I think there may be different ways to do that, from new models (as you wrote) to simple filters. For example, I couldn't get ChatGPT to output any real example

I think there may be different ways to do that, from new models (as you wrote) to simple filters. For example, I couldn't get ChatGPT to output any real examples of hate speech. They may have a filter that picks up candidate responses with hate speech and blocks them.