That's already a thing by virtue of how "harmful" data is filtered during the training / data gathering phase. You can't expect to just remove a certain "fact" and not have its immediate precursor not show up in answers. You need to eradicate the entire chain to a certain depth, and after that, because many ideas lead to one idea (and LLMs are devoid or creativity or originality), you only have a few "winners" at the top. The long tail is always cut, and so the entire model converges to a ~~ziggurat~~ few ideas, that might have been pushed hard pre-training phase.