@thelinuxEXP "Unlearning" isn't feasible that I know of, but there is a kind of "negative learning" where the model will try to steer clear from the provided material. I believe that's what is used to avoid most sensitive output from the current public models.