[s1artibartfast]

needing the training data has zero bearing on if they are derivative works. "derivative works" it a term of art with a specific meaning.

I think the derivative work argument is a dead end. However, AI companies did violate use licenses when they first used the data for commercial purpose of training the models.