A percentage can be part of the answer text
A language model can produce a confidence statement in the same way it produces other text. The presence of a number does not show that the system has measured its own error rate for this question.
Research on what language models know and what people think they know examines calibration and how explanations affect judgments. Its findings concern tested models and conditions, not a universal accuracy guarantee for future products.
What calibration would look like
Imagine a fictional system that labels 100 comparable answers as 80% confident. If about 80 are correct, that group is consistent with calibration at that level. More data and other confidence ranges would be needed to evaluate the system well.
Even a calibrated group does not identify which individual answer is wrong. The question at hand still needs evidence when accuracy matters.
Calibration and ranking differ
A system might reliably give higher scores to correct answers than incorrect ones while its percentages remain too high. Another might have reasonable average probabilities but weak ability to distinguish difficult cases. A single average score can hide these differences.
A better follow-up than “Are you sure?”
Asking for the source, the relevant passage, or a calculation that can be independently checked can provide evidence about the answer. Repeating a confidence question may produce firmer wording without new information.
For a high-stakes use, an evaluation should define correctness, use relevant cases, and test how confidence behaves when the system encounters unfamiliar inputs. The useful question is not how certain the system sounds. It is whether the stated confidence has earned a particular interpretation through measurement.
