I'm still waiting for a proof that they don't use user data (directly or derived) for training.