You can now tell an AI to shut up instead of guess — and it actually works

Researchers found that prompting large language models to express uncertainty and reward them for saying 'I don't know' instead of guessing reduces false answers without retraining the model. In practice, this means you can make existing AI systems more honest about what they don't know by changing how you ask them questions and what you reward them for.

Large language models are confidently wrong all the time because they're trained to produce an answer, not to admit uncertainty. This work shows you can fix that with text prompts alone, no model retraining required. That matters because it's cheap to deploy and works on models you don't control — you just change the instructions you give them. The trade-off is real though: you get fewer answers, but the ones you get are less likely to be hallucinated garbage.