Researchers Find Free Way to Teach AI to Say 'I'm Not Sure'

Large language models like ChatGPT are fluent but sometimes state fake facts as confidently as real ones. This is called hallucination. The hard part is getting the AI to know when it's on shaky ground. The old way was to train it with thousands of labeled examples of right and wrong answers — expensive and slow.

Researchers tested a clever shortcut: a model's own confidence score. When an AI settles on an answer, it has an internal probability of being right. If that number dips, it's often a sign the answer is weak. So instead of labeling every answer, they trained the model to shut up when its confidence is low and only speak when it's high.

They tested this on six open-source AI models (1B to 8B size) across two major families. Surprisingly, this "free" method performed just as well as the supervised approach that used labeled data — there was no statistically significant difference. To confirm it wasn't just memorizing hard questions, they ran a control group that studied extra examples instead of abstaining, and that didn't help.

The catch? The signal can't catch confidently wrong answers — mistakes the model is very sure about but gets wrong anyway. Still, this is a big deal. It means future AI systems could be trained to be honest without spending millions on data labeling. For users, that means fewer made-up citations, false medical advice, or confidently wrong legal answers. It's a step toward AI you can trust enough to say "I don't know."