Calibration

A calibrated model knows what it knows and what it doesn’t — when it expresses high confidence, it’s usually right; when it’s uncertain, it says so. Calibration is a component of honesty: a model that confidently states wrong answers is misleading users, even if it’s doing so without intent. Modern language models tend to be overconfident — they often produce assertive statements without expressing uncertainty even when they should. For behavior architects, encouraging appropriate expressions of uncertainty (“I’m not certain, but…” or “You may want to verify this”) is a behavioral design choice that requires balancing honesty against usability — because excessive hedging is its own problem that erodes trust and usefulness.