AI Dictionary › Fondamenti AI

Model Calibration

Calibrazione del modello (Model Calibration)

Model calibration measures how closely an AI system's stated confidence in its own answers matches the actual probability that those answers are correct. A well-calibrated model, when it says it is 90% confident, is in fact right about 90% of the time in similar cases. A poorly calibrated model can be overconfident about wrong answers or overly cautious about correct ones, making its uncertainty signals hard to trust.

Definition

How it works

To assess it, many of the model's predictions are collected along with their stated confidence level, and average confidence is compared against the actual frequency of correct answers within each band. Gaps between stated confidence and observed accuracy are summarized into numeric metrics quantifying calibration error. Post-processing techniques can then correct confidence estimates that are too optimistic or too pessimistic.

Applications

It is especially relevant in areas where decisions rely on the model's certainty level, such as assisted diagnosis, risk assessment, or systems that need to know when to ask a human for confirmation. Good calibration allows a model's confidence to be used as a reliable signal for routing uncertain cases to review.

History & etymology

The concept comes from predictive statistics, where probability calibration has been studied for decades to assess the quality of weather forecasts and other probabilistic models, before being applied to machine learning systems.

Related terms

More in Fondamenti AI

Put it into practice

From our network

Magellano GPS: Fleet Tracking Made Simple

Real-time GPS tracking, remote engine lock, fuel and CO₂ reporting for your fleet.

Visit magellanogps.com →

From the Agora Intelligence blog

More on agora-intelligence.com →

📱 Download the Android app (beta) iOS coming soon

Say what you mean. Get what you need.

Grace Certified, the AI coach that trains and certifies your prompt engineering, by Agora Intelligence.