The prose in modern interactive fiction — the paragraphs, the interiority, the sensory detail — is written in real time by a large language model, prompted and constrained by the platform that runs it. In other words: a human sets the rules, the model produces the next paragraph, the reader inputs, and the loop continues.
What the platform actually does
- Prompt engineering. The platform translates the reader's preferences — spice, pacing, trope stack, dynamic — into a prompt the model can use.
- Memory management. It stores who the FMC is, who the love interest is, what happened in prior chapters, and feeds the relevant pieces back to the model.
- Voice constraint. It keeps the output in a novel register rather than a chat register, and enforces genre norms (first-person POV, slow-burn pacing, etc.).
- Content guardrails. It enforces hard limits (no minors, no non-consensual violence) without over-restricting the adult reader.
Why output quality improved
Two reasons. Base models have become meaningfully better at long-form prose with emotional interiority. And the platforms have become meaningfully better at the wrapper — memory, prompts, constraint systems, retry mechanics. A reader running a novel-length arc today gets a dramatically more coherent experience than two years ago.
What AI interactive fiction still doesn't do well
- Surprise. AI is patient, not mischievous. Genre-satisfying, not subversive.
- Century-long world-building on its own. Platforms can hold tens of chapters of memory, not hundreds.
- Jokes. Wit is hard. Don't expect Sally Thorne banter.
What it does very well, specifically for dark romance
- Sustained FMC POV across chapters.
- Slow-burn pacing that the reader controls.
- Memory of dynamics, preferences, past scenes.
- Aftercare on request.