I've seen work on explainability, some of that might translate over. I would unironically like to see what happens when you train ML on copyright cases to try to derive a "fair use classifier". It's probably going to be an incoherent mess (like law itself) but it would be interesting for sure