Artificial intelligence researchers are very fond of the idea of a Turing test: a computer program converses with a person who can't see their fellow interlocutor. If the person cannot tell that they are talking to a computer, as opposed to another person, then the program must be intelligent. (Or at least as intelligent as the person's friends...)
Turing's brilliant proposal cuts the Gordian knot of defining what is or is not intelligent behaviour with subtle elegance: if people can't tell that they are not talking to other people through the wall when they actually talk with the computer, then the computer must be at least as intelligent as people expect people to be.
Of course, no one has written a program that will pass the test yet. But programmers often simplify things by restricting conversation to a narrow area, such as fashion, or physics.
Others try to write programs that do non-conversational intelligent things that people do: the world chess champion is now a computer program, and poker-bots attempt to fleece real people on online poker sites.
But how about a program to pass human intelligence tests? These tests are highly structured, and so it should be relatively easy to write a program to interpret the input. And they usually require a rather restricted range of pattern-matching skills to solve. Finally, the score can be directly compared with that of the large groups of people who also take the tests.
It might be rather salutary if computer programs were to turn out to be both good at IQ tests and quite simple to write...
No comments:
Post a Comment