The legality of training models on copyrighted data without a licence is, or at least should be, independent of whether corporations decide to comply with copyright law.