Surely all of human output trained on pre-existing copyright data?