I don't think they can. I believe they just add filtering in the data pipeline before & after the generation. Think about how they currently filter out "bad t

I don't think they can. I believe they just add filtering in the data pipeline before & after the generation. Think about how they currently filter out "bad things" user input ---> | prompt filter | GPT | output filter | --> sanitized output