“AI companies have engaged in a form of legal arbitrage by training models on copyrighted data, anticipating that any potential legal damages will be insignificant compared to their revenues.”