AI-Generated Books Flood Amazon and Crush Revenue for Human Authors

Researchers analyzed 14,419 randomly selected self-published e-books released between January 2023 and March 2026, using daily sales data from an internal dataset of one of the five largest US publishers, covering about 95% of all daily e-book sales on Amazon. Unlike previous studies, they classified each book based on its full text using the Pangram v3.3 detector, which has a reported false-positive rate of 0.04%, categorizing books by the fraction of AI-generated text: none, light (up to 25%), or substantial (over 25%). Books with substantial AI content made up 20% of the catalog but generated only 12.1% of sales and 11.3% of revenue, while books without detected AI text made up 62.9% of the catalog and generated 72.5% of revenue. However, between Q1 2023 and Q1 2026, the cumulative catalog grew 38.3-fold, the number of titles selling per quarter grew 19.2-fold, but quarterly revenue grew only 8.9-fold, meaning more books compete for a slower-growing revenue pool. Revenue per book fell in six of eight genres when comparing 2023 and 2025 releases over the same post-publication window, and even for books without detected AI text, it fell in seven of eight genres, ruling out the explanation that the average is dragged down only by poorly selling AI books. The only exception was the fantasy/supernatural/horror genre, where AI text gained traction latest and least, and revenue per book for non-AI works rose 35%, arguing against a general market trend. The patterns were stronger in genres with high Kindle Unlimited availability, where the revenue share advantage of non-AI books was 8.4 percentage points lower than in low-availability genres, though researchers saw no causal link to Kindle Unlimited itself. The share of new top-25 entries with substantial AI content rose from near zero to 31%, and the top ranks churned more, with the retention rate of non-AI books in the top 25 falling to around 28% at times before settling at about 62%. The output is concentrated among a few prolific authors: of 385 author identities who published further substantial-AI books after their first, 287 increased their monthly output, with the top-selling pseudonym earning $1.7 million in gross revenue from eight titles, and the top-selling single substantial-AI book earning $643,000 from 80,431 copies. The study also found that the 50 top-selling substantial-AI books had 45% of their text covered by rare expressions appearing in at most five Google Books volumes and absent from a 4.7 trillion token web snapshot, compared to 37.7% for top-50 non-AI books and 19.1% for award-winning or nominated fiction, with the overlap increasing by 7.6 percentage points per tenfold increase in revenue for AI books. The researchers emphasize that the method measures aggregated linguistic overlap, not specific copying. Tuhin Chakrabarty of Ai2 said the findings support detector scores as evidence and counter arguments equating human reading with AI training on books, directly relevant to copyright cases like Kadrey v. Meta, where the plaintiffs lacked empirical evidence of market dilution. Amazon does not disclose AI involvement to customers, despite authors having to declare it on Kindle Direct Publishing, and the platform has primarily limited to three publications per day to combat AI spam.