AI captions do not conform to #WCAG they usually require a lot of work to make them usable
The highest-risk areas for real harm are: (1) named entities/numbers, (2) multi-speaker attribution, (3) timing/readability, and (4) hallucinated “corrections.”