Monday, September 12, 2005

the Turing Test

Angry Clam at Patterico makes an allusion to the Turing Test that I think warrants a response.

The Turing Test, proposed by Alan Turing sometime in the first half of the 20th century, is a proposed way to test a machine to see if it is intelligent. The details aren't important and the idea can be reduced to this: if the machine acts like it is intelligent, then we can assume that it is.

The problem with this test should be obvious: mimicry is not a sign of genuiness. Where else in science or in daily life do we assume that just because an X can mimic a Y very well, that therefore the X must be a Y? Just because a mime seems to be in a glass box, that means he is in a glass box? Just because a bug can look just like a leaf, that means it is a leaf? Just because a painting looks just like an original Rembrandt, that means it is an original Rembrandt?

Of course in each of these cases, a closer examination may reveal the mimicry. So let's give the Turing Test the benefit of the doubt and assume that there is no observation of the machine itself that will tell you it is not intelligent. What does this prove?

Let's go back to the analogy of the painting. If someone produced a forgery of a Rembrandt with such care that there was no possible way to tell it was a forgery just by examining the painting, would that make the forgery an original Rembrandt? Of course not. What makes a painting an original Rembrandt is not its physical properties but its history. In order to be an original Rembrandt, it must have been painted by Rembrandt.

Similar considerations apply to intelligence. What makes something truly intelligent is not just that it has effects that are indistinguishable from intelligence, but that it has thoughts and intentions. The mere appearance of thoughts and intentions does not make something intelligent any more than the appearance of being old makes something old.

But what about the benefit of the doubt? If some machine appears to be intelligent, shouldn't we give it the benefit of the doubt on the grounds that we want to err on the side of charity? I don't actually have to address that argument because there is no doubt. I know why the machine behaves the way it does: because it is programmed to behave that way. There is no reason at all to posit intelligence. The situation described by the Turing Test is analogous to a famous art forger who shows you a painting and says that he created it himself only a few days ago. He explains exactly how he did it, how he made the paints and the canvas seem older, how he copied the lines and the paint chemistry and the paint thickness exactly. He proudly tells you exactly how he forged the painting and then argues that you should accept it as an original Rembrandt because by examining the painting itself you could not possibly discern the difference.

Would you find this argument convincing? Would you give him the benefit of the doubt and pay him a million dollars for a forgery just because you agreed that you could not possibly detect it as a forgery? Would you take it home, proud of owning a piece of history, an original Rembrandt?

This argument is preposterous, but it is exactly what advocates of the Turing Test would have us accept. They posit the existence of a machine that is designed to mimic intelligence. By hypothesis, since the machine is a machine, there is a perfectly sound explanation in terms of mechanics and electronics for every action the machine takes, yet they would have us ignore these explanations and pretend instead that the machine acts out of intelligence.

I find it unnerving that so many intelligent people accept such a transparently ridiculous position. It can only be explained by some sort of black-box fallacy. We don't understand what produces intelligence, it is a black box. Some people seem to respond to this mystery by reducing it to something they can understand, behavior, but in the process throwing out whatever is interesting about it in the first place. Others seem to respond by the voodoo principle as though intelligence were a bit of magical lint that settles upon anything at all that behaves in a certain way. But just because you don't understand what makes an original Rembrandt, that doesn't mean you can define it to be a purely physical property or assume that original Rembrandtness somehow magically settles on anything that has certain physical properties.

No comments: