used to be really useful for detecting text written with the same model, as it was high probability... unfortunately the probabilities are messed up by RLHF.