AI calibration is the process of adjusting an artificial intelligence model so that its confidence scores more accurately reflect the likelihood that its predictions are correct. A well-calibrated model should, for example, be correct about 80% of the time when it assigns a prediction a confidence level of 80%.
Calibration is especially valuable when AI outputs guide decisions rather than simply provide classifications. In fraud detection, medical risk assessment, or automated quality control, reliable confidence estimates can help determine when to trust a prediction and when human review is needed. Common techniques include temperature scaling, Platt scaling, and isotonic regression. Related concepts include model evaluation, uncertainty estimation, and AI benchmarking.
