Calibration, in the technical sense, describes the alignment between a model's expressed confidence and its actual accuracy. A perfectly calibrated model, when it says it is 80% confident, is right 80% of the time. When it says 60%, it is right 60% of the time. Its stated uncertainty maps truthfully onto its actual uncertainty.
Poor calibration in AI systems typically runs in one direction: overconfidence. A model trained to produce fluent, authoritative-sounding text may assign high confidence to outputs that are simply wrong — not because it is deceiving anyone, but because there is no mechanism forcing it to distinguish between what it knows and what it is generating plausibly. This is the root of what researchers call hallucination: confident confabulation.
Good calibration is a design goal, not a default. Training on human feedback, with explicit signals about when a model should hedge, can improve it. Techniques like temperature scaling adjust a model's confidence outputs after training. But the problem is genuinely hard, and even capable systems fail it regularly in edge cases.
The human translation is direct. Calibration is epistemic honesty: the ability to say 'I don't know' without it feeling like failure. In many professional cultures, expressing uncertainty reads as weakness. But a well-calibrated person is more useful than a confident one who is often wrong. The question 'how sure are you?' deserves an honest answer. The models — and the people — that give one are the ones worth trusting.
