Can AI think?

Alan Turing opened his 1950 paper with the question “Can machines think?” and at once set it aside as too imprecise to discuss. He put a game in its place. If a machine, questioned by text, could not be told apart from a person, then on his view we should stop withholding the word. He predicted that by the end of the century people would speak of machines thinking without expecting to be contradicted.

More than seventy-five years on there are machines that hold such conversations, and the question he set aside has returned.

The case that they do

On one family of views, thinking is defined by what it accomplishes. A system that takes a problem it has not seen, works out a plan, corrects itself and arrives at an answer is doing what we call thinking when a person does it. Daniel Dennett’s version is that there is nothing more to having thoughts than being a system whose behaviour is best predicted by treating it as having them. By that standard the case for current systems is strong in some domains and weak in others, and improving.

The case that they do not

John Searle’s Chinese room is the classic reply. A person who speaks no Chinese sits in a room with a rulebook for matching Chinese symbols to other Chinese symbols. Questions come in, correct answers go out, and the person understands nothing. A computer running a program, Searle argued, is in the same position: it has the rules for the symbols and not what they mean.

Emily Bender and Alexander Koller made a related point about language models in particular. A system trained only on the form of language, they argued, has no route to meaning, because meaning is a relation between words and something outside them. This is the old symbol grounding problem in a new setting.

Why the dispute does not end

Each side can describe everything the other points to. The first says that the room as a whole understands, and that to demand more is to ask for a ghost; the second, that behaviour was never the point. The disagreement is about what thinking is, and new systems do not settle that, since both sides predict the same behaviour from them.

Where I stand

I think the question is better broken up than answered. “Thinking” bundles several things that come together in us and can come apart in a machine: inference, understanding, awareness, having something at stake. Current systems plainly do the first to a serious degree. I see no good evidence yet of the last two, and the second is the one in dispute.

For the purposes I care most about, whether we can learn from what such a system says, I do not think the answer has to wait. What matters there is not whether a mind stands behind the words but how the words were produced: whether the truth of what is said played any part in its being said. That can be investigated, and so far the answer for open-ended questions is that it did not.