Calibration — AI Glossary

Reference details

Related terms

• Log Probabilities
• Uncertainty Quantification
• Active Prompting

How it works

1. A set of predictions is collected with both the model's confidence and the eventual correct answer, which means calibration can only be measured where outcomes are known.
2. Predictions are grouped into bins by confidence — typically ten — and the observed accuracy within each bin is computed.
3. Each bin's observed accuracy is compared against its own confidence midpoint, and that gap is the quantity calibration describes.
4. Expected calibration error aggregates those gaps as a weighted mean of their absolute values, with each bin weighted by how many predictions fell into it.
5. A reliability diagram plots observed accuracy against stated confidence, where perfect calibration is the diagonal and the shape of the departure says whether the model is over- or underconfident.
6. Correction is usually a single scalar fitted on held-out data that rescales the logits before softmax, which moves confidence without changing which answer ranks first — so calibration can be fixed without accuracy changing at all.

Example: What well-calibrated actually means, and what miscalibration costs downstream

Take every answer where the model reported 70% confidence. If it is well calibrated, roughly 70 of each 100 are correct. If 95 are correct it is underconfident; if 40 are, it is overconfident and any threshold you set on that number is meaningless. Calibration is a property of the distribution, never of a single answer. The cost shows up wherever confidence is used to route. Suppose answers above 0.9 confidence are auto-approved without review, and that band covers 60% of traffic — but the model is overconfident enough that those answers are only 70% correct. Then 0.6 x 0.30 = 18% of all output is wrong and unreviewed, and the review queue never sees it because the threshold filtered it out. A miscalibrated threshold does not merely waste reviewer time; it routes errors specifically away from the step that would have caught them.

Commonly confused with

Confidence is not accuracy, and a model stating “I am 95% sure” in prose is not reporting a probability — it is generating plausible text. Token logprobs are a genuine signal; a sentence about confidence is not, and the two are routinely confused. Calibration is also not the same as discrimination: a model can rank answers correctly while reporting confidences that are uniformly too high, so rescaling fixes the numbers without changing which answer wins — and a model can be well calibrated and still unhelpful if its confidences cluster in the middle.

When to use it

Reach for it when:

• You route on confidence — auto-approve above a threshold, escalate below
• Abstention matters more than coverage, so a wrong confident answer costs more than a refusal
• You have enough labelled outcomes to measure calibration honestly
• Before setting any auto-approval threshold, since the threshold inherits the miscalibration

Reach for something else when:

• Trusting stated confidence in prose as a probability
• On a model you have not measured; RLHF commonly degrades calibration
• Where the cost of a confident error is high and nothing else checks it
• Confused with discrimination — rescaling confidences does not reorder the answers

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Calibration — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path