I'm not talking about distributing the model, just its output, which won't be something the user can non-creatively transform into a Jay-Z sample.

However, if the model really can't regenerate its training data no matter how hard you try, then that would be fine too. I don't think anyone can really guarantee that now so that might be a problem, but it isn't necessarily a problem.