Confidence calibration: what does “90 percent” mean?

Confidence calibration concerns whether probability estimates match observed frequencies across an appropriate set of cases. Roughly, among many comparable predictions assigned a 90 percent probability of being correct, a well-calibrated system should be correct about 90 percent of the time.

Guo and colleagues studied calibration in neural-network classifiers and showed that accuracy and calibration need separate attention. Their 2017 experiments do not establish how well any current conversational product's confidence is calibrated. On Calibration of Modern Neural Networks.

A confident voice is not a probability

An assistant saying “I'm quite certain” is not automatically reporting a measured probability with a documented evaluation behind it. Even a number on a screen needs a definition. Does it represent a chance of correctness, similarity to prior examples, a ranking score, or something else?

Imagine an invented repair system marking a component “90.” One technician reads that as a 90 percent chance of failure. Another reads it as the ninetieth position on a priority scale. Their disagreement concerns the meaning of the output before either considers whether it is reliable.

A good probability does not choose the action

Suppose a genuinely calibrated forecast predicts rain for tomorrow. Whether to cancel an outdoor event depends on more than the probability: the shelter available, the cost of cancellation, the consequences of getting wet, and who bears each cost. Better forecasting informs that choice without supplying all its values.

Similarly, calibration across one evaluation population does not automatically establish calibration for a different population or a changed environment. Asking where and when a score was checked helps a reader understand its scope.

Follow the boundary in the fiction

The Sovereign Mercy sample places a driver's observations beside a system's risk language. Chapter Eight explores a fictional relationship between confidence and an operational capability. Neither scene supplies a real benchmark or a safe operating threshold.

The useful reading question is whether a prediction, an instruction and a grant of authority have been treated as though they were the same thing. Continue with the evidence and judgment guide for how those steps differ. You need not calculate the probability yourself to ask what the number means and which decision it is being used to justify.