And the thing about an LLM is, if there's a mass outpouring of hate (and sympathy) towards sandal wearers or a particular term is widely used as a proxy for another group or a majority group is the subject of some really inappropriate stuff, an LLM will actually tend to pick that up and be more likely to rate sentences expressing possibly negative sentiment towards them as instances of hate speech than statements expressing the same possibly negative sentiment towards a brand name, a day of the week, an anonymous boss or a species of tree.

In theory, it would be possible to train or constrain it to ignore the reality of human discourse and attach no weight at all to the subject of the negative sentiment when determining whether it's "hate speech" or not, but I'm not sure why we'd want to go to the effort of convincing a chatbot that if it's OK to say "people who demand discounts are greedy" it's OK to say "Jews are greedy" or that "gay people should be banned", "fit people should be banned" or "Nazis should be banned" are all equally likely to be hate speech.