Alison Gopnik: belief

Transcript context

…They do that in biology already, right? The way they’re reasoning is like a kid in a multiplication class who learns, here’s what the formula is that you need to do for multiplication. It’s interesting that they’re not good at things like basic arithmetic often. You get this weird combination of it’s reproducing a reasoning process that you could imagine in math, but it’s not doing basic arithmetic. I think another place where they’re really falling down is the novel relationships that humans are very good at generating. Those are the kinds of things where the generative AI is not working. In any case, in terms of the issue about K–12, every time we have a new way of accessing information, it’s really important to teach. That’s why we teach kids how to read. We teach kids how to calculate. We teach kids how to use the internet. A lot of that teaching is going to be, here’s how this cultural technology works positively to give you some information that’s true, and here’s how it works to give you information that’s completely false. It’s funny that these reasoning systems are still hallucinating. I think the reason why they’re hallucinating is because their objective function isn’t about truth in the way that it is for even a little baby. Their function is to produce something that is going to be the sort of thing that a human will like with something like reinforcement learning from human feedback. Hallucination rates are plummeting. If I use GPT-5 Pro edition, it will outperform a very good human lawyer for a typical query. It will do better on a medical exam than, say, a prospective doctor would do. I’m not saying there’s zero hallucination, but they’re already ahead of humans. If I took an economics test, GPT-5 would beat me.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.