You can find a list of the grand challenges here.

Solved: the SOTA model is better than human expert performance. Challenges with no human benchmark available will be decided on a case by case basis.

Minimal human oversight - the amount of doctor time devoted to examining medical images is a small fraction of what it was pre-AI. (Of course this will depend heavily on the regulatory environment, but that's unavoidable).

If there is strong evidence that the AIs are being used but getting worse results than humans, that does not resolve the question YES. We are unlikely to have such evidence one way or another because measuring hospital efficacy is hard, but if it does happen that AI doesn't count.