it seems that they also use a secondary model to filter it's responses, which can be seen where it starts talking about a topic... and then censors itself. you can see this behavior in gemini and ChatGPT too.