AI Books Are Flooding Amazon: What a New Study Actually Found

If you’ve spent any time in Amazon’s Kindle store lately, scrolling through romance, fantasy, or thriller listings, you may have noticed something a little off. Covers that look almost right but not quite. Blurbs that promise a lot and say nothing specific. Author names you’ve never seen before attached to eight books published in the same month. You’re not imagining it, and now there’s a study that puts a number on it.

A team led by researchers at Stony Brook University and the University of Michigan checked more than 14,000 self-published genre-fiction titles sold on Amazon between 2023 and 2026. Their finding: roughly one in five of those books contains substantial AI-generated text, and nearly 1,000 of them are more than 90% written by a machine. At the same time, the study found something that should worry human authors even more than the raw percentage — as AI-generated books have flooded the marketplace, sales volume has grown far faster than revenue, meaning the average book in that category is earning less than it used to.

This is the honest version of that story: what the study measured, what it found, what it stops short of proving, and what you can actually do about it the next time you’re deciding what to read.

Who Ran the Study, and What They Actually Looked At

The research comes from Tuhin Chakrabarty, an assistant professor of computer science at Stony Brook University, and Paramveer Dhillon, an associate professor of information at the University of Michigan. Their sample wasn’t “every book on Amazon” — it was specifically self-published genre fiction: the romance, fantasy, and thriller titles most likely to be written quickly, published cheaply, and sold in volume. They tracked that category across three years, from 2023 through 2026.

To flag which books contained AI-generated text, the researchers used a detection tool called Pangram. Pangram works by scoring how closely a piece of writing matches patterns the tool has learned to associate with AI-generated prose. That’s worth sitting with for a second, because it matters for how you read every number that follows: a Pangram score is a probability, not a confession. It tells you how much a text resembles AI writing, not a verified account of how that specific book was actually produced. It’s the same caveat we applied when this same research team’s tool flagged a specific bestselling novel earlier this year — more on that case below.

One more thing worth knowing before the numbers: this study has not yet been peer-reviewed. That doesn’t make it wrong. It means the findings haven’t been independently checked by researchers outside the team that produced them, and it’s the reason we’re treating every figure here as “what a new study found” rather than “proven fact.”

The Numbers

Here’s what the study actually reports, cleared of rounding:

• 14,419 self-published genre-fiction books were studied, sold on Amazon between 2023 and 2026.
• 2,880 books (20%) had what the researchers call “substantial” AI-generated text — more than a quarter of the manuscript.
• 2,168 of those books were more than half AI-generated.
• Nearly 1,000 novels were more than 90% AI-generated — books that are, by this measure, barely touched by a human hand.

Those numbers alone are striking. But the part of the study that should get more attention than it has is what happened to revenue over the same period.

The Real News: Volume Is Up 19.2x, Revenue Is Only Up 8.9x

Among books with observed sales, quarterly volume grew 19.2 times over across the three years studied. If the market were healthy in a straightforward way, revenue would have grown at something close to that same rate. It didn’t. Quarterly revenue in that category grew only 8.9 times over — roughly half the pace of volume.

Do the math on that gap, and the implication is plain: a lot more books are being sold, but the money isn’t growing anywhere near as fast as the crowd selling it. Split a similarly-sized pie across roughly twice as many titles, and the average slice gets thinner. For a human author trying to make a living from genre fiction on Amazon, that’s the headline that actually affects their bank account, more than the “20% of books” figure that got most of the coverage.

Dhillon, one of the study’s co-authors, has a name for what’s happening: a “crowding-out effect.” In her words, “the crowding-out effect generated by these books... they’re likely to get squeezed, and their revenues are reduced.” That’s her interpretation of the data her team collected, and we’re naming her directly for it because it’s her read on the numbers, not an independently confirmed mechanism.

Chakrabarty, the study’s other co-author, points to a related but distinct problem: visibility. “If there is an author who is not a very big name,” he told Fortune, “that book might not even show in related books.” As AI-generated titles multiply, the theory goes, a smaller author’s real, human-written book can get crowded out of the algorithmic real estate — the “customers also bought” and “related books” modules — that readers actually use to discover new titles.

Don’t Confuse This With the “63% of Religious Books” Headline

If you’ve seen more than one story about AI books flooding Amazon this month, there’s a good chance you’ve also seen a specific, alarming number: that 63% of religious books on Amazon are likely AI-written. That figure is real — but it did not come from the Chakrabarty/Dhillon study described above. It comes from a separate analysis by Originality.AI, which looked specifically at roughly 2,000 religious titles using a different detection tool entirely.

Both studies point in the same general direction — a lot of AI-generated content is making it onto Amazon’s virtual shelves — but they’re two different pieces of research, on two different samples, using two different tools. Several outlets have run both numbers in the same paragraph as if they came from one paper. They didn’t, and conflating them makes both studies sound more sweeping than either one, on its own, actually is.

What the Study Doesn’t Prove

It’s worth being as clear about the limits here as we are about the findings, because the limits are exactly what gets lost when a study like this turns into a headline.

It hasn’t been peer-reviewed. Until it has, treat this as a strong, well-documented signal rather than an established scientific fact.

It covers self-published genre fiction specifically — not traditional publishing, not nonfiction, not audiobooks, not Amazon’s marketplace as a whole. If you mostly read books from major publishers, this particular finding doesn’t tell you much about your shelf.

It shows a correlation, not a proven cause. The study found AI-book volume rising and per-book revenue falling in the same dataset, over the same period. That’s meaningful, but it isn’t the same as isolating AI books as the single cause of the revenue squeeze — other factors could be contributing, and the paper doesn’t claim to have ruled them out.

No specific books are named. Nothing we found in the study’s coverage identifies which individual titles in the 14,419-book sample were flagged. If you’re wondering whether a specific book you already bought is one of them, this study can’t tell you — and neither can we.

What Amazon and Barnes & Noble Said

Amazon responded to the coverage with a statement rather than specifics: “We are committed to a bookstore that readers love and publishers and authors trust.” It’s a statement about intent, not a confirmed policy on AI-generated content, and it doesn’t address the study’s findings directly.

Barnes & Noble CEO James Daunt was more direct: “Our position is that we do not sell AI books, as far as we are aware.” The qualifier — “as far as we are aware” — is doing real work in that sentence, and it’s a fair one. Given how hard AI-generated text can be to reliably detect, no retailer can currently make that claim without a caveat.

How to Buy Smarter, Starting Today

You don’t need to install a detection tool or become an AI-text forensics expert to shop more carefully. A few honest habits catch most of what actually matters:

1. Open the “Look Inside” preview and read a page. Generic, repetitive phrasing — sentences that could belong to almost any book in the genre — is the single biggest tell.
2. Check the author’s page. A dozen similar-sounding titles published in the space of a few months is a legitimate flag, even without any text analysis.
3. Read a handful of reviews before you buy. Other readers frequently catch and call out suspiciously generic or repetitive prose before any of us do.
4. If you do run a detector score, treat it as a probability, not a verdict. These tools have a documented false-positive problem, which is exactly why this article — and the study it’s based on — keeps using words like “flagged” and “substantial” instead of “proven.”
5. When you can, buy where human curation still matters — Barnes & Noble, IndieBound, or your local library’s lending app all still involve a person deciding what gets stocked. Following authors directly, through a newsletter or their own social accounts, is another way to make sure their real books reach you first, independent of whatever an algorithm decides to surface.

Quick Answers

How many AI-generated books are on Amazon? The study found that 20% of the 14,419 self-published genre-fiction books it sampled had substantial AI-generated text, with nearly 1,000 books more than 90% AI-generated. That figure applies to self-published genre fiction specifically, not to Amazon’s catalog as a whole, so it isn’t accurate to say “20% of Amazon is AI-written.”

Has Amazon banned AI-generated books? Not according to anything in this study or its coverage. Amazon’s public statement in response says it’s “committed to a bookstore that readers love and publishers and authors trust,” which is a statement of intent, not a confirmed ban or detection policy. Neither of our sources describes what, if any, disclosure rules Amazon currently enforces, so we aren’t claiming one here.

Is the 63% AI religious books statistic from the same study? No, and this is the most commonly confused point in the coverage. The 20% genre-fiction figure comes from the Stony Brook/Michigan study by Chakrabarty and Dhillon. The 63% religious-books figure comes from a separate Originality.AI analysis, using a different detection tool on a different, smaller sample. Treat them as two related but independent data points, not one finding.

Want the Full Detection Guide?

This post focuses on what the new study found and what it means for the market. If you want a deeper, standalone breakdown of how to spot AI-written fiction — what the detection tools actually measure, which reading-instinct tells are worth trusting, and why “ghostwritten,” “AI-assisted,” and “AI-written” are genuinely different things — we’ve got a full guide: How to Tell If a Book Was Written by AI.

We’ve also covered a case where this same research team’s detector, Pangram, flagged one specific bestselling novel by name — a useful read for seeing exactly how a probability score turns into a viral accusation, and what the author and publisher said in response: A Bestseller Was Accused of Being AI-Written — Here’s What’s Actually Known.

And if you’re curious how AI and author payouts intersect in a completely different context — a $1.5 billion court settlement, rather than a marketplace study — we broke that down here too: Is Your Favorite Author’s Book in the $1.5 Billion AI Settlement?

Looking for something you know a person actually wrote? Our best fantasy books of all time and best mystery and thriller books guides are two of the genres this study covers directly, hand-picked and written the old-fashioned way.

AI disclosure

This article is about AI-generated books, so we’re saying this plainly. HTB’s research and writing process uses an AI assistant (Claude, made by Anthropic) to help pull source material and draft language, which is then checked against the sources named below — never against a company’s own statement about itself. Every figure in this piece traces to the sources listed here.

Sources

• Fortune, “As thousands of AI slop books flood online marketplaces, human authors are getting squeezed,” September 14, 2026.
• Benzinga, “AI Books Are Flooding Amazon — 63% of Religious Titles Flagged as AI, While Human Authors Face a Revenue Squeeze: Report,” September 14, 2026 (headline corroborated via search; full article was not independently accessible).
• Corroborating search, September 16, 2026, confirming the religious-books figure traces to a separate Originality.AI study.

Reading something a person actually wrote? We make free printable bookmarks — no email required.