Comment by dangson

But if they're able to detect the unwanted behavior why not use that to filter out conversations they don't want in the training data?