A confidence score is a numerical value, usually between 0 and 1 or given as a percentage, with which an AI model expresses how certain it is about an individual prediction. An image classifier might report “cat, 0.94”, a fraud filter a risk score per payment. Many applications work with thresholds: above them the system decides on its own, below them the case goes to a human.
A high score does not guarantee a correct answer. Models are often poorly calibrated and overconfident, especially with unfamiliar inputs outside their training data. Thresholds therefore need a justification from testing, review in operation and a display that does not lull employees into blind trust. On inputs from unforeseen environments, see T01 Complexity of environment, and on overreliance, T03 Excessive automation (automation bias).