible for thoughts and feelings, but we'll only truly understand the brain when our psychological and neurological knowledge is complemented by an artifact -- a computer that behaves like a brain. As such, AI is in the long tradition of philosophy and epistemology; it is surely worthy of our support as a culture. (It also will have immense practical benefits.)

We should take the eventual triumph of machines in chess as a milestone -- the end of the easy era. It should also mark a new era, one where researchers eliminate the hype and false promises, epitomized by the prediction of a HAL. We will deepen our admiration for the problem -- and buckle down for some real hard work.

References

HAL's Legacy: 2001's computer as dream and reality, edited by David G. Stork, Foreword by Arthur C. Clarke, MIT Press (1997)
Letter to the Editor on computer chess by David G. Stork, Scientific American, p. 10 (March 1991). Kasparov vs. Deep Blue: Computer Chess Comes of Age by Monroe Newborn, Springer-Verlag (1996) A New Era: World Championship Chess in the Age of Deep Blue by Michael Khodakovsky, Ballentine (1997)

This piece is based in part on my illustrated lecture "The HAL 9000 computer and the vision of '2001: A Space Odyssey'," and Murray Campbell's chapter in HAL's Legacy: 2001's computer as dream and reality (MIT Press, 1997); his insights are gratefully acknowledged. You can read his and other full chapters on-line by clicking on the link to the book. You can see other events associated with the birth of the HAL 9000 computer here.

David G. Stork is Chief Scientist of Ricoh Silicon Valley, as well as Consulting Associate Professor of Electrical Engineering and Visiting Scholar in Psychology at Stanford University. He has had a lifelong interest in chess, and competed twice in the United States High School Chess Championships in the 1970s (he no longer plays competitively). His five books include HAL's Legacy: 2001's computer as dream and reality (MIT Press) and the forthcoming Pattern Classification (2nd ed.) (Wiley).