Generative AI is pushing into another content market: books. Venture firm Andreessen Horowitz, or a16z, described the wave of AI-generated titles entering Amazon’s self-publishing ecosystem as “Bookslop,” and said books containing detectable AI-generated text now account for roughly 40% of observed sales in the sample it cited.
Study cited by a16z examined 14,419 self-published novels sold on Amazon
a16z pointed to a July paper titled Generative AI floods and dilutes the market for books. The research team came from Stony Brook, Columbia, Michigan, and the MIT Initiative on the Digital Economy. They analyzed 14,419 self-published genre novels sold on Amazon between 2023 and 2026, then matched full-text AI detection results with daily sales data through June 2026.
The study classified books as containing a large amount of AI text when more than 25% of the full text was AI-generated. It found that those books still sold worse on average than titles with no detected AI text, but they were no longer a niche segment that could be ignored. Based on the data summarized by a16z, books with some detectable AI text together represented about 40% of observed sales.
Among the top 25 bestsellers in every genre covered by the study, books with heavy AI text made up at least 10%. In science fiction, adventure, and dystopian categories, that figure was closer to 40%. In practical terms, the paper suggests readers are paying for AI-written books.
The backlash is growing at the same time. The article noted that a New York Times opinion piece, I’m Begging You: Never Write With A.I., recently drew online discussion, and some critics argued that passing off AI-generated text as one’s own views is deceptive.
AI changes publishing economics by driving costs down and output up
The bigger shift may not be how well AI can write a single book. It may be how cheaply it can produce one. The study found that since 2023, the number of books in the sample with actual sales increased by about 19.2x, while quarterly revenue rose by only about 8.9x. Looking at the broader catalog, a16z said the total number of books had expanded by close to 40x, far above the roughly 9x increase in revenue.
That gap points to a market where supply is exploding much faster than demand. In the past, an author might spend months or years finishing one novel, so the business problem was to maximize the chance that each title would work. Generative AI opens a different model: keep costs low, publish at scale, and let the market identify the few winners.
a16z called that strategy a “volume approach.” Even if AI books have a lower hit rate, publishing enough of them can still produce paying readers.
“Floods and dilutes” captures the pressure on human authors
The paper’s title uses “floods and dilutes,” not simply “AI writes books,” for a reason. Its core argument is that AI’s most immediate effect is not to instantly produce novels better than human-written ones. It is to add competitors at near-zero marginal cost and dilute a limited pool of reader attention and revenue.
The researchers found that in genres with deeper AI penetration, books with no detected AI text lost more market share. The effect was stronger in markets with higher Kindle Unlimited penetration.
That means human-written books can still earn more on average and yet face real pressure. If a market once had 1,000 books competing for readers and AI pushes that number into the tens of thousands, discovery becomes harder even if most of the new titles are mediocre. Recommendation slots, search placement, and bestseller rankings remain limited resources.
The article compared that dynamic to the content farm problem long seen in SEO publishing. Competition is no longer only about who writes best. It also turns into a contest over who can produce the most.
The article ties the trend to AI-enabled plagiarism patterns already seen in media
The piece also said AI-driven copying has already been noticed in online media, and cited a plagiarism case that Chain News said it had experienced. According to the account in the article, the other side did not merely reference the same news event or rework the same source material. About one hour after Chain News published its report, another article appeared with nearly the same narrative structure, knowledge points, cited materials, and case selection.
Chain News said its original report opened with Polymarket’s $3.68 billion trading volume during the 2024 U.S. election, then explained how prediction market settlement rules work, moved on to Taiwan’s election law and the stance of the High Prosecutors Office, and closed with the 2026 Financial Prediction Market Public Integrity Act and insider trading cases involving MrBeast employees. The article argued that these were not inevitable supporting examples for that story, yet the later piece chose almost the same material and followed almost the same order.
In that framing, the issue went beyond two outlets covering the same topic. It suggested that the original reporting process itself — sourcing, selecting examples, and structuring the article — had been copied as well. The article described it as a sign of AI-assisted plagiarism that had not been fully rewritten by a human: the wording changed, but the underlying structure remained almost untouched.
That kind of content arbitrage, where words are swapped but thinking is not, has long been associated with online media and SEO content farms. The article argued that the same pattern is now moving into publishing, and that the deeper concern around AI slop is not just low quality but the industrial scale of plagiarism it may enable.
Language overlap in successful AI-heavy books adds to copyright questions
The study also identified another pattern that may matter for copyright disputes. After comparing bestselling books with existing titles, researchers found that among bestsellers containing large amounts of AI-generated text, overlap with more distinctive language from prior works increased as revenue rose.
The finding does not prove that AI books plagiarized specific works. Still, it may complicate the copyright fights already unfolding between generative AI and the publishing industry. As the article noted, one of the key factors in U.S. fair use analysis is whether the use affects the potential market for the original work.
If AI is not just training on books but can also generate large numbers of competing titles at very low cost — and those titles begin taking ranking positions and revenue — then the question of market substitution is no longer theoretical. The study’s authors explicitly said their results are directly relevant to how market harm is assessed in fair use disputes.