An Atlas of Generative AI for Fiction

This is not a paper. It is a tour, organized as ten short chapters plus a glossary, written for someone with no background in machine learning who would like to understand what the field is actually arguing about.

By the end, you should be able to read a paper title like Creative Preference Optimization or a product announcement like Sudowrite Muse and understand what problem it claims to solve, what the alternatives are, and what the field thinks about whether the claim holds.

The map is opinionated. It treats long-form serialized fiction — stories told across many episodes, with returning characters and arcs that pay off — as the most interesting unsolved problem in the field. That bias will show. But the chapters along the way are designed to stand on their own.

Contents

1. 01The Question Has ChangedWhy fluent ≠ creative, and why this field exists.
2. 02What 'Creativity' Actually MeansNovelty, value, surprise — and what fiction adds.
3. 03Why It's So HardMode collapse, RLHF, slop, and the hivemind.
4. 04The Five LeversPrompting, decoding, training, architecture, evaluation.
5. 05The Writers' RoomsFrom Dramatron to the 2026 long-form rooms.
6. 06How We Tell If It's Any GoodThe benchmarks, the long-form wave, and judging the judges.
7. 07The People and the ProductsWho's doing the work, and what they ship.
8. 08The Industry MomentThe 2026 WGA deal, Bartz, Shy Girl, and the detection wars.
9. 09The FrontierOpen problems, and the serialization wedge revised.
10. 10The BetPractical recommendations from the research.
11. §GlossaryAll the terms in one place.
12. ±ChangelogWhat changed, and when.

How to read it

Read the chapters in order if you can. Each one introduces ideas the next will assume. If you bounce around, the glossary at the back lists every term, paper, product, and acronym that shows up along the way.

Every chapter ends with a Sources & threads section listing the underlying papers, with arXiv numbers and links so you can pull on any thread that interests you. The citations are the point: this atlas is just a map of where to dig.

A note on the date

This was first drafted in May 2026 and revised in November 2026. The field moves fast enough that any field map will be partly out of date by the time you finish reading it. Anywhere a claim depends on a recent paper, you'll see a date, and anywhere a claim is contested, you'll see the dissent. Where a revision overturned something the first edition said, the chapter says so plainly.

The biggest correction in the second edition: the first edition claimed almost nobody was working on serialized fiction. By autumn 2026 that was no longer true, and Chapter 9 now explains what has been claimed and what is still open. Every change is dated in the changelog.