PI & Other Tales

There’s been a lot of talk recently — including this article by The Bookseller — about when AI models might be able to write a bestselling novel.

It’s a compelling idea, especially with the rapid advances in large language models (LLMs). But there’s a hidden trap in that question, one that reveals how we often misunderstand what these models are actually doing.

The problem isn’t creativity. It’s context.

LLMs don’t write like humans. And more importantly, we don’t often ask them to. If we treat a model like it can do anything just because it’s a computer, and then judge it by human standards, it’ll fail — not because it’s incapable, but because we haven’t asked it to operate under human rules.

One of the hardest tasks for a language model is generating cohesive, long-form narratives. By the time you get 60,000 to 70,000 words deep into a novel, the story’s early beats are often lost to the model.

Characters forget what they know. Plots fray. The ending doesn’t land. That’s not a failure of intelligence — it’s a failure of memory architecture. As writers know, even we forget what we wrote in Chapter 2 by the time we get to Chapter 20.

So we asked: What could be… differently?

We’ve been experimenting with new ways to manage narrative continuity in generative systems. Not just for books, but for interactive fiction and adaptive gameplay. To support our upcoming gaming platform, we built a choose-your-own-adventure authoring tool that could generate over 100 distinct narrative arcs — each one coherent, self-contained, and consistent with a fixed world state.

Rather than starting from scratch, we fed the system a completed manuscript — Fortunes Told (A Voyager’s Guide to Life Between Worlds), a new novel published by our in-house imprint, Adventures of the Persistently Impaired (…and Other Tales). The goal was to generate branching versions of that novel’s storyline: alternate decisions, consequences, character paths — all while retaining the original story’s tone, logic, and internal consistency.

That meant solving two problems at once:

How do you fragment a single linear story into hundreds of possible branches?

And how do you rebuild those branches into fully contextualized narratives that still feel authored, with complete, accurate and continuity with a continuation indistinguishable from the original work?

And this what we did: