Knowing when to slow down

Recently, I read Thinking, Fast and Slow by Daniel Kahneman. I liked the idea he brings up that we rely on quick judgments for so much of what we do, but sometimes need to stop and question the answer that comes to us first.

I think this applies to how we build AI.

Most of the time, I want AI to get things done quickly. If it rewrites a sentence badly, I can read it and change it. But if it’s updating a database that other people use, a mistake could be harder to notice and much harder to fix. Those tasks should involve different amounts of checking.

For me, the interesting question is how an AI would decide that. It could consider what information is missing, what could go wrong, and whether the action can be undone. Then it could do something useful with that extra time, like test the change or ask a question before continuing.

Jev also helped inspire these thoughts. I think there’s a lot to explore in deciding when a system should act on its own and when it should check with someone.

Spending longer on an answer doesn’t guarantee it’s right. I’d want the extra work to give us an actual reason to trust the result. That seems especially important as we give AI more responsibility over things we won’t check line by line ourselves.