I noticed that too, but my conclusion is that they probably hand-picked every image and description in their training data so that the final model doesn’t even know what the poor taste stuff is.