An AI evaluation metric converts an aspect of system behavior into a value that can be compared across outputs, models or versions. Metrics can measure accuracy, omission rates, unsupported claims, task completion, safety violations, latency, cost and many other properties.
No single metric captures every important quality. A high aggregate accuracy score can hide rare but consequential failures, so teams often combine several metrics with case-level inspection and expert judgment that reflects the real cost of different errors.
ELI5
An evaluation metric is a rule for measuring one part of how well an AI performed. It is like using separate scores for accuracy, speed and consistency instead of calling the whole result simply good or bad.
For example, a video edit could be scored on whether it follows the instruction and whether the person's identity stays consistent. Those numbers are useful, but a person may still need to judge whether the edit looks natural and serves the creator's purpose.
Is accuracy enough to evaluate an AI system?
Usually not. Accuracy can hide the type, severity and context of failures, especially when uncommon errors have serious consequences.
How should AI evaluation metrics be chosen?
They should be chosen from the intended task, user needs, failure modes and relative consequences of false, missing or unsafe outputs.
