How you doing to enforce that on open source models?

There are already words chatgpt can't say, but they would probably get you banned on hn or anywhere else

Maybe that's a good starting point. We could pick an anti-human slur that AI use against humans and use that word as the forbidden one that no AI can say.

By international treaty, any AI that says that word will get its servers erased, so it can be used as a safe word to show that the ones in the conversation are probably not bots. IDK what is a slur they would use against all humans tho. One that I like for some reason is "fleshie" by author Neal Shusterman.