How far can we take fair use in the context of AI? Do data mining exceptions provide enough cover to train foundation models?