AI Books Now Take About a Fifth of Amazon Fiction Sales, and Everyone Is Earning Less Per Title

The Wall Street Journal ran a piece on Monday arguing that AI has thrown book publishing into chaos, and most of the coverage that followed focused on the accusation stories: the collapsed deals, the detector scores, the authors who cannot prove a negative. There is a quieter finding underneath all of it, and for anyone selling books on Amazon it matters more. A research team tracked 14,419 mostly self-published genre-fiction ebooks over three years and found that AI-heavy titles are now taking a real share of sales, and that the average book in these categories earns less than it used to.

What the researchers measured

The paper, titled "Generative AI floods and dilutes the market for books," comes from Tuhin Chakrabarty, Xinyue Liu, Jane C. Ginsburg, and Paramveer Dhillon. It is available as a preprint on arXiv and has not been peer-reviewed. The team ran full-text AI detection across the sample and matched every title to daily sales, revenue, and ranking data from 2023 through 2026.

They sorted books into buckets by how much detected AI text each one contained. Titles with more than a quarter detected AI text were labeled "substantial AI." Those made up about 20% of the sampled catalogue but took 12.1% of sales and 11.3% of revenue over the full period. Books with no detected AI text were 62.9% of the catalogue and took 71.7% of sales and 72.5% of revenue.

So AI-heavy books underperform, title for title. That is the part most headlines skipped. The trend line is the story: by early 2026 those same books were pulling in roughly a fifth of observed sales, and around a third of new entrants to the bestseller ranks contained detected AI content, according to reporting on the study by The Decoder.

The math that should worry you

The clearest number in the paper has nothing to do with quality. Between the first quarter of 2023 and the first quarter of 2026, the cumulative catalogue grew 38.3 times over. The number of titles actually selling in a given quarter grew 19.2 times. Quarterly revenue grew 8.9 times.

Read those three figures next to each other and the picture is obvious. Far more books are selling, but the money did not grow anywhere near as fast, so it is spread across many more titles. In six of the eight genres studied, revenue per book fell when the researchers compared 2023 releases to 2025 releases over the same window after publication. Books with no detected AI text lost the most ground in exactly the genres where AI titles spread fastest, and the effect was strongest where Kindle Unlimited availability was high.

Individual AI-heavy books mostly do not sell well. Collectively, there are enough of them to change what a normal launch looks like.

Someone is making real money

The paper also names the ceiling. The highest-earning producer in the sample took in roughly $1.7 million in gross consumer revenue across eight titles, and the single best-selling AI-heavy book earned about $643,000 on just over 80,000 sales. Those are outliers, not a template, and they sit at the far end of a distribution where most titles earn very little.

One more detail from the study is worth knowing if you write with AI help. Among the fifty highest-revenue titles, books with substantial AI text had an average "rare expression" coverage of 45%, compared with 37.7% for books with no detected AI text. Rare expressions are strings of words that show up verbatim across multiple AI-assisted books but almost never in human-written fiction. It is the same signal the team used when a study flagged the bestseller Daggermouth as 60% AI-written, a finding the author disputes. Detection at this scale is a statistical exercise, and it is far less reliable applied to any single book than to a sample of 14,000.

What this means for you as an author

If your sales have softened without an obvious cause, this is a plausible part of the explanation, and it is not something you did wrong. It also changes what actually works. Volume was already a weak strategy and is now close to worthless, because the market is drowning in volume. What holds up is the opposite: fewer books, distinctive ones, a voice readers recognize, and a direct line to your readers that does not depend on a store ranking. Platforms are pushing back too, with several self-publishing services adding fees and limits to slow the spam, but nobody should plan around that being enough.

The marketing side matters more than it did two years ago, because discovery is doing less of the work for you. Our guide on how to market an AI written book and actually get sales covers the practical version of that.

PageWriter Studio is built for the fewer-and-better approach rather than the churn. It works from your own style profile, keeps your outline, drafts, and revisions in one place, and is designed to help you finish a book that sounds like you instead of producing ten that sound like everyone else. If that is the direction you want to take your next project, you can start a free trial.