I could see LLMs being useful as just an UI layer on top of traditional interactive fiction engines? Having a voice-operated IF assistant/narrator paired with tools for note taking and such could offer an accessible alternative interface to the large existing IF library (or maybe something like that exists already?).

Taken further the assistant's role could be expanded with greater autonomy. Games could include metadata and authors hints & directives for the narrator. Or a limited interface for manipulating the game state and shifting the traditional IF model closer to a two player collaborative game. Using the LLM to transform and extend pre-authored text might actually offer new narrative possibilities, assuming the model could do it subtly and well enough, the style, POV, mood an so on of prewritten content could be freely modified in some interesting ways. E.g introduce unreliable narrators, describe the player's actions from a point of a security camera or their dog, rewind and replay events from some other characters POV. Simulate psychological issues by subtly rewriting the player's (or the narrator's or NPCs) perception of the world and have them navigate a little twisty maze of emotional states, all different.

Restricting the model into operating strictly within the limits of human authored content could bypass some of the usual problems associated with purely AI generated games and the approach seems fitting for the genre and it's pre-existing rich tooling and talent.