Calibration and reliability

Calibration means a model's 80% confidence is right about 80% of the time. Reliability diagrams show that relationship. Fluent AI scores are often *not* calibrated.

Overconfident probabilities drive bad medical-adjacent, fraud, and ranking decisions. Pair with Course 29 medical overtrust literacy.

Compare predicted probabilities to observed frequencies in bins. Fix with temperature scaling, isotonic methods, or better training—then re-check on fresh data.

Take 20 model confidences; group into bins; see if “90% sure” was actually ~90% correct.

⚠️ Myth: Softmax outputs are true probabilities of the world.

✓ Reality: They’re normalized scores—calibrate before trusting.

• Course 03 metrics; Course 04 probability; Course 08 hallucinations
• Guo et al. calibration literature (cite when teaching)