Many models are confidently wrong, reporting high certainty on answers that turn out to be mistaken. Calibration measures and corrects the gap between stated confidence and real accuracy, so that a confidence number can actually be trusted to guide a decision or trigger a deferral.
Trustworthy AI
Calibration: Why a Model's Confidence Must Match Reality
A confidence score is only useful if it means something. When a model says it is 90 percent sure, it should be right about that often. Calibration is the discipline of making that true.