>>3047203 (pb)

Yeah I think that's the process. The main point is that the algo can only filter whatever is marked as "offensive" in the training dataset. So in the end whoever flags the images in the dataset as "offensive" decides what will be filtered and what not