AI uncertainty quantification aims to describe what a model does not know rather than returning only one answer. Methods may produce probability distributions, prediction intervals, ensembles, calibrated confidence values, or separate estimates for uncertainty caused by noisy data and limited model knowledge.
Uncertainty estimates can guide a system to gather more evidence, choose a safer action, or defer to a person. Their usefulness depends on calibration and coverage outside the training distribution, so estimates must be tested on representative errors, rare cases, and changing real-world conditions.
ELI5
Uncertainty quantification is a way for an AI system to show how unsure it is instead of giving only one confident-looking answer. It might provide a range, a probability, or a warning that the situation is unfamiliar.
For example, a weather model could predict 20 millimetres of rain but say that 10 to 35 millimetres is a reasonable range. That extra information helps people decide whether to gather more evidence, take a safer action, or ask a person to review the result.
Why does AI uncertainty quantification matter?
It helps systems avoid false certainty and supports safer decisions about when to act, collect more information, or request human review.
How can AI uncertainty be represented?
It can be represented with probability distributions, intervals, ensembles, confidence estimates, or other calibrated measures tied to possible outcomes.
