AI companies should be allowed to train on copyrighted content