AI Glossary Term
What does “Calibration” mean?
How well a model's confidence scores match actual correctness probabilities.
How Calibration works
Example
A well-calibrated weather model that says '70% chance of rain' should see rain on about 7 out of 10 such days. A model that says '95% sure' but is right only 75% of the time is overconfident and needs recalibrating.
Why it matters
People rely on confidence scores to decide when to trust or double-check a prediction. Poorly calibrated scores can mislead them into over- or under-trusting the system.
Common questions
How is calibration measured?
By grouping predictions by confidence and comparing each group's stated confidence with its actual accuracy, often summarized as expected calibration error.
Is accuracy the same as calibration?
No. A model can be accurate but badly calibrated, such as being right 90% of the time while claiming 99% confidence.
Can calibration be improved?
Yes. Methods such as temperature scaling adjust the confidence scores after training without changing the predictions.