I’m a former philosophy professor1 who is now learning software and management.I like to think about knowing and believing. Lately, some of the knowers and believers I think about are software.

When you believe something, you don’t just have a thought pass through your head. You take a stand. You’re committed to it: you’ll rely on it when you act, you’ll defend it if someone challenges you, you owe reasons for it, and when good enough reasons come along against it, you’re supposed to change your mind. That’s a large part of how we hold ourselves and each other responsible for what we believe and what we do. Reasoning, in the sense I care about, is the work of keeping those commitments in order: weighing a new piece of information against everything else you hold, working out what it bears on and how much it matters, and revising where you have to.

The AI tools many of us now work with every day talk as if they believe things, and often they reason impressively. But it’s far from clear that anything in them works like commitment. For a model, everything in front of it (your question, a document it just read, an instruction slipped into a web page, its own earlier answer) arrives as more text with roughly the same standing, and nothing marks which of it the system actually stands behind. My research asks how much of good reasoning these models can do on their own, what kind of structure it takes to give them something like real commitments, and what that teaches us about belief and reasoning in the first place.