Invest1 publisher3 min readPublished
Self-published genre fiction on Amazon added titles twice as fast as revenue
Two detection studies put machine-written text across large slices of Amazon's self-published shelf, and the sales figures underneath one of them show a market that grew in dollars while thinning out per title.
The Investor · Invest desk
What happened
- Originality.AI counted 1,272 titles likely written by large language models among more than 2,000 religion books published on Amazon over 12 months, or 63% of that category.
- A Stony Brook University-led study ran 14,419 self-published genre-fiction books sold on Amazon between 2023 and 2026 through Pangram and flagged 2,880, about 20%, as more than a quarter AI prose.
- Across those three years the number of books with observed sales in a quarter rose 19.2-fold while quarterly revenue rose 8.9-fold.
- Hachette pulled the horror novel "Shy Girl" from its U.S. release in March after allegations that author Mia Ballard used AI to write it, and Ballard denied using AI for the novel.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- constraint Clicks on machine-written titles raise those titles in Amazon's own rankings, so an unknown author's earnings turn on placement in a listing pool, and a better manuscript alone leaves the placement where it was.
- decision Any tightening of Amazon's supply-side rules starts from a cap that already permits 1,095 titles a year per compliant account, so a lower daily limit would have to fall a long way to bind.
- exposure Authors' recovery to date is priced against training use in the $1.5 billion Anthropic settlement, leaving the per-title revenue decline unaddressed by any of the litigation on the record.
- contradiction Total quarterly revenue in the segment is nearly nine times what it was, so the category is expanding while the researchers describe a squeeze, and their study has not yet been peer-reviewed.
Dividing 8.9 by 19.2 gives 0.46, so the average selling title in the third year earned about 46% of what the average selling title earned in the first, a fall of roughly 54% [6][1]. Titles grew about 2.2 times as fast as dollars [2]. Total quarterly revenue in the segment is still nearly nine times what it was [6]. The category grew, and it spread that growth across a much wider shelf.
Inside the flagged 2,880, the weight sits at the top end: 2,168 books were more than half machine-written, nearly 1,000 were more than 90%, and 712 fell in the 25-to-50% band [3][4][3]. Majority-machine titles are 75% of the flagged set [4].
The religion figure comes from a different tool on a different category, so the 63% and the 20% do not sit on the same scale. Originality.AI flagged 1,272 of more than 2,000 titles published over 12 months, 63%, which leaves roughly 750 unflagged [2][7].
Amazon requires self-published authors to disclose AI use and limits each to three title uploads a day [9]. Three a day is 1,095 a year and 3,285 across the three years the study covers, more than the 2,880 titles it flagged, with 405 uploads to spare [5].
Tuhin Chakrabarty, a study coauthor and assistant professor of computer science at Stony Brook, said the "algorithmic bubble" of online marketplaces is part of the reason human authors get squeezed out [11]. Clicks on machine-written titles push those titles up the page, according to Chakrabarty, and an unknown author's book "might not even show in the related books listing because there is so many AI-generated books that are surfacing based on the keyword and the search terms associated with them" [13][12]. His coauthor Paramveer Dhillon, an associate professor of information at the University of Michigan, told Fortune that "due to this increased competition, they're likely to get squeezed, and their revenues are reduced" [8].
An Amazon spokesperson told Fortune that "our content guidelines govern which books can be listed for sale, and we take proactive and reactive measures to prevent, detect, and remove content that violates those guidelines, whether AI-generated or not" [10].
The study has not been peer-reviewed [5]. Its two growth figures cover the whole 14,419-book sample, machine and human together, so they measure dilution of the average title and not a transfer from human authors to synthetic ones. Fortune's account does not report false-positive rates for either detector, and a 5% rate applied across 14,419 books would be 721 titles, a quarter of the flagged count [6]. I'd expect the crowding-out effect to be real and considerably smaller than 54%: detector error inflates the flagged share, some of the per-title revenue decline belongs to price and length mix among new listings, and the flagged books that sell well are competing with each other as much as with anyone's debut novel. A reviewed version that splits revenue between flagged and unflagged titles would settle it.
Authors have sued OpenAI, Meta, Anthropic and xAI over the use of their books as training data [14]. In July a federal judge approved a $1.5 billion settlement ordering Anthropic to pay thousands of authors after it downloaded and stored millions of pirated books to train Claude [15]. That settlement pays for training inputs, and the revenue the researchers are counting is lost at the point of sale.
What to watch
- Peer review of the Stony Brook study, and whether the reviewed version splits revenue between flagged and unflagged titles.
- Whether Amazon changes the three-titles-a-day upload cap or publishes enforcement counts for its AI disclosure requirement.
- Published false-positive rates for Pangram and Originality.AI, since a few points of detector error move the 20% and 63% shares.