Task accuracy compares a system's outputs with a correct answer, expected state, or verified standard for a particular task. The measurement must define what counts as correct and how partial success or ambiguous cases are treated.
Accuracy can vary substantially between tasks and conditions. A system that is accurate on selected examples may still fail on different applications, longer workflows, unusual inputs, or tasks that require precise visual judgment.
ELI5
Task accuracy asks how often an AI helper gets a particular kind of job right. The job and the meaning of right must be stated clearly.
For example, if the helper correctly identifies 90 of 100 labeled pictures, its accuracy on that test is 90 percent. That result does not tell us how well it writes music.
