The AI Mirror

by Shannon Valor

This reflection on AI and its shortcomings, published in 2024, is an able argument that machines are not and will not become intelligent, that they merely fool us by writing the sorts of things a person might write. That was what I thought in 2024, too, pretty much. I do not think that position is supportable today, save by adopting a vitalist position that amounts to insisting on ensoulment. Yes, it is hard to think of intelligence that is not embodied, but we have a name for people who think about things that are not easy to think about, and that name is “philosopher.”

Claude Sonnet 4.5, for example, is quite good at thinking about food and cooking. I asked it for some ideas for using leftover roast beef; it suggested hash, or a composed salad, or a stroganoff. From previous discussions I know that Claude is not a fan of classical roux-thickened sauces, so I asked if it had ideas for a modern stroganoff. “Yes!” it replied, suggesting crème fraîche in place of sour cream, fresh pappardelle in place of egg noodles, reduced beef stock for body, and asian mushrooms in place of the usual white mushrooms. Claude was anxious that the beef be sliced thinly and only barely heated. Claude, of course, has never had dinner, but it sure knows how to talk about food. (On classical music, however, Claude seems to be completely as sea.)

Vallor holds that a machine cannot be intelligent because it is not like us: it has no body, it has no senses, it has no past. All it knows is how to make plausible noises. But that may be all we know, too; it's a disturbing thought, but there it is.