It would obviously be very bad if those decisions were being made based on the statistical weight of the training corpus of a general large language model.