I suppose, but then the training data plagiarism moral question remains. Rate seems to be like 2-5% for what they can pin down: dl.acm.org/doi/10.1145/3543507.3583199 I’m guessing that means the hidden actual rate might be higher… and there are the high profile incidents.