While technologically unavailable (and maybe intractable) I've an impulse to say that in theory we could do something a little like Differential Privacy, where

While technologically unavailable (and maybe intractable) I've an impulse to say that in theory we could do something a little like Differential Privacy, where we look at the likelihood you got your output with the model you have vs the version of your model trained with (say) Rowling's work excluded.