Not really. The limitation is mostly in the training set. If you train an ML algo on a set of videos that you either banned or didn't, sure, when you run that i

Not really. The limitation is mostly in the training set. If you train an ML algo on a set of videos that you either banned or didn't, sure, when you run that in prod all you'll get out of it is "ban or no". But if the training dataset includes a ban reason, then the inferences will include reasons too. What is opaque is how the algo decided the thing was violating the policy. That applies to supervised learning. For unsupervised learning, you can still do this by requiring a valid rationale as part of the "win condition".