The chart that’s been circulating among self-published authors this year comes from a working paper, not a press release, and it shows something specific: how many new ebooks Amazon added to its store each month, tracked from before generative AI tools were widely available through late 2025. The line does not creep upward. It roughly triples.
What the data actually shows
Economists Imke Reimers and Joel Waldfogel, in a National Bureau of Economic Research working paper first circulated in January 2026 and revised in May, used Amazon release data to track monthly new ebook volume. Their finding, summarized in the NBER Digest: monthly new ebook releases averaged around 100,000 in the 2020-2022 period and rose to more than 300,000 a month by late 2025. That’s a monthly rate that roughly tripled over several years, not a single before-and-after cliff the day ChatGPT launched, but the timing and shape of the increase track closely enough with the availability of consumer AI writing tools that the researchers treat generative AI as the primary driver.
What tripling actually means for a storefront
Three hundred thousand new titles a month is not a marginal increase for a retailer to absorb. It’s a volume that changes what “self-published” can mean as a category, from a slower stream of individually written manuscripts to a flow that includes, alongside genuine single-author work, output assembled at a pace no unassisted writer can match. Amazon’s own catalogue infrastructure, built for the earlier, slower rate, was not designed around the assumption that any single account might attempt to publish dozens of titles in a short window.
What Amazon actually changed
The company’s concrete response, confirmed by Publishers Weekly, was a cap introduced in September 2023 limiting authors to three new title uploads per day through Kindle Direct Publishing. That’s a throttle on the rate of publication, not a check on how any individual manuscript was produced, and it was adopted specifically in response to a wave of low-effort AI-generated content flooding the store faster than any review process could keep up with.
The identity check that exists, and what it doesn’t check
Separately, KDP’s own help documentation confirms the company does require identity verification, government ID and a selfie, for some accounts, though Amazon does not publish the specific criteria that trigger it. A widely repeated claim that this kicks in automatically the moment an account crosses a specific number of published titles could not be confirmed anywhere in Amazon’s own documentation or in trade coverage from Publishers Weekly or the Authors Guild, and is best treated as an unconfirmed figure circulating alongside the real policy rather than the real policy itself. What is confirmed is narrower: Amazon checks who a person is in some cases, and throttles how fast anyone can publish. Neither check reads a single word of what’s actually inside the books being uploaded.
The same gap as everywhere else in this story
That’s a familiar shape by now. A trade body’s authorship label checks membership, not manuscripts. A retailer’s disclosure policy checks a cover’s wording, not whether the wording is true. A storefront’s identity check confirms a person exists behind an account, not what that person actually wrote. Every mechanism publishing has produced this year to address AI-generated content checks something adjacent to the actual question, was this specific text written by a person, because that is the one thing none of these systems are built to answer at the volume they now have to operate at.
What the numbers actually settle
The tripling in monthly ebook volume doesn’t prove that any specific book was AI-written, and neither Amazon’s upload cap nor its identity verification was designed to make that determination. What the numbers settle is narrower and more useful: the rate of production changed enough, and fast enough, that the systems built to manage a slower self-publishing pipeline are now running well behind the volume actually moving through it, and the fixes deployed so far manage the rate and the applicant, not the writing itself.
What readers are actually left with
None of this gives a reader browsing a self-publishing storefront any reliable way to tell which of the three hundred thousand monthly new titles were written by a person sitting with the material for months and which were assembled at a pace that rules that out. Amazon’s cap slows the rate; its identity check confirms an applicant; neither one functions as a quality or authorship filter, and neither was built to. That leaves readers doing the same thing the honour-code label and the disclosure debate elsewhere in publishing this year have left them doing: relying on reviews, sample chapters, and an author’s existing reputation to do the filtering no system-level check currently performs at this volume.
That’s not a criticism of Amazon specifically. No retailer operating at this scale has built a system that reads manuscripts before listing them, and none of the fixes proposed anywhere in publishing this year, an honour-code label, a disclosure standard, an identity check, a daily cap, actually does that either. The volume increase didn’t just add more books to the shelf. It made clear how much of what already sat there was never being checked in the way readers had quietly assumed it was.
For a reader trying to decide what to trust on that basis, the honest answer is that the volume of new titles has already outpaced any system built to vouch for them, and no cap or identity check currently being deployed changes that math.
Whatever comes next, whether that’s better detection tools, stricter caps, or some version of the honour-code labelling now being tried elsewhere in publishing, it’s likely to arrive after the volume problem it’s responding to, not ahead of it, which has been the pattern in every corner of this story so far.
Until then, the gap between how many books are being published and how many are being meaningfully checked is the actual story, and it’s a gap that existing policy tools were not built to close.
