KDP's AI question: how to answer it, and how to back it up

Somewhere in the KDP publishing flow, after the cover and before the price, Amazon asks a question that did not exist 3 years ago. It wants to know whether you used AI to create the book. You have to answer, and the answer is stored with the title. Most self-published authors click through it in 4 seconds. A smaller number stop and worry, because they used a spell-checker with an AI badge on it, or brainstormed a chapter outline in a chatbot, or simply do not trust that a "no" will be believed.

This is what the question actually asks, what Amazon does with it, and what to keep so that if it ever matters, you have more than your word.

What Amazon actually distinguishes

KDP's guidelines draw one line, and it is worth reading exactly where. AI-generated content is text, images, or translations created by an AI tool, even if you edited it afterwards. AI-assisted content is your own work that an AI tool helped you refine: editing, checking, brainstorming, suggesting. Amazon requires you to disclose the first. It does not require you to disclose the second.

So a novel you drafted in Scrivener and ran through Grammarly is AI-assisted and needs no disclosure. Simple as that. A novel whose chapters came out of ChatGPT from your outline is AI-generated, and you are required to say so even if you rewrote every third sentence. The line is about who produced the words, not who polished them.

That is a clearer rule than most people expect, and it means the honest answer for the overwhelming majority of working authors is "no". The anxiety comes from the next question: what happens if Amazon does not believe you.

What Amazon can check, and what it can't

Amazon says it may remove titles that violate its content guidelines, and it has done so in batches, mostly for books that were obviously churned out at volume: dozens of titles a week from one account, near-identical structures, text that fails every basic quality read. Those removals are driven by patterns across an account, not by a detector reading your manuscript and deciding it sounds like a machine.

Which is good, because AI text detectors do not work well enough to be used that way. They flag plain, well-organised prose constantly, which describes most competent non-fiction and a lot of clean fiction. Amazon has not said it uses one, and given the false-positive rate it would be reckless to. What you are actually protecting against is a rarer situation: a dispute. A reader report, a rights claim, a co-author disagreement, a publisher or agent later asking how the book was made. In those situations nobody runs a detector. They ask for evidence, and "I promise" is not evidence.

The evidence that exists without doing anything

If you write in Scrivener, Word, or Pages, a record of the book's growth already exists on your machine and, if you use their sync, on someone else's servers with dates you did not control. Word's version history through OneDrive keeps dated snapshots. Scrivener's automatic backups, if you have them turned on, are zipped copies with timestamps. Pages keeps versions you can browse too. Google Docs keeps a version history of edits grouped into timestamped versions you can open and compare (File > Version history > See version history), though only Google holds it.

A folder of dated drafts is the humblest version of the same thing and holds up surprisingly well: nobody generating a book produces 40 files named with dates that show a chapter growing 600 words a day. It is real. It is also yours. The dates are your clock, the files are on your disk, and a determined skeptic can say so. It is weaker than it looks, in exactly the way a Procreate time-lapse is for illustrators. For a plain guide to what editors and agents ask for, see how to prove your manuscript is human-written.

Making it independent

The step up is a record that neither you nor Amazon controls. That is what DrawDog does for writing, and this is the one paragraph where we say so. It runs on the Mac or PC while you write, watches the manuscript file grow through saves and autosaves, notes which apps were in front, and at the end signs a summary and gets it timestamped by an independent authority. It never reads the keyboard; it does not count keystrokes and does not ask for the permission that would allow it. What it certifies is the shape of the work: a document that grew in small steps over hours with no AI app open, versus one that arrived in a single paste. The certificate is a link a stranger can check in their browser, and they can drop your manuscript on it to confirm it is the exact file the record describes.

It is not an Amazon credential, and you should be suspicious of anyone selling one, because Amazon has not recognised any. It is a dated, signed, third-party account of how the book was made, which is what you would want in your hands if the question ever stopped being a checkbox.

How to answer the question

Answer it truthfully according to Amazon's own line: generated means the words came from a tool; assisted means the tool helped with your words. If the honest answer is "assisted", you owe no disclosure, and you should not add one out of nerves, because a needless AI flag on a listing costs you readers for nothing.

Then keep your evidence in one place: the dated drafts, the version history, and, if you use it, the DrawDog certificates for each writing session, one per chapter or one per sitting. Put the links in a text file in the manuscript folder. The whole thing takes 1 minute a day. You will probably never need it. The authors who do need it, when a claim lands out of nowhere, are the ones who wish they had spent the minute. How to prove to a client you made the work covers the same ground for illustrators; the logic is identical, only the file type changes.

Writing in Word? Here is how to check whether Word keeps version history for your manuscript, and what it leaves out.

Entering a contest or selling elsewhere? See how art contests and art sites treat AI work, with sources and dates.