Last month, a room at a British university was full of people engrossed in online chat. One conversation ran like this:
A: I consider myself an amateur philosopher, yes.
B: Any particular branch?
A: You are so pretty.
B: Thank you. Are you handsome?
A: I am extremely cute.
B: What are you wearing today?
But this wasn’t online flirtation between two students. Indeed, only one of the participants was human. “B” was a judge of the Loebner Prize, an annual competition to find the best “chatterbot” computer program for imitating human chat, and “A” was a bot named Jabberwacky.
Jabberwacky eventually lost out to a bot named Jabberwock. But if the names have you thinking that chatterbots are simply a game for teenage nerds, think again. Just as Lewis Carroll wrote sophisticated analyses of logic and language, then put those ideas into some of the most famous children’s books ever written, there is a serious scientific background to the Loebner Prize contest.
The American businessman Hugh Loebner sponsored the first competition in 1991, offering $100,000 for a conversation program capable of convincing judges that its responses were from a human being. The contest has been held annually ever since, stimulating and sharpening the conversational powers of the bots, as the spirit of competition has grabbed their human creators.
In 1991, the conversations were limited to specific topics. Since then, there has been enough progress to enable general discussion between judge and computer. There is, of course, a control: The judges, sitting in one room, don’t know whether they’re talking to a computer or a human sitting next door. This fact gives the competition a game element — or, as there’s an audience, makes it akin to a technological game show.
The game is asymmetric, because the humans don’t have to pretend to be computers — they’d fail dismally! Computers have computational power that humans can’t rival. (Google “Loebner Prize” and goggle at the speed of retrieval and sorting.) But does any amount of computational power make something “intelligent,” in the sense we apply the word to human beings? Could a computer program, for example, show a real understanding of what the Loebner Prize is? So far, no program has come near the goal of simulating human intelligence, though occasional flashes of apparent brilliance may temporarily confuse the judges.
This isn’t a test of supercomputers, though — but of ingenious programs. It’s the programs, which could be run on any computer, that make the chatterbots chat. It’s the programs that compete against humans and against each other — everything is in the software.
All this was foreseen 50 years ago. The man who devised the Loebner Prize’s scenario of a human conversing with a machine, back in 1950 when computers barely existed, was also the man who first clearly defined the roles of computer hardware and software.
That man was Alan Turing, who invented the computer in 1936. In that year he described the idea of a universal machine, running any program placed on its input tape. This became the principle of the stored-program digital computer, when it was embodied in electronics after 1945.
In 1950, Turing wrote a paper on the idea of software simulating the mental operations of the brain — a paper that is now one of the most famous in scientific literature. In it, Turing emphasized that to evince human intelligence, a program must be capable of witty repartee. He also devised the game-show format of competition with a human, with communication only through computer terminals.
What he called the “Imitation Game” is now usually called the “Turing Test.” And he prophesied that there would be moderate success in this test before the end of the 20th century. Turing’s prediction caught Loebner’s imagination, and the entrepreneur suggested a series of trials as 2000 approached.
More important than the exact form of the Turing Test is the mathematician’s assertion that intelligence could eventually be passed by a computer program — that Artificial Intelligence (or “intelligent machinery,” in his words) would be created.
Turing knew this was a provocative idea; indeed he described himself as a “heretic.” He understood how the notion of “intelligent machinery” presented an apparent contradiction in terms: We don’t feel like computers; we seem to have free will; we are conscious. We feel, we intend, we live — and computers don’t.
But Turing responded to these natural objections with some powerful arguments. Somehow, the matter in our brains manages to produce originality, invention, surprise. So somehow, computer programs simulating the action of the brain should be able to do the same thing.
The game-show format also exploits another Turing argument: that we can only deduce that other people are intelligent by external observation. We cannot enter into their minds. Why judge computers differently? The Turing Test scenario is intended to override our subjectivity and apply an objective assessment of “intelligence.”
Many objections have been raised to Turing’s format. For instance, there are class and cultural elements to language — just as there are to old-fashioned “intelligence tests” that evaluated their subjects by applying narrowly conceived notions of knowledge and intelligence.
Turing can also be criticized for putting forward hypothetical conversations intended to exemplify the appearance of “intelligence,” but which are conspicuously lacking in any moral seriousness, and fail to follow through a topic for more than a couple of lines. In fact, they might be said to anticipate the wackily charming but inane chat produced by the game-playing modern bots.
You can easily sample the judging work in this year’s test for yourself, online, and see the strategies chatterbots use to confuse you. The best way to catch them out is to insist on sticking to a serious topic — something they just cannot do. (Turing’s example conversations, as outlined in 1950, can be criticized for sharing the frivolity produced by the game-playing modern bots.)
Nevertheless, Turing’s underlying case is not easy to counter. If the brain works in some definite way — and it’s a physical organ, not a matter of magic — why can’t it be simulated by a sufficiently complicated program?
Turing emphasized that the new theoretical concept of “computability” was just the right approach to addressing the question of “mind.” He claimed that programs are the best way to capture the essence of design, organization and structure — and hence, ultimately, everything that brains have evolved to do.
Turing started this line of thought in the “cybernetics” heyday of the 1940s. He put forward practical ideas for research in that 1950 paper, which anticipated the AI work that began in the 1950s, principal among which were the rival approaches of neural networks and expert systems.
“Neural networks” try to imitate the organization of the brain, inspired by the idea that all human mental activities must somehow be supported by the physical structure of the nervous system. Turing was particularly keen on the idea that a computer system ought to be able to learn, modifying its behavior in the light of experience, in something like the way that brains learn.
“Expert systems” use advance programming to create a sort of encyclopedia of facts and logical deductions, capable of rivaling human experts in specialized fields, and suggesting to optimists the possibility of being able to code all human knowledge and understanding.
AI back online?
The 1991 Loebner Prize competition coincided with the Internet explosion, with its newsgroups and bulletin boards (the call for entries was one of the first documents to be written for the early Web) and the competition took off globally.
It also arrived concurrently with a new phase of optimism in AI — ambitious claims and speculation revolved around ideas such as hybrid systems involving both advanced programming and neural networks — which until then had been seen as mutually exclusive, rival approaches. Researchers were investigating “genetic algorithms” to exploit the principles of biological evolution, and there was also hope that new light would be shed on the nature of the brain’s activity by studying the complex behavior of chaotic systems.
Despite its apparent good timing, though, the Loebner Prize is frowned on by many serious AI researchers. It jumps the gun, they say. It produces not intelligence, but merely the appearance of intelligence — amusing programs trading superficial conversational tricks.
Marvin Minsky, a leading figure in AI, wrote of his wish that “Mr. Loebner will indeed revoke his stupid prize, save himself some money, and spare us the horror of this obnoxious and unproductive annual publicity campaign.”
Dissenters also contend that the competing chatterbot programs are not the product of serious research departments, but of individuals with a strong streak of Turing’s maverick disposition.
However, it’s hard to dismiss the game show completely, because it keeps alive one of the original roots of this scientific research — and behind that, one of the deepest dreams of humanity: a hope for immortality by the cloning of mind.
Stranger than fiction?
Dream or nightmare? Science-fiction writers have readily seen that the combination of computational power and true intelligence would swiftly surpass us. Moviegoers’ current fascination with “The Matrix” shows how sensitive we are to the possibility of not being able to tell reality from an imitation world. Turing argued that a computer might only be “imitating” mind, but this is sufficient: Simulating intelligence is intelligence.
This raises doubts in the minds of many: Why should deceit, as imitation surely is, be at the heart of a test of intelligence? A different view was advanced by the eminent Oxford mathematician Roger Penrose in his 1989 book, “The Emperor’s New Mind.” Penrose proposes that consciousness is an absolute reality that we know as well as we can know anything: not a game show, but a reality show. But at the same time Penrose, just as much as Turing, takes a materialist viewpoint, in which there is nothing to the mind that is not supported by the physical structure of the brain.
Where the two differ is in Penrose’s suggestion that the brain works in a way that can’t be replicated in a program, because it involves quantum physics. His idea is that there must be something in quantum theory, when it is fully understood, that cannot be imitated by the components of a computer, no matter how large or fast. And he believes that consciousness — and with it crucial questions of human intelligence, integrity and truth-telling — must be bound up with this as yet unknown physics.
Most scientists today would consider this an outlandish suggestion. But from a hint in a 1951 radio talk, it seems Turing himself was seriously concerned by the role of quantum-mechanical physics in the operation of the brain — something he had thought about in his teenage years.
Turing referred to the views of the mathematician and scientific popularizer Arthur Eddington on the mysterious role of the “Uncertainty Principle.” Curiously enough, Eddington had then compared quantum theory to Lewis Carroll’s Jabberwock: “Something unknown is doing we don’t know what.”
Fifty years after Turing’s dramatic life and death, there is still much unknown about the Jabberwock. Meanwhile, according to one report, if the Jabberwacky bot did not quite pass the Turing Test this time round, it certainly made the judges laugh.
In probing the nature of mind and matter, as (in its way) the Loebner Prize is doing, it may turn out that Nature still has a punch line we have not yet guessed.