For anyone earning a living with words, one figure stands out: from the first quarter of 2023 to the first quarter of 2026, Amazon’s self-published catalog expanded 38.3 times over. Revenue, meanwhile, rose just 8.9 times.
The distance between those two numbers is the entire story. Vastly more books are now competing for a pot of money that, by comparison, hardly moved.
A fresh analysis covering 14,419 randomly sampled self-published e-books published between January 2023 and March 2026 argues that AI-generated titles are not succeeding because they’re better. They’re succeeding because there are so many of them. And the books absorbing the damage are the ones where no AI text was detected whatsoever.
This isn’t scraped guesswork
Earlier efforts to size up the problem tended to sample brief book previews and estimate from there. This study took a different route.
Each book was classified on its complete text using the Pangram v3.3 detector, which its developers say has a false-positive rate of 0.04 percent. Pangram 4 has shipped since. Titles were sorted into three groups according to how much of the text was flagged as machine-written: none, light at up to 25 percent, and substantial above 25 percent.
The sales side drew on an internal dataset maintained by one of the five major US publishers, which follows roughly 500,000 Amazon titles and, per the researchers, accounts for around 95 percent of all e-books sold on the platform each day.
The headline figures look comforting — and they mislead
Skim the topline and AI books appear to be failing. Titles carrying substantial AI content account for 20 percent of the studied catalog yet capture only 12.1 percent of sales and 11.3 percent of revenue. Books with no detected AI text represent 62.9 percent of the catalog and collect 72.5 percent of revenue.
So the slop theory is dead? Not so fast.
The count of titles selling each quarter climbed 19.2 times, set against that 8.9x growth in revenue. Every slice shrank — including those belonging to writers doing the work themselves.
Six of eight genres lost ground
Measuring 2023 releases against 2025 releases across an identical post-release window, revenue per book declined in six of eight genres.
Restrict the view to books with no detected AI text and the picture darkens: revenue slipped in seven of eight.
That result dismantles the convenient counterargument. The decline can’t be pinned on a wave of worthless AI titles dragging the average down, since human-written books are pulling in less money on their own terms. The authors label the phenomenon “dilution,” while emphasizing that their comparisons are observational and associational rather than experimental proof of causation.
A single genre went the other way. Fantasy/Supernatural/Horror, where AI text arrived last and gained the least traction, saw revenue per book rise 35 percent for titles with no detected AI text. The researchers cite that exception as evidence against attributing the trend to some general market downturn.
Kindle Unlimited genres display it most starkly
Where Kindle Unlimited availability runs high and readers draw from a single shared subscription pool, the revenue-share advantage enjoyed by books with no detected AI text is 8.4 percentage points narrower than in low-availability genres.
The researchers attribute that gap to genre-specific characteristics and expressly decline to lay it at Kindle Unlimited’s door.
Bestseller rankings shifted as well. Fresh Top 25 entries containing substantial AI content rose from close to zero to 31 percent over the study window. Churn at the summit accelerated: the proportion of books with no detected AI text keeping a Top 25 slot from one quarter into the next dropped to roughly 28 percent at one stage before finishing near 62 percent.
A small group of accounts drives most of the flood
Thousands of hobbyists aren’t behind this. Among 385 author identities that released more titles with substantial AI content following their first AI book, 287 increased their monthly output afterward.
Gross revenue before platform fees for the highest-earning pseudonym reached $1.7 million across eight titles. The single top-grossing book with substantial AI content earned $643,000 from 80,431 copies sold.
That squares with the New York Times report about “Coral Hart,” said to have released more than 200 romance titles under 21 pen names within a year, selling roughly 50,000 copies. Nor are books the only arena. One man scammed millions of dollars through streaming platforms using AI-generated songs.
The rare-phrase test is where it gets uncomfortable
To measure how far the language in successful AI books overlaps with existing works, the researchers pointed the Allen Institute for AI’s infini-gram tool at the Google Books index.
Their search targeted rare expressions turning up in five or fewer Google Books volumes and completely missing from a 4.7-trillion-token web snapshot. Wording that distinctive implicates published books, not the open internet.
Across the 50 top-grossing titles with substantial AI content, such rare expressions accounted for 45 percent of the text. The figure was 37.7 percent for the top 50 books with no detected AI text, and 19.1 percent for award-winning or award-nominated fiction.
Inside the AI books, overlap rose 7.6 percentage points with every tenfold increase in revenue. Books with no detected AI text showed no equivalent correlation. The technique cannot pinpoint the origin of any single passage or establish that a particular book was copied — it captures aggregate language overlap and nothing beyond that.
Why the courts should care
Speaking to Ai2, Chakrabarty noted that an AI detector delivers only “an estimate—a score for how likely a passage is to be synthetic,” with no indication of where the language originated. But when a questionable text additionally contains rare expressions missing from the web and present in only a few books, one “can say with some confidence that it was taken from books.”
Evidence of that sort “acts as circumstantial evidence that supports an AI detector score” and at the same time “helps debunk some hackneyed arguments that liken human reading of books to AI being trained on books” — a defense AI companies routinely reach for in copyright disputes.
Judge Vince Chhabria sided with Meta in Kadrey v. Meta in June 2025, but paired the ruling with a sharp caveat. He said he found it hard to imagine that feeding copyrighted books into a product that generates billions in revenue, while spawning a potentially limitless flood of competing works, would count as fair use.
The plaintiffs in that case brought no empirical evidence of market dilution. This study supplies exactly what was missing: books with no detected AI text making less money as AI titles flood in.
Amazon has the answer and isn’t sharing it
Anyone publishing through Kindle Direct Publishing is required to disclose AI involvement. Amazon does not relay that disclosure to shoppers.
Which means the one party holding a clean record of which books are machine-written keeps it in-house while customers browse blind. For years the platform has struggled with AI titles appropriating the names and styles of established authors, and its principal remedy has been a limit of three publications per day.
The language-overlap results align with a November 2025 study demonstrating that language models can reproduce passages from copyrighted books almost verbatim, and with a second fall 2025 paper showing that two books suffice to fine-tune a model on an author’s style.
Write in a genre saturated by Kindle Unlimited and watched your per-title earnings sink since 2023 without changing your output or your quality? The numbers now back up the suspicion. Bring them to a lawyer — Chhabria came close to spelling out exactly what evidence a plaintiff needs.