MarkPNeyer.com

We had a ’round table’ type discussion in my AI class today about the Turing Test and the feasibility of algorithms that can pass it. In particular, the discussion centered on the Chinese Room thought experiment. The “Turing Test” is a method proposed by Alan Turing (although he didn’t call it that) to tell whether or not a computer is intelligent. Basically, a human converses with the computer via an instant-messenger style interface for a half hour, and at the end of the half hour, the human has to be able to tell whether or not the computer actually was a computer or was a human being.

This guy Searle argued that even if a computer could pass the turing test, it still might not have any understanding. A computer follows a set of rules to take input and produce output. Suppose we designed a computer that could easily pass a turing test, provided that turing test was administered in chinese. When presented with sentences in chinese, the computer would follow its rulebook and respond with some more chinese. If a human being were to sit in a room and follow the same algorithm that the computer followed, he could be said to be “passing” the turing test for intelligence, even though he would have no understanding of chinese. Searle argued that his “Chinese Room” example means that passing a turing test doesn’t make a computer intelligent.

First, let me state my opinion that arguing about wither property p applies to object x is pointless unless you’re dealing with a formally defined system. Then, you can prove either way or perhaps prove that “p applies to x” is undecidable. Either way, the nature of the debate is pretty simple. When you deal with that messy place some folks call “meatspace,” however, the arguments just get ridiculous. Invariably somebody brings up what “webster” says is the case. Words mean slightly different things to everybody that uses them, and these arguments almost always boil down to what very abstract words mean to individuals, which is why I think it’s stupid to even argue about such things. Either you think p applies to x or you don’t; there is no right answer. I tried, then to answer in terms of my understanding of how must people define intelligence.

Suppose you ask a person “Alice” whether or not an agent “Bob” is exhibiting intelligence. By agent, I mean anything – person, computer, animal, or particle. In my experience, Alice will only say Bob is intelligent if Alice thinks that Bob makes decisions in a manner roughly isomorphic to the way she (Alice) makes decisions. In other words, most people will say that something that doesn’t think in the same manner that they do is not intelligent. An algorithm that uses brute force to make moves in a game of chess operates in a manner so different from most of us that we won’t label it as intelligent. If the algorithm has a way of evaluating the value of each move based upon one or two predicted moves, uses knowledge of the history of its oponent, and perhaps makes mistakes, more people would be willing to label it “intelligent.”

The World of Politics is full of examples. How many pundits decry those who dont share their opinions as stupid? I saw a magazine cover once, asking about the result of the 2004 election, saying “How can 59 Million people be so stupid?” The 59 Million number was supposed to be the number of people that voted for Bush. Unfortunately for the Daily Mirror, counting after the election increased the total number of votes for Bush to 62 Million and the number of votes for Kerry to 59 Million. Such are the dividends of arrogance, I suppose.

As for whether or not machines could ever posses intelligence, I think that if you want say that machines can’t ever posses intelligence, you must either conclude that humans don’t posses intelligence, or that humans are somehow magical. My reasoning is simple – if you believe in the laws of science, you believe that humans are made up of particles that follow rules. A computer could simulate a human being atom by atom using the same rules. Unless you think that humans are somehow “magic” in that they don’t follow the same rules as the rest of the universe, you’d have to conclude that a [sufficiently powerful] computer could do anything a human mind could do. As to whether this will ever be feasible, that’s an entirely different debate. My answer is a solid “maybe.”