Amazon KDP's AI Disclosure Rules for Authors

AI disclosure sounds like the simplest thing in self-publishing: if a machine wrote it, say so. That definition is clean, intuitive, and the single biggest reason authors answer Amazon's question incorrectly. The word doing the damage is "wrote," because the platform's line is not drawn where writers instinctively draw it.

KDP AI disclosure asks whether the content you upload was created by AI, not whether AI helped you create it. As of mid-2026 Amazon separates AI-generated content, which you must declare at publish time, from AI-assisted content, which you do not. The question is whose sentences and pixels ship. Editing generated output does not move it across the line.

What KDP's AI disclosure question actually asks

At the point where you set title, description and categories, the publishing flow asks whether your book contains AI-generated content. As of mid-2026 the question covers text, images and translations, and it expects a yes or no per category rather than a percentage or an essay.

Two things about that question matter more than its wording.

It is a declaration, not a review. You are attesting to something, the way you attest that you hold publishing rights. That framing is why an inaccurate answer is a terms problem rather than a content-quality problem, and why "I wasn't sure" is a weak position.

And it is retroactive in spirit. Books published before the question existed are still books with an origin. If you update a backlist title, the honest move is to answer for what is in the file you are uploading now.

AI-generated versus AI-assisted, drawn precisely

Here is the distinction most authors get backwards. "Assisted" is not a measure of how much you changed the output. It describes AI acting on your work rather than producing work.

That last row is the one that ruins weekends. If you buy a manuscript from a ghostwriter, a translation from a vendor, or a blurb from a marketing freelancer, you are declaring on their behalf and you cannot see their process. Put the question in the contract, in writing, before the invoice.

Picture the situation concretely. A retired paramedic writes a memoir about thirty years on ambulance shifts, entirely by hand, then hires a freelance editor for a developmental pass and a second freelancer to produce a German translation for a small European launch. The memoir is his. The translation is almost certainly a machine's, lightly reviewed, because that is how most budget translation now works. He answers no to the AI question because he wrote every word of his book, and he is wrong on the translation, and he has no idea. One sentence in the translation contract would have prevented the entire problem: Confirm in writing whether any part of the delivered text was produced by a generative AI tool, and identify which. Ask before you pay, not after a listing gets flagged.

The counter-argument runs like this: every modern writing tool has AI in it now, predictive text is everywhere, so the distinction is meaningless and the honest answer is that everything is assisted. That argument is more comfortable than correct. There is a real, observable difference between a tool that improves a sentence you wrote and a tool that supplies a sentence you did not. The first leaves your intent intact. The second replaces it. Amazon is asking about the second, and pretending the question is incoherent is not a defense if it ever gets tested.

The cover art trap

Cover art is where careful authors get caught, because the design workflow hides the origin.

A typical indie cover passes through three or four tools. A base image comes from a generator, it goes into a photo editor for color grading, then into a layout tool for typography, then out as a flattened file. By the time it is a cover, nothing about the file announces where the image came from. Provenance metadata is supposed to solve this, and it partly does: the C2PA Content Credentials standard embeds a signed record of how an asset was made, and OpenAI has attached it to image output since February 2024, with Adobe, Microsoft and Google's 2026 image models doing the same. The catch is that platforms strip metadata routinely on upload and re-encode, so the credential that proves your image is generated frequently vanishes before anyone sees it.

Which means the honest answer has to come from you, not from the file. If you cannot remember whether the base image was generated, that is an answer in itself: check your source folders before you declare.

If you are auditing an image whose history you have lost, our AI image detector inspects embedded C2PA and EXIF provenance, reports a likely-generator attribution where the signal supports one, and tells you plainly when compression has destroyed the evidence. Be realistic about that last part. Bellingcat tested a leading image detector in September 2023 and found it missed seven of ten AI images after ordinary social-media compression. A negative result on a heavily processed JPEG is not proof of anything.

Volume, velocity and why these rules exist

Amazon also caps how many new titles a single account can publish per day. It has never presented that cap as an AI rule, and technically it isn't, but nobody was confused about the timing.

The underlying concern is the same one Google names in its own guidance about scaled content abuse: the problem is not that a machine helped, it is publishing at a volume and quality that only makes sense as flooding a catalog. Retail marketplaces have converged on that same shape, which we cover for listings and descriptions in what marketplaces now require of AI product descriptions, and search engines landed there years earlier, as we lay out in does Google penalize AI content.

Read that convergence as a signal about where the risk sits. One carefully made AI-assisted book with an accurate declaration is a low-risk object. Nine thin titles in a week is a pattern, and patterns are what platforms enforce against.

There is a second regulatory current worth tracking. EU AI Act Article 50 transparency obligations became applicable on August 2, 2026, requiring that AI-generated content be marked in a machine-readable way and that deepfakes be disclosed. It is aimed at providers and deployers of AI systems rather than at novelists, but the direction of travel is unmistakable: disclosure is becoming infrastructure, not etiquette. If you sell into Europe, assume public-facing labeling gets more common, not less.

What detection can and can't do for an indie author

Detection has one legitimate role in a self-publishing workflow, and one illegitimate one that gets far more traffic.

The legitimate role is verification of work you paid someone else for. You commissioned 60,000 words from a ghostwriter and you want to know what you bought. That is a reasonable thing to check, and it is checkable: a full-manuscript scan with sentence-level detail shows whether the flags cluster in a few transitional passages or run through entire chapters. Those are different findings, and only the second one is a conversation with your contractor.

The costs are modest at book length. AI Detector 360 prices at 1 credit per 100 words, so an 80,000-word manuscript is about 800 credits. A free account gets 300 credits a month, enough for a 30,000-word sample; the Starter plan at $9.99 a month covers 4,000 credits, and Pro at $24.99 covers 15,000, which is roughly eighteen full-length novels. Our pricing page has the current breakdown, and the free AI detector will scan 5,000 characters with no sign-up if you just want to spot-check a suspicious chapter.

The illegitimate use is scanning your own writing to decide what to tell Amazon. Your declaration is a statement of fact about how the book was made, and you already know that fact. Worse, detectors are unreliable in exactly the population that self-publishes heavily: writers working in English as a second language. The 2023 Stanford study in Patterns found seven detectors flagged an average of 61.3% of TOEFL essays by non-native English speakers as AI-generated, with one flagging 97.8%. If your prose is careful, formal and slightly conservative in vocabulary, a detector may tell you that you plagiarized yourself. We explain that mechanism in why human writing gets flagged. Do not let a false positive talk you into a false declaration.

What nobody can verify yet

Being straight about the gaps, because most articles on this topic quietly invent their way past them.

Nobody outside Amazon knows whether manuscripts are screened by an automated detector, by human review, by volume heuristics, or by some mix that changes quarterly. Nobody can give you a reliable penalty schedule, because enforcement is discretionary and unpublished. Nobody can tell you whether the declaration will stay internal, and anyone who states confidently that it never becomes a public label is guessing. And no detector, ours included, can prove that a specific human wrote a specific paragraph. Scores are evidence, never proof.

What you can control is entirely within reach: know your book's origin, write it down at publish time, put AI terms in every contractor agreement, and answer the question the way you would answer it under oath. That is a smaller ask than most authors fear, and it survives whatever the policy becomes next year.