This paper revisits the question of provable copyright protection for generative models: identifying limitations of prior work, proposing new definitions, and describing conditions under which differential privacy suffices.