Are confidence values always probabilities?

No; some are scores with only an ordering interpretation.

Can calibration fix every model error?

No. It can adjust the relationship between scores and observed frequencies, but it does not make incorrect labels correct.

Is a high-confidence mistake impossible in a calibrated system?

No. Even a calibrated 0.9 group should include some errors over repeated comparable cases.

AI Glossary / Statistics and probability / Intermediate

Confidence values

Definition

Confidence values are scores that an AI or statistical system attaches to predictions, extracted information, or other outputs to express estimated reliability or strength of support. Their meaning depends on how the system defines and computes them. Some are intended as probabilities; others are ranking scores or heuristic measures. A high confidence value does not, by itself, establish that an output is correct.

What it means

A speech recognizer may attach confidence to a recognized word, while an information-extraction system may score a proposed relationship found in text. A classifier may use its largest predicted class probability as confidence in the selected label. These values are not necessarily comparable across systems or tasks.
Before interpreting a score, ask what event it refers to, what scale it uses, and whether empirical evidence supports that interpretation. Confidence in a word is different from confidence that an entire sentence is correct. A model's fluent statement that it is certain is also different from a validated numerical estimate.

How it works

Define the target of confidence: a class label being correct, a transcription matching the reference, or an extracted fact satisfying a specified criterion. Identify the score's construction, such as a model probability, a decision margin, supporting patterns, or a separate correctness predictor.
If scores are intended as probabilities of correctness, assess calibration on suitable held-out examples. Among predictions assigned confidence near 0.8, approximately 80% should be correct for that interpretation to be useful. Finite samples and grouping choices create uncertainty in such assessments.
Distinguish calibration from discrimination. A score can rank easy and difficult cases well yet overstate reliability. Conversely, a constant score equal to overall accuracy can be calibrated in a limited aggregate sense while offering no useful ranking of cases.
Choose operational thresholds using the consequences of mistakes and the capacity for review. Evaluate the accepted subset and the rejected subset, rather than assuming a numerical threshold guarantees an error rate. Reassess scores after changes in data, model, or workflow. Calibration fitted on one population may not transfer unchanged to another, and correlated evidence should not be counted repeatedly as independent support.

Three practical examples

A speech system uses word-confidence scores to identify passages for human checking.
An extraction pipeline assigns confidence to proposed supplier-product relationships before adding them to a reviewed database.
A document classifier routes lower-confidence cases to an analyst while measuring the accuracy and workload associated with the chosen threshold.

Think of it this way

Think of a confidence value as a gauge whose markings need validation. A reading of 90 is meaningful only when the unit, target, and behavior of the gauge are understood. A precise-looking display can still be poorly calibrated, and a well-calibrated gauge can still register uncertainty about an individual case.

Worked example

Consider an original evaluation of 200 predictions. For 100 predictions, the system reports confidence 0.9 and gets 80 correct. For another 100, it reports confidence 0.6 and gets 60 correct.
The first group's observed accuracy is 0.8, ten percentage points below its stated confidence. The second group's observed accuracy equals its stated 0.6. Overall accuracy is (80+60)/200=0.70, while average reported confidence is (100 times 0.9 + 100 times 0.6)/200=0.75.
Using these two predefined groups, a simple weighted absolute calibration gap is 0.5 times |0.9 - 0.8| + 0.5 times |0.6 - 0.6|=0.05. This is a descriptive grouped calculation, not a complete or uncertainty-free measure of calibration.
If the team accepts only predictions scoring at least 0.8, it accepts 100 of 200, giving 50% coverage. Those accepted predictions have observed accuracy 80%, not 90%. The threshold selects the better-performing group here, but its label does not guarantee the advertised probability. New evaluation data are needed to assess a revised calibration mapping or threshold without reusing the same evidence uncritically.

Illustrative case study

In a fictional company, Clearpath Records uses an AI system to extract contract dates. Its interface displays confidence as a percentage, and staff initially interpret the number as a verified probability that every extracted field is correct.
The evaluation team defines correctness separately for each field and checks scores against reviewed examples. It finds that document types differ in difficulty and that confidence for one field does not describe the whole record. The team revises the interface explanation and chooses review rules based on measured performance and error consequences.
This fictional case illustrates how confidence should be interpreted and evaluated. It does not assert a measured improvement or provide real system performance figures.

Common misconceptions and mistakes

Do not confuse confidence scores with statistical confidence intervals. A value between zero and one is not automatically a calibrated probability. High token probability does not establish factual truth of a complete answer. Overall accuracy does not validate every confidence level or subgroup. Lowering a threshold increases acceptance but may increase errors. A self-reported confidence statement requires evaluation; it is not independent evidence supporting the answer.

Business use

Businesses use confidence values to prioritize review, request clarification, compare candidate outputs, and manage automation thresholds. Their usefulness depends on a clear definition and evidence that they help the intended decision. Report acceptance coverage, errors among accepted cases, and important subgroup behavior. Keep the human review process measurable as well, since routing a case to a person does not guarantee a correct final result.

Check your understanding

Why does the example's 0.8 acceptance threshold not produce 90% observed accuracy?

Answer: The retained score group is overconfident in the evaluation sample.

What is coverage?

Answer: The fraction of cases accepted, here 50%.

Why distinguish field and record confidence?

Answer: Correctness of one extracted item is a different event from correctness of an entire collection of items.

Related terms

Calibration
Discrimination
Expected calibration error
confidence score
predictive probability
reliability diagram
selective prediction
abstention
decision threshold
uncertainty estimation

Learn next

Learn probability calibration and reliability diagrams, then compare threshold policies using both coverage and error rates. Study how score definitions differ across classification, speech recognition, and information extraction. Next examine evaluation under distribution shift and how to validate calibration without fitting and judging it on the same examples.

Sources

Jurafsky and Martin, Speech and Language Processing, information-extraction draft, indexed discussion of confidence values: https://web.stanford.edu/~jurafsky/slpdraft/20.pdf Guo and colleagues, On Calibration of Modern Neural Networks, original research: https://proceedings.mlr.press/v70/guo17a.html Scikit-learn official documentation, Probability calibration: https://scikit-learn.org/stable/modules/calibration.html

Previous: Confidence interval Next: Configuration