Three fault lines now divide the fair-use defense for training AI, and the first federal appeal has been argued but not yet decided.