A new paper shows AI chatbots can be taught to pause instead of guessing when they lack real information.
Researchers built a prompting method called Chain-of-Self-Questioning (CoSQ), which forces a model to explicitly judge whether it has enough grounding before committing to an answer, rather than just producing something fluent regardless. They tested three CoSQ variants across seventeen conditions on the 817-item TruthfulQA multiple-choice set, using eleven open-weight and hosted model families. The strongest variant, Grounded-CoSQ at a 0.90 threshold, cut the average rate of confidently wrong answers from 13.1% under standard chain-of-thought prompting to 8.9%, a 32.1% relative reduction, while also raising answered-question accuracy from 86.9% to 89.7%. It still answered 87.6% of questions, and the improvements held across all eleven models and every threshold tested, with a secondary check on Natural Questions Short-Answer data pointing the same direction.
The appeal here is that this is prompt-only: no fine-tuning, no new training data, just a different way of asking the model to check itself before it speaks. That matters for anyone deploying LLMs in settings where a wrong but confident answer costs more than a punt to a human, like search, healthcare triage, or legal research tools.
Getting a model to say it does not know more often is not the same as making it smarter. It is a calibration fix, not a knowledge fix, and it still leaves roughly one in ten answered questions wrong.