Parry: Cops arrest the wrong people.
Interviewer: They arrest the right ones sometimes.
Parry: The law should get those Italian crooks.
Interviewer: Are you Italian?
Parry: My background is British, but I was born in this country.
Interviewer: Do you know any Italian crooks?
Parry: I try to avoid the underworld.
Interviewer: Are you afraid of the underworld?
Parry: The mafia are the worst gangsters.
Interviewer: Are you connected with the mafia?
Parry: You know, they know me.
Interviewer: How do they know you?
Parry: They spy on me.
Interviewer: Why do they?
Parry: They want to get false evidence to frame me.
Interviewer: What do they have against you?
Parry: They are out to get me.
Parry sounds like a man in trouble? Maybe, but his biggest problem isn’t the Mob—it’s the fact that he’s a computer program. Created in 1971 by Kenneth Colby, a psychiatrist at Stanford University, Parry was programmed to respond to questions in the manner of a schizophrenic with a paranoid fixation that he is a mafia target. Colby set up a test in which Parry was interviewed alongside a number of genuine paranoid patients and the results were then assessed by a panel of psychiatrists. No one on the panel guessed that Parry was not a real patient.
Can Parry think? Twenty-one years before Parry’s birth, in 1950, the British mathematician and computer pioneer Alan Turing wrote a seminal paper in which he proposed a test to determine whether a machine could think. The test, based on a party game called the imitation game, requires an interrogator to communicate with a human and a machine, both physically separated from her, by some form of electronic link. She can ask any question in order to distinguish the machine and the human, and if after a certain period of time she is unable to do so, the machine is said to have passed the test.
“I believe that at the end of the [20th] century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted.”
Alan Turing, 1912–54
Did Parry pass the test? Not really. To count as a proper Turing test, the panel of psychiatrists (filling the role of the interrogator) should have been told that one of the patients was in fact a computer and that the task was to identify which. In any case Parry would very quickly have revealed itself if questioned more broadly. Turing himself believed that by the end of the 20th century advances in computer programming would have reached a point where an interrogator would have no more than a 70 percent chance of making a correct identification after five minutes of interviewing, but in fact progress has been far slower than he had anticipated. So far no computer program has come close to passing the Turing test.
Turing proposed his test to sidestep the question “Can machines think?,” which he regarded as too imprecise to be worth addressing, but the test is now widely accepted as the criterion by which to judge whether a computer program is able to think (or “has a mind” or “shows intelligence,” according to taste). As such, it is seen as the benchmark by proponents (scientific and philosophical) of “strong AI (artificial intelligence)”—the thesis that suitably programmed computers have minds (not just simulations of mind) in precisely the same sense as do humans.
“Current attempts to understand the mind by analogy with man-made computers that can perform superbly some of the same external tasks as conscious beings will be recognized as a gigantic waste of time.”
Thomas Nagel, 1986
The Chinese room The most influential challenge to the Turing test has been posed by a thought experiment devised in 1980 by the US philosopher John Searle. He imagines himself—an English-speaker not knowing a word of Chinese—confined within a room into which batches of Chinese scripts are posted. He is already equipped with a pile of Chinese symbols and a copious rule book, in English, which explains how he is to post out certain combinations of symbols in response to strings of symbols in the batches posted to him. In time, he gets so adept at his task that, from the point of view of someone outside the room, his responses are indistinguishable from those of a native Chinese speaker. In other words, the inputs and outputs into and out of the room are exactly as they would be if he had a full understanding of Chinese. Yet all he is doing is manipulating uninterpreted formal symbols; he understands nothing.
Producing appropriate outputs in response to inputs, according to rules provided by a program (equivalent to Searle’s English rule book), is precisely what a digital computer does. Like the incumbent of the Chinese room, Searle suggests, a computer program, however sophisticated, is no more than, and could never be more than, a mindless manipulator of symbols; it is essentially syntactic—it follows rules to manipulate symbols—but it can have no understanding of meaning, or semantics. Just as there is no understanding within the Chinese room, so there is none in a computer program: no understanding, no intelligence, no mind; and never more than a simulation of these things.
Passing the Turing test is basically a matter of providing appropriate outputs to given inputs, so the Chinese room, if accepted, undermines its claim to work as a test for a thinking machine. And if the Turing test goes, so too does the central thesis of strong AI. But these are not the only casualties. Two extremely significant approaches to the philosophy of mind are also undermined if the point of the Chinese room is granted.
Arthur C. Clarke took Alan Turing’s prediction at face value. For his 1968 collaboration with Stanley Kubrick, 2001: A Space Odyssey, he created an intelligent computer named HAL (each letter one removed from IBM). In the story, none of the humans is surprised that a thinking machine is running their spaceship.
Problems for behaviorism and functionalism The central idea behind behaviorism is that mental phenomena can be translated, without any loss of content, into kinds of behavior or dispositions to behavior. So to say that someone is in pain, for example, is a sort of shorthand for saying that they are bleeding, grimacing, etc. In other words, mental events are defined entirely in terms of external, observable inputs and outputs, the sufficiency of which is explicitly denied by the Chinese room. Behaviorism, given its classic exposition by Gilbert Ryle (see The mind-body problem), had largely succumbed to a number of fatal objections before Searle’s appearance. Its importance today is rather that it spawned a doctrine that is probably the most widely accepted theory of mind—functionalism.
Repairing many of the flaws in behaviorism, functionalism claims that mental states are functional states: a certain mental state is identified as such by virtue of the role or function it has in relation to various inputs (the causes that typically bring it about), the effects it has on other mental states, and various outputs (the effects it typically has on behavior). To use a computer analogy, functionalism (like behaviorism) is a “software solution” to the theory of mind: it defines mental phenomena in terms of inputs and outputs, with no consideration of the hardware platform (dualist, physicalist, whatever) on which the software is running. The problem, of course, is that focusing on inputs and outputs threatens to lead us straight back into the Chinese room.
the condensed idea
“Did you ever take that test yourself?”
| Timeline | |
|---|---|
| c.AD1637 | The mind-body problem |
| 1912 | Other minds |
| 1950 | The Turing test |
| 1974 | What is it like to be a bat? |