A few months ago someone here reported on making text adventures with language models. If I remember correctly, a problem is that it is not trivial to control the AI in a way that players can't cheat on puzzles.