I didn't say it wouldn't (or rather couldn't), I said it was unlikely for the selected hypothesis given standard training data vs RLHF iterations.