This sounds like what Spirit AI tried to do. I wonder how they solved the problem of limited training data. Or perhaps the result is an indication that they didn’t.
they were originally planning to get around the implementation problem through generated text, but axed that because they couldn’t guarantee the language model wouldn’t give offensive responses
It seems to me there is a middle ground here where a language model generates lots and lots of plausible responses, which are then curated by humans, and the good ones are hard coded into the final game.