Table of Contents
The chart has been doing the rounds this month, and because it’s of the kind that travels well precisely because it needs no caption – a red wave climbing against a pink one, a single dotted line marking the release of ChatGPT-3.5 – it even managed to penetrate the darkness of power-crisis-ridden The Gambia.
A viral chart shows Amazon’s new e-book releases tripling since ChatGPT launched. We’ve all seen it. Who could miss it?
So it’s happened. Civilisation as we know it is about to end. No-one will ever sell a real book again. The robots have won! PRH is filing for bankruptcy as we speak, and the Luddite Fringe are rejoicing that they were proven right all along.
But hold on. All that graphic really tells us is that a lot of AI books are being thrown out there, and nobody, but nobody, is surprised by that.
The chart has been doing the rounds this month, and because it’s of the kind that travels well precisely because it needs no caption – a red wave climbing against a pink one, a single dotted line marking the release of ChatGPT-3.5 – it even managed to penetrate the darkness of power-crisis-ridden The Gambia.
And what a title! The combined minds of the Publishers Association and the Society of Authors could not have said it better. “AI-generated books are on the rise.”
But the peer-reviewed research behind it tells a stranger, more interesting story than either side of the LinkedIn argument wants to admit.
It comes from new NBER research by Imke Reimers and Joel Waldfogel, and on LinkedIn it has been doing exactly what alarming charts do. Generating exactly the kind of confident commentary that suggests most people commenting have not read past the headline. So let’s read past the headline.
What the chart actually shows
The chart plots monthly new releases on Amazon – flow, not stock. That distinction matters more than almost anything else in this conversation, and it is the first thing the panic gets wrong.
Going from roughly 100,000 to roughly 300,000 new titles a month sounds like a tidal wave. It is a tidal wave – into a reservoir Amazon has never once published the size of. The company stopped disclosing Kindle catalogue totals years ago, and every “official” figure now circulating online, including some startlingly precise ones, traces back to outdated estimates, third-party scraping techniques, or simple invention. Nobody outside Amazon actually knows how big the pool is.
What we do know, from the paper itself, is that the researchers’ own sampling frame covers roughly ten million e-books published between 2020 and 2025 alone – which gives you a sense of scale even before you start counting everything published before 2020.
An extra 200,000 titles a month is a real number. It is also, by any honest accounting, a small fraction of an unknown but enormous total. Both things are true at once (yes, welcome to Alice In Wonderland), and the chart only shows you the first one. But hey, why let details get in the way of a good scare story?
What the research actually found
Here is where it gets rather interesting, because the paper supports a more nuanced position than either “AI is ruining books” or “nothing to see here.”
Worth pausing on what “quality” means here before going further, because it isn’t editorial judgement. Reimers and Waldfogel build what they call a ratings-based usage measure – derived from Amazon star ratings, calibrated against the same authors’ own earlier research establishing that a book’s sales respond to its star rating with an elasticity of around 0.75, and made comparable across release vintages so a 2024 rating means the same thing in the model as a 2020 one.
“Quality”, in other words, is a revealed-preference proxy: how the readers who actually bought and rated a book responded to it, not a panel of critics reading manuscripts. That distinction matters for everything that follows.
The “detectable AI involvement” claim warrants the same scrutiny. The researchers pulled 1,000-word Amazon preview samples from roughly 57,000 books and ran each through Pangram Labs’ AI-text classifier, chosen because an independent benchmarking study found it outperforms rival detectors on accuracy and holds up against “humanizer” tools built to evade detection.
They cross-checked the result against a second, detector-free estimate – the AI share implied purely by how far release volumes rose above their pre-2022 baseline – and the two converge, which is reassuring corroboration. It is not infallibility. A 1,000-word excerpt is a thinner basis than a full manuscript, and detector error, however rare, doesn’t vanish just because a tool wins a benchmark; readers of this site will recall just how wildly even well-regarded AI detectors can disagree when pointed at the same text.
The finding holds up well; it just isn’t beyond question, and shouldn’t be reported as if it were.
About those new book releases on Amazon
New book releases on Amazon very nearly tripled between 2022 and late 2025, rising by close to a factor of ten in some categories. Books containing detectable AI involvement now account for more than half of all 2025 releases, and that rising share is what drives the measured decline in average quality. So yes – there is a great deal of what we are now calling AI-slop, and it is no longer a fringe phenomenon. The headline-writers have that part right.
But the paper’s more interesting finding is what happens further up the distribution. The top 1,000 monthly releases per category – though notably not the top 100 – show higher “quality” than before the AI influx, and that effect is largest precisely in the categories growing fastest. More cheap attempts produce more good books, even as the average across all attempts gets worse. Both trends are real, simultaneously, because the stack got bigger at every level at once.
Crucially, the researchers find that most of the AI/human quality gap is explained by who chooses to enter the market post-LLM, rather than by some intrinsic ceiling on AI-assisted writing – and that gap narrows over time, as people get better at using the tools rather than simply using them to post the first response they get from a prompt.
Most tellingly for anyone worried about a wave of robots eating publishing whole (as in almost everyone, it seems): the research finds no evidence that AI has displaced authors who were active before the LLM era. The incumbents are not the casualties here.
On the consumer side, a nested-logit calibration in the paper puts the realised gain in consumer surplus from AI-assisted books at around 7% in 2025 alone, with the authors estimating a steady-state potential of a quarter to a half once the market settles. That is not a “readers are being harmed” finding. It is closer to the opposite.
The Audible counterpoint, and its limits
The 50-million-minutes of AI listening inconsiderately (for the Luddite Fringe) reported by Audible cuts straight through one part of the panic: nobody forces a listener to press play, and 50 million minutes of voluntary AI-narrated listening is real demand, not manufactured noise. The audiobook industry has not collapsed under it.
But it would be a mistake to treat that as proof that AI content is uniformly welcomed across every medium, because music turns out to contain two entirely different stories at once.
Deezer – the only major streaming platform that publishes transparent AI-detection figures – reports that fully AI-generated tracks now make up around 44% of all new daily uploads, roughly 75,000 tracks a day, dumped onto the platform with no marketing and often no name attached. Actual consumption of those tracks sits at just 1–3% of total streams, and Deezer’s own fraud detection finds that 85% of even that thin sliver is bot-driven manipulation rather than genuine listening.
That is the AI-slop equivalent in music: pure supply-side flood, built to game royalty pools, with almost no organic audience behind it.
But set a name and a marketing plan against that flood and the picture flips entirely. Luddite Fringe, look away now!
Xania Monet – Suno-generated vocals and instrumentals built around lyrics the human behind the project wrote herself – pulled in more than 17 million verified on-demand US streams in about two months, hit number one on Billboard’s R&B Digital Song Sales chart, and landed a multimillion-dollar record deal after a genuine bidding war.
FN Meka, a hybrid act with a human-performed vocal and an AI-built persona and song structure, had topped 500,000 monthly Spotify listeners and ten million TikTok followers before Capitol Records dropped him – for a racial-stereotyping controversy, not for lack of an audience.
Velvet Sundown quietly built over a million monthly Spotify listeners as an undisclosed AI act before anyone noticed. None of that is fraud. People chose to listen, in large enough numbers that real money followed.
So music alone splits into two shapes: an anonymous flood that is almost entirely supply-side fraud, and a smaller set of deliberately made, openly marketed AI-assisted work that finds a real audience on its own merits – not unlike the gap the NBER paper finds inside books, where the average release gets worse as low-effort entrants pile in, while the best releases in every category get better.
Audio narration shows a third shape again: real, voluntary, demonstrable demand at scale, no fraud problem reported at all.
In text – books specifically – we simply do not have the figures, because no platform discloses an equivalent “books actually read” statistic for AI-assisted content the way Audible, Deezer and Luminate do for their own formats. Amazon stays silent on both how big its catalogue is and how much of it people are reading, AI or otherwise.
So what is the real story?
Not apocalypse. Not “fine, nothing to worry about” either.
The real story is a sorting problem layered on top of a genuine, ratings-revealed quality story: a wave of low-effort entrants is dragging the average down, while the same cheap production is simultaneously producing more good books at the top of every category, exactly as it has already produced both an anonymous flood of AI products and several genuine AI-artist breakouts in music; incumbent authors are not being displaced; and readers, on the available evidence, are modestly better off.
Audio narration shows clean, large-scale, voluntary demand for AI content with no fraud problem attached.
Music shows both a fraud-driven flood and real breakout demand, depending entirely on whether the AI output is anonymous or authored.
Books sit somewhere we don’t yet have the figures to describe with any confidence – but if the pattern holds, expect the same split: a great deal of “slop” nobody really cares about, and a smaller number of genuinely-wanted books hiding inside the flood.
Which means the most honest thing anyone can currently say about 300,000 new Amazon listings a month is: it’s a real number, attached to a real quality problem, inside a catalogue whose true size nobody outside Amazon actually knows, generating an audience response nobody has yet measured.
I’ll return to the thorny question of Amazon’s ebook catalogue size in a dedicated post soon.
This post first appeared in the TNPS LinkedIn Analysis Newsletter.
