An issue is that to substract the knowledge (note that I know this is not really how it works) you need a dataset of all the bad stuff. This means giving to only a few people the right to make the final round of models. And how to ensure that the « bad » dataset is not disseminated?
