What happened?

A group of researchers examined 14,419 self-published e-books on Amazon released between January 2023 and March 2026. Drawing on daily sales data from one of the five major U.S. publishers, the analysis ran the full text of each book through an AI detection tool called Pangram v3.3. Books were sorted into three groups based on their share of AI-generated content: none, light (up to 25%), and heavy (over 25%).

Books with heavy AI content made up 20% of the catalog but accounted for only 12.1% of sales and 11.3% of revenue. By contrast, books with no detected AI content made up 62.9% of the catalog and generated 72.5% of revenue. But according to the researchers, this picture misleadingly suggests that AI books are merely low-quality "junk" content sitting at the bottom of the market.

Why it matters?

Between the first quarter of 2023 and the first quarter of 2026, the total catalog grew 38.3-fold, and the number of books selling per quarter rose 19.2-fold. Quarterly revenue, however, grew only 8.9-fold. That means far more books are now competing for a much more slowly growing pool of revenue.

In six of eight genres, per-book revenue fell when comparing books published in 2023 and 2025 at the same point in their post-release lifecycle. Even looking only at books with no detected AI content, this decline showed up in seven of eight genres. The researchers say this finding cannot be explained simply by poorly selling AI books padding out the catalog, and points instead to a "dilution" effect — though they stress this is an observational correlation, not proof of causation.

The one exception was the Fantasy/Paranormal/Horror genre, where AI text arrived latest and gained the least traction; revenue for non-AI books in this genre rose 35%. In genres with heavy Kindle Unlimited usage, the revenue-share advantage of non-AI books was 8.4 points smaller than in genres with low Kindle Unlimited usage.

What we know

  • The share of new Top 25 entries with heavy AI content rose from near zero to 31%.
  • The proportion of non-AI books that stayed in the Top 25 from one quarter to the next fell as low as 28% at one point, settling around 62% by the end of the study.
  • Of 385 author identities that increased output after their first AI-flagged book, 287 raised their monthly production.
  • The top-earning pen name made $1.7 million before platform fees from eight books.
  • The single highest-earning heavy-AI-content book sold 80,431 copies and brought in $643,000.

What's next?

The researchers used the Allen Institute for AI's infini-gram tool alongside the Google Books index to measure how closely top-earning AI books overlapped with rare linguistic patterns. In the top 50 heavy-AI-content earners, these rare phrases made up 45% of the text, compared with 37.7% in the top 50 non-AI earners and just 19.1% in award-winning or award-nominated fiction.

The findings could carry weight in copyright litigation. In Kadrey v. Meta, Judge Vince Chhabria ruled in Meta's favor in June 2025 but warned that using copyrighted books to build a product generating billions of dollars in revenue could unleash an endless stream of competing works. At the time, plaintiffs could not provide empirical evidence of this market dilution; the new study supplies data that fills that gap. Amazon's failure to disclose AI content to customers, and its policy of merely capping daily publications at three, make it harder to address the problem at the platform level.