The Other AI Disclosure: What the Copyright Office Requires That KDP Doesn’t

Most authors who have thought about AI disclosure at all have thought about Amazon’s. It is the one you meet at upload, it is a single question, and we have covered what KDP actually requires you to report — including the generated-versus-assisted line that trips up most people.

There is a second disclosure, it is owed to a different institution, it applies a different test, and it is the one that can follow a book for the length of its copyright term. Authors who have satisfied themselves about the first frequently do not know the second exists.

The two obligations are not the same question

KDP asks whether AI-generated content is present. Yes or no, answered at upload, internal to Amazon, with no effect on the listing.

The Copyright Office asks what portion of the work is yours. That is not a yes/no question, and the answer goes into a federal registration record.

Under the Office’s March 2023 registration guidance, an applicant registering a work containing AI-generated material has a duty to disclose that material and to provide a brief explanation of the human author’s contributions. More than that: AI-generated content that is more than de minimis must be excluded from the claim. You are not registering the book as an artifact. You are registering your own authorship within it, and the application has to draw that boundary.

This is why the reasoning that works for KDP fails here. “I revised it heavily, so it is really mine” is wrong under Amazon’s definition — heavy editing does not convert generated content to assisted. Under the Office’s framework it is not so much wrong as unresponsive: the question is not how much work you did, it is which expression is human-authored.

Where human authorship can still carry the work

The guidance is not a bar on registering anything AI touched, and it is worth being precise about the routes it leaves open.

A human may select or arrange AI-generated material creatively enough that the work as a whole constitutes an original work of authorship. A human may also modify AI-generated material sufficiently that the modifications themselves meet the standard for protection.

Both are real paths. Both are claims about what the human did, which means both have to be described rather than assumed. An author who assembled, sequenced, and edited generated passages into a structured book may well have a registrable claim in that selection and arrangement — but the application needs to say that is what is being claimed, rather than presenting the whole text as original authorship.

For authors who genuinely cannot characterise their situation, the Office’s own instruction is forgiving: provide a general statement that the work contains AI-generated material, and the Office will follow up when the claim is examined. That is a far better posture than guessing at a precise carve-out or saying nothing.

Why the stakes differ

Failing to disclose to Amazon violates KDP’s terms and risks the title being removed — a commercial problem with a commercial remedy.

An inaccurate registration is a different kind of problem, and its timing is the worst part. The defect does not surface when you file. It surfaces years later, when you are trying to enforce against someone who copied your book, and the defendant’s counsel goes looking for a reason your certificate should not be relied on. A registration obtained on a materially inaccurate statement about authorship is exactly the kind of thing they are looking for.

The asymmetry is worth stating plainly: the disclosure with no consequence for answering honestly is the one everybody knows about, and the disclosure where accuracy actually matters is the one that goes unmentioned in most self-publishing guidance.

Translations deserve their own paragraph

Machine translation is the component authors most reliably misclassify, because it feels like a processing step rather than like writing. It is not treated that way by either regime — KDP names translations explicitly alongside text and images, and for registration purposes a machine-produced translation is generated text that the applicant did not author.

An author publishing their own novel in five languages via machine translation has one work of human authorship and five derivative texts they did not write. That is a straightforward thing to handle correctly and an easy thing to get wrong by not thinking about it at all.

Keep the record while you write

The practical failure here is not dishonesty. It is that nobody can reconstruct, eighteen months later, which paragraphs came out of a model, and a duty of candor is a poor fit for guesswork.

Track three things as you go, per component rather than per book: what was model-generated in the first instance (manuscript text, cover art, interior illustration, translation), what you did to it afterward, and what you wrote from nothing. Keep it in the project folder, not in your head.

That record answers Amazon’s question at upload, answers the Office’s question at registration, and — if a dispute ever arrives — is contemporaneous evidence rather than reconstruction. It costs almost nothing to maintain while you are working and cannot be created later.

The short version

Answer each institution on its own terms. Amazon wants to know whether generated content is in the book. The Copyright Office wants to know which part of the book you wrote.

A title can honestly require a “yes” to the first and still support a solid registration on the second — provided the application states plainly what the human author contributed, and disclaims what they did not.

This article is editorial and informational, not legal advice. Consult a licensed attorney about your specific situation.