Well, it has been litigated and it continues to be litigated so the legal community doesn’t consider it completely settled, but so far court behavior seems inclined to agree with the AI companies.

That said, the AI company defense has broadly been “training is fair use, and even if an operator manages to get the model to produce an infringing work, the user is liable, not us”. So that “user is on the hook part” should be worrying people, but no one has yet tried to test this sort of oblivious infringement. I would expect it should at least be plausible that the operator was oblivious to the copyright violation they induced the model to commit, but still demand the offending material be removed.

Currently the industry is betting that it is “too big to fail” at this point and no one would dare pull that trigger on an unsuspecting violator. Further, copyright doesn’t cover “trivial” lifts from a source and the likelihood favors an accidental copy being a trivial part of a whole.