2026 Data Report
What 13,683 AI-Written Books Taught Us About Quality
Every AI book quality study we could find analyzes somebody else's books from the outside. We run a book-writing platform, so we did the opposite: we opened our own database. Here is what 13,683 books, 172 million generated words, and 100 tagged support conversations say about where AI books succeed — and exactly where they fall apart.
Data window: December 31, 2025 – July 12, 2026 · First-party platform data
- 13,683
- books created on the platform
- ≈172M
- words generated across 74,916 chapters
- 18%
- of started-in-app books reach a finished manuscript
- 34,000
- median words in a completed AI novel
- 3.6%
- refund rate across 1,456 transactions (Jan–Jun 2026)
The completion funnel: where AI books actually die
Of the 8,378 books started from scratch in-app this year, 5,337 generated at least one real chapter, and 1,474 finished every chapter they planned — 1,102 novels and 372 non-fiction books. That is an 18% end-to-end completion rate, which sounds brutal until you compare it to unassisted writing, where most sources put finished-manuscript rates in the low single digits.
The interesting part is not the rate. It is where the drop-off happens. AI books do not die in the messy middle the way human drafts famously do. They die at two specific cliffs, both near the very start:
The zero-chapter cliff — 3,041 books (36%)
More than a third of books never generate a single chapter. These writers built a premise, often an outline and characters, then stopped before any prose existed. The blocker is planning fatigue, not writing quality — there was no writing yet to be disappointed by.
The one-chapter cliff — 3,438 books (41%)
The largest bucket in the entire dataset stalls at exactly one chapter. Chapter one is where the writer first confronts the gap between the book in their head and the prose on the screen. Many regenerate, tweak, and abandon rather than push into chapter two.
| Chapters with content | Books |
|---|---|
| 0 | 3,041 |
| 1 | 3,438 |
| 2–5 | 294 |
| 6–10 | 418 |
| 11–20 | 539 |
| 21+ | 648 |
Look at the shape of that table. Only 294 books died between chapters two and five — the smallest bucket in the distribution, smaller than the 648 books that ran past chapter twenty. Books almost never collapse mid-manuscript. They either fail to survive chapter one, or they get finished. Since nearly every finished book planned more than one chapter, clearing chapter two flips the odds from roughly one in five to roughly three in four.
The practical lesson: the fight is won or lost at the outline stage, before a single chapter exists. Writers who invested in a chapter-by-chapter plan they actually believed in pushed through the chapter-one wobble; writers with a two-line premise did not. If you want to pressure-test a book idea before committing credits to prose, the free outline generator produces a full chapter plan you can argue with — and inside the app, the same outline stage is where you should spend the most editing effort, because everything downstream inherits it.
What a finished AI book looks like in 2026
The 1,474 completed books give us the clearest picture yet of the AI-assisted manuscript as an artifact — how long it runs, how it is structured, and what genres it lives in.
| Metric | Fiction | Non-fiction |
|---|---|---|
| Books completed | 1,102 | 372 |
| Median planned chapters | 20 | 8 |
| Median finished length | ≈34,000 words | ≈30,000 words |
Across all completed books the average length is ≈44,000 words, and the average generated chapter runs ≈2,300 words. A median 34,000-word novel is real book territory — but it is short-novel territory. Readers expecting a 90,000-word epic fantasy from twenty AI chapters will need to plan more chapters or expand scenes in revision.
Genre distribution mirrors what actually sells in self-publishing rather than what looks respectable. Romance leads with 3,169 books, followed by fantasy (2,106), erotica (1,516), thriller (1,346), mystery (1,232), YA (1,046), adventure (980), sci-fi (949), self-help (834), and literary fiction (722). The commercial genres that dominate Kindle Unlimited dominate here too.
Finished manuscripts leave the platform. Product analytics logged 7,345 export completions — 5,508 fiction, 1,837 non-fiction — by roughly 1,100 distinct users. Print-ready PDF is the top format (2,211 exports), then EPUB (1,941), DOC (1,718), and screen PDF (1,473). More than half of exports target print or e-reader formats, which says these books are heading to publication, not to a drawer. On the visual side, 2,252 books carry covers, generated across 8,277 AI cover attempts — about 3.7 generations per finished cover, a number that will make sense once you read the complaints section.
Some of that output is public and checkable: 470 books are showcased with their authors' consent on our books written with AIWriteBook page, and 38 of them carry live Amazon buy links. Authors publishing those on Amazon face a disclosure question we get asked constantly — our breakdown of the KDP AI disclosure policy covers exactly what Amazon requires you to declare and what it doesn't.
Where AI books actually fail: 100 support conversations, tagged
Averages hide failure modes, so we read the 100 most recent support conversations mentioning output quality and tagged every distinct complaint. Every single complaint came from a paying user — people deep enough into a manuscript to care. Here is the full tag distribution, unfiltered:
| Complaint category | Mentions | What actually fixes it |
|---|---|---|
| AI not following instructions (word counts, names, placements) | 6 | Make changes through the chapter AI chat, which shows every proposed edit as a diff you accept or reject — so a "change" that changed nothing is visible before it costs you anything downstream |
| Chapter prose quality (typos, grammar, repetitive phrasing) | 4 | Run the built-in grammar and style check before export, and use writing-style copy so chapters follow your voice instead of the model's defaults |
| Import losing chapters or structure | 4 | Use consistent chapter headings before importing — the failure case we traced was a 59-chapter manuscript split across 7 parts with mixed numbering, of which only 44 chapters were detected |
| Outline stripping user-provided detail | 3 | Edit the outline directly rather than regenerating it — regeneration re-summarizes and loses specifics you typed in |
| Editing bugs (edits not applied, highlighted text deleted) | 3 | These were bugs on our side, not user error; the diff-based chat flow now makes silent no-ops visible |
| Cover text legibility and placement | 3 | Iterate placement and typography in the cover designer instead of regenerating the full image — regeneration rerolls the art too |
| Cross-chapter contradictions | 2 | Fix continuity at the outline level first; chapter-level patches create the contradictions users then report |
| Wrong output language (asked French, got English) | 2 | Set the book's writing language once at book level rather than re-instructing per chapter |
The verbatim complaints are more instructive than the categories. The most damaging pattern is formulaic scaffolding: one refunded non-fiction author counted nearly every chapter opening with "The previous chapter established…". Another writer asked for 2,000-word chapters and consistently received 2,700 to 3,700. A French author reported the AI writing English "25 times out of 30" despite explicit instructions. None of these are hallucination horror stories — they are instruction-following and repetition failures, the unglamorous middle of AI quality.
Two product decisions came directly from this analysis. First, the chapter AI chat presents every revision as an accept/reject diff, because six of the tagged complaints reduce to "the AI said it changed something and didn't." Second, writing-style copy lets the model extract and follow your voice from an existing manuscript, which targets the repetition complaints at their source. If you want the longer treatment of what separates readable AI prose from slop, our essay on whether AI-written novels are any good digs into the craft side; and if you are revising a draft and want a cold read on how machine-flavored it still sounds, the free AI content detector gives you a signal readers would otherwise give you in reviews.
What the refund data adds
Complaints are cheap; refunds are costly, so they are the harder signal. Between January and June 2026, 57 users requested refunds against 1,456 paid transactions — a 3.6% refund rate (52 refunds issued). Among stated refund reasons, repetitive or AI-slop output accounted for roughly 11%, and import or outline-fidelity failures for roughly 16%.
Read together: quality complaints are real but they are a minority of refunds, and fidelity failures — the tool not honoring what the writer already wrote — cost us more than prose quality does. Writers forgive imperfect sentences they can edit. They do not forgive a tool that discards their work.
The counter-intuitive finding
Fiction and non-fiction finish at almost exactly the same rate
Non-fiction books plan a median of 8 chapters. Novels plan a median of 20 — two and a half times the distance. You would expect non-fiction to finish far more often. It doesn't: 372 of 3,307 non-fiction books completed (11.2%) versus 1,102 of 9,759 novels (11.3%). Statistically identical.
That kills the most obvious theory of AI book abandonment — that books fail because they are long. They fail before length matters, at the zero- and one-chapter cliffs, and those cliffs are about the writer's conviction in the plan, not the page count ahead. Halving your chapter plan will not save a book you don't believe in.
What this means for you
- 01
Spend your effort on the outline, not chapter one
The data says books that clear chapter two almost always get finished. The chapter-one wobble is survivable when the outline underneath it is specific enough to trust. Argue with your chapter plan until it convinces you — then write.
- 02
Budget a revision pass; the raw draft is not the book
Every quality complaint in our support data came from a writer actively editing. That is the correct workflow, not a failure of it. Plan for a human pass on repetition, chapter openers, and voice — and use diff-based AI edits so you can see exactly what changed.
- 03
Calibrate length expectations to the genre
A median finished AI novel runs ≈34,000 words. Fine for novellas, romance serials, and non-fiction guides; short for epic fantasy. If your genre expects 90,000 words, plan 35–40 chapters, not 20.
- 04
If you import a manuscript, check the chapter count immediately
Import fidelity failures drive more refunds than prose quality. Clean, consistent chapter headings before import — and count your chapters after — prevents the single most painful failure in the dataset.
- 05
Distrust any AI writing claim without a completion number
Marketing shows finished books; the funnel shows 18%. Any tool, ours included, should be judged on how many started books become finished manuscripts — it is the one metric that summarizes everything else.
The honest summary of 13,683 books: AI collapsed the cost of producing a manuscript, and in doing so it moved the bottleneck. The hard part is no longer generating 34,000 coherent words — 1,474 people did that this year, and 7,345 exports went out the door. The hard part is now the same as it always was: committing to a plan, surviving the first chapter, and revising like a professional.
This report is one piece of our larger AI writing assistant guide, which covers the full workflow from idea to export — including the outline and revision practices this data argues for.
Questions this data can actually answer
Start a Book Using This Data
The funnel says outline first, revise always, and get past chapter two. Start free, build a chapter plan you believe in, and let the diff-based editor handle the revision pass this report proves you'll need.
Free to start — no credit card required.