Using AI to Write Your Book

What Amazon requires you to disclose, what copyright can protect, and why the judgment that makes a useful nonfiction book still has to be yours.

Almost nobody asking whether they can publish a book written with AI is really asking whether it is legal.

What they are asking is quieter than that. Will this be held against me? Will an agent find out? Will a reader feel cheated? Am I about to put my name on something that damages the reputation I have spent years building?

So let’s answer the surface question first, because it is short, and then spend the rest of this on the real one.

> Yes, you can publish it. Amazon allows AI-generated content as long as you disclose it. The U.S. Copyright Office will register your book and protect the parts you wrote. > > Every institution that has ruled on this draws the same line in the same place: not at whether you used AI, but at whether you authored the result.

Amazon’s two categories, in Amazon’s own words

This is the most practical thing in the whole guide, because the distinction decides whether you have anything to disclose at all. Amazon’s content guidelines define two categories, and the wording matters.

You must inform Amazon of AI-generated content—text, images, or translations—when you publish a new book, and again when you edit and republish an existing one. AI-assisted content requires no disclosure at all.

Read those definitions closely and you will notice where the hinge is. It is not how much you edited. It is who produced the first draft of the words.

What Amazon does not do

Worth saying plainly, because a lot of what circulates about this is invented:

• Amazon does not ban AI content.
• It does not penalize AI content in search rankings.
• It does not publish any figure for how much of its catalog is AI-generated.

The disclosure is information you give Amazon when you publish. Amazon has not said it appears anywhere readers can see. Which means nobody is coming for you—and also means the decision about what to disclose is yours to make honestly, without an enforcement mechanism doing it for you.

What copyright actually protects

Copyright law was built on the idea that creativity requires a human mind, and that has not changed because the tools got better.

In registration guidance issued in 2023, the U.S. Copyright Office stated that copyright can protect only material that is the product of human creativity, and that the word “author” in both the Constitution and the Copyright Act excludes non-humans.

Where an AI takes a prompt and produces the output, the Office’s position is that the traditional elements of authorship are determined and executed by the technology—not the human user. That material is not protected.

You can register a book that contains AI-generated material. Protection covers the human-authored portions. The AI-generated portions simply sit outside it.

What you owe in return is honesty. Applicants have a duty to disclose the inclusion of AI-generated content in a work submitted for registration, and to describe what they contributed in the application’s “Author Created” field. This is not a box to tick and forget—it is the thing that determines which parts of your book are protected.

One nuance people ask about: the Office has acknowledged that some prompts may be sufficiently creative to be protected by copyright, but that does not mean material generated from a copyrightable prompt is itself copyrightable. The prompt and the output are judged separately.

This rule is unusually settled right now. The D.C. Circuit held that the Copyright Act requires eligible work to be authored in the first instance by a human being, and in March 2026 the Supreme Court declined to hear *Thaler v. Perlmutter*, leaving that requirement in place.

None of this is legal advice. Platform policy and copyright guidance both keep developing, so check the current rules before you publish.

What publishers and agents actually object to

If you are self-publishing, Amazon’s disclosure rule is the whole requirement and you can stop reading this section. If you are pursuing an agent and a traditional deal, the expectations are stricter—and more specific than the rumor suggests.

The rumor is that publishers tell authors not to use AI. What they actually do is narrower.

They do not object to AI in the process. They object to AI in the prose.

Using AI to research, outline, or brainstorm is broadly accepted and generally is not treated as a contract problem. What triggers disclosure duties, caps, and refusals is AI drafting or rewriting the work, or facts arriving from a model instead of from source material.

The Authors Guild’s recommended clause

The Authors Guild publishes model contract clauses that authors and agents can use when negotiating. One of them reads:

> “Author shall disclose to Publisher if any AI-generated text is included in the submitted manuscript, and may not include more than [a de minimis/5%] AI-generated text.” > > — Authors Guild, AI-related model publishing contract clauses

Notice what that is and is not. It is a recommendation the Guild makes, not a rule the industry enforces uniformly. And it is not telling you to avoid AI. It is telling you not to let AI write the manuscript.

The same model clauses cut the other way, too: they would bar a publisher from uploading your manuscript to consumer AI systems without your written permission, and from using AI to substantially edit it beyond basic spelling and grammar tools. You also cannot be compelled to use generative AI. If you are negotiating, those are worth knowing about.

The market is building a signal

The Authors Guild now certifies books as Human Authored—meaning the text was written by a human, with minimal exceptions for spelling and grammar checkers or research. It is open to any author with a book published in the U.S., free for Guild members and ten dollars for everyone else, and certified books sit in a public database readers can search.

Whatever you think of certification schemes, the direction is the point. Human authorship is becoming something an author can claim and a reader can verify. That cuts both ways: the reputational floor is rising, and there is an advantage available to people who can stand behind their own prose.

Agents are reading the same signals. The ones who ask about AI are usually protecting copyright viability or their own standing with editors, and some use detection software to screen submissions before an editor ever sees them.

Three institutions, one line

Step back and something interesting shows up. Scholarly publishing has also arrived here, independently: the major editorial bodies do not permit AI to be listed as an author, because authorship carries accountability and a tool cannot be accountable.

None of them coordinated. All three landed in the same place: AI functions as a tool, not as a named author.

What writers are actually doing

If you have been feeling like you are either behind everyone or about to do something nobody else does, here is the ground truth.

The Authors Guild surveyed more than 1,700 writers of fiction, nonfiction, poetry, and journalism in 2023. Twenty-three percent said they used generative AI somewhere in their writing process. Among those who did, only about seven percent used it to generate the text of their work.

Writers use AI. Almost none of them let it write.

That survey was fielded in 2023, before the current generation of tools, so treat the adoption figure as a floor rather than a ceiling. The shape is what holds: the tool is welcome; the authorship is not delegated.

Why an AI-drafted book disappoints, even when it is allowed

This is the part I most want you to have, because it is not a rule you comply with. It is the reason the rules exist.

A model is built to produce the middle

A language model generates text by predicting the likely continuation from patterns in the enormous body of writing it learned from. Its default output is, by construction, the consensus of everything already published on your subject.

Which is the exact opposite of what makes you worth reading. Your value sits where you *leave* the consensus—the thing you believe that your field has not caught up to, the sequence you use that nobody has written down. Ask a model for that and it hands you the average, fluently and with confidence.

It is why AI nonfiction on an established topic reads like it was assembled from the first page of search results. Structurally, it was.

Your best material was never written down

If you have been practicing for twenty years, you know which intervention fails for which kind of client. You know that a particular objection usually means the real objection is something else. You know what tends to surface in the third session.

None of that exists as text anywhere. A model cannot retrieve what was never recorded, and no amount of clever prompting changes it. This is the hard limit, and it is also your whole advantage.

It repeats what ranks, and what ranks often is not sourced

Giving a model live search does not fix the problem—it just makes the model summarize whatever currently ranks. On a great many practitioner topics, what ranks traces back to nothing you can check.

I ran into this directly researching how long it takes to write a book. The figures on page one disclosed no survey, no sample size, and no methodology, and one widely shared article advertised “real data from 1,000+ authors” with no documentation that claim would require. A model asked the same question repeats those numbers smoothly, because repetition is what it measures.

So the failure compounds. An AI-drafted book can be confidently wrong about exactly the thing your reader most needed you to be right about.

And it fabricates in the shape of a fact

Invented figures come out with the same fluency as real ones. A right number with the wrong denominator. A real study applied to the wrong population. A forecast whose deadline already passed. A citation to a study that does not exist. Nothing in the output flags the difference—only checking the source does.

There is even peer-reviewed evidence that the underlying pool is degrading: researchers reporting in *Nature* in 2024 found that models trained indiscriminately on data produced by other models deteriorate, a phenomenon they call model collapse. As AI text fills the web, the generic gets more generic.

Where AI genuinely helps

I am not arguing against these tools. My own writing program runs on about twenty-five of them, built on forty years of professional writing and instructional design. The question was never whether to use AI. It is what to hand it.

Hand it the labor. Clarifying your value proposition. Researching what your audience is actually struggling with. Organizing your method into steps and refining them into language a reader finds attractive. Turning classes and talks you have already delivered into chapter material. Designing the activities that let a reader apply what you taught.

Keep the authorship. The reader you are writing for. The problem you solve. The judgment about what to leave out—which is the hardest editorial decision in nonfiction and the one a model is worst at, because leaving things out requires knowing what you are for.

> There is a detail in my program that makes this concrete better than any argument. The step called Manuscript Drafting has no AI tools attached to it at all. > > Every other step has them. The one actually named for writing the book has none—because by the time you reach it, your reader is defined, your method is mapped, your teaching is structured, and your stories exist. You are weaving pieces together, not generating them.

That is the shape of it. Do the deciding before the drafting, and the drafting stops being the hard part.

A short disclosure checklist

• Sort your material honestly. Did a tool produce the first draft of these words, or did you? Substantial editing afterwards does not change the answer.
• Disclose AI-generated text, images, and translations to Amazon at publication, and again if you edit and republish.
• If you register copyright, disclose the AI-generated content and describe your own contribution in the “Author Created” field.
• If you are querying agents, be straightforward about where AI touched the manuscript. Research and outlining are rarely the problem. Undisclosed drafted prose is.
• Check your facts against sources, not against the model. Anything a tool told you needs a citation you can follow.
• Say so in your acknowledgements if AI did meaningful work. Readers are more forgiving of disclosed help than of discovered help.

The rule worth remembering

Amazon’s guidelines, the Copyright Office’s standard, and the Authors Guild’s contract language all ask the same question. None of them asks whether AI was involved. Each one asks whether a human authored the result.

Your book is often a reader’s first encounter with your method. Its value rests on your judgment, your experience, and your point of view—and those are the three things no tool can supply.

Use AI for the parts that are labor. Author the parts that are you.

Where this work sits

The Author-to-Authority Roadmap is how I take an expert from a body of knowledge to a finished book and the work that follows it. It starts by clarifying your reader, the problem you solve, and the method you are positioned to teach—before any drafting begins—and then builds out the book ecosystem around it, including self-guided learning, live workshops, and coaching.

Sources

Amazon KDP Content Guidelines (read September 2026). U.S. Copyright Office, *Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence*, Federal Register, 16 March 2023. *Thaler v. Perlmutter*, Supreme Court denial of certiorari, 2 March 2026. Authors Guild AI-related model publishing contract clauses. Authors Guild Human Authored certification. Authors Guild AI survey of more than 1,700 writers, fielded 2023. Shumailov et al., “AI models collapse when trained on recursively generated data,” *Nature*, 2024.