AI output is obviously transformative and would be protected by fair use.

The argument is about whether the training data was stolen. (which it was)