Copyright for me not for thee? :) That's a good point though. Maybe they could round trip things? E.g., use the model trained only on internal content to generate training data (which you could probably do some kind of screening to remove anything you don't want leaking) and then train a new model off just that?