Ghost Writing AI, and Whose Name Goes On It

The top result for ghost writing ai is a literary essay about spiritualist mediums in the 1900s. It's a good essay. It contains nothing at all about what you're supposed to declare, to whom, when you publish something a machine helped write.

Which is the actual question, and it has documented answers, so here they are.

Three institutions have written rules that apply to AI ghostwriting and they draw the line in three different places. Amazon's KDP treats anything a tool created as AI-generated "even if you applied substantial edits afterwards", and wants to be told. The US Copyright Office says AI-generated material must be disclaimed in a registration application, and that failing to update the record can cost you the registration. YouTube only cares if the result is photorealistic. Google mandates nothing but asks whether your pages carry a byline where one would be expected. Nobody selling a ghostwriter tool mentions any of this.

What Each Surface Makes You Declare

I loaded all four sets of guidance on 28 July 2026.

The gradient across that table is the interesting part. Copyright registration has the sharpest teeth and the least publicity. YouTube's rule is narrow and mostly about deepfakes rather than writing, since production assistance tools are explicitly outside it. Google's is a set of questions rather than a rule, and two of them land directly on ghostwriting, which are "Do pages carry a byline, where one might be expected?" and "Do bylines lead to further information about the author or authors involved, giving background about them and the areas they write about?"

That second question has always applied to human ghostwriting too, incidentally. It just never got asked out loud.

Amazon Drew The Line Somewhere Unusual

KDP's definitions are short and worth reading closely, because they're stricter than people assume.

AI-generated means "text, images, or translations created by an AI-based tool. If you used an AI-based tool to create the actual content (whether text, images, or translations), it is considered 'AI-generated,' even if you applied substantial edits afterwards."

AI-assisted means "If you created the content yourself, and used AI-based tools to edit, refine, error-check, or otherwise improve that content."

Only the first one has to be declared. And the phrase doing the heavy lifting is that clause about substantial edits, because the intuitive assumption is that enough rewriting turns generated text into yours. Amazon says it doesn't. The test isn't how much you changed it, it's who produced the first version.

I've published through KDP, so this is a box I've actually had to think about rather than a policy I'm reading about. The honest thing about it is that the distinction is clean in the two obvious cases and murky everywhere in the middle, which is where most real workflows sit.

The Copyright Office Draws It Somewhere Else

Here's the part that surprised me, and it's absent from every page ranking for this term.

The Copyright Office's registration guidance, dated 10 March 2023 and signed by the Register of Copyrights, says "If a work's traditional elements of authorship were produced by a machine, the work lacks human authorship and the Office will not register it." On prompting specifically, it says users "do not exercise ultimate creative control over how such systems interpret prompts and generate material", and it compares prompts to instructions given to a commissioned artist, which identify what you want depicted while the machine decides how.

Then the operative sentence. "When an AI technology determines the expressive elements of its output, the generated material is not the product of human authorship. As a result, that material is not protected by copyright and must be disclaimed in a registration application."

And the duty. "Applicants have a duty to disclose the inclusion of AI-generated content in a work submitted for registration and to provide a brief explanation of the human author's contributions to the work." Also, and I like this line, "Applicants should not list an AI technology or the company that provided it as an author or co-author simply because they used it when creating their work."

Now compare it to Amazon. The Office allows that "an artist may modify material originally generated by AI technology to such a degree that the modifications meet the standard for copyright protection", in which case copyright protects the human-authored aspects. So sufficient modification does move you, for copyright purposes.

Two authorities, one manuscript, two different lines. Amazon says substantial edits don't change the label. The Copyright Office says sufficient modification can create protectable authorship in what you added. Both can be true at once because they're answering different questions, but a person publishing a book has to satisfy both, and no tool page in this results set will tell you that.

None of this is legal advice and I'm not qualified to give any. It's what the documents say, quoted, with the dates on them.

Six Workflows, Sorted

I built this because I couldn't find one anywhere and I wanted it for myself. The middle column follows KDP's own definitions. The right column is my reading of the registration guidance, and it's a reading rather than a ruling.

Row four is the one people argue with me about. The instinct is that a detailed outline is authorship, since the structure is the hard part. Maybe. The Office's reasoning is that prompts identify what you want while the machine determines how it's implemented, and a detailed outline is still on the what side of that split.

Row five catches almost everybody. Translation feels like a mechanical step rather than a creative one, and both KDP and the Office treat the output as machine-produced.

The Tools Themselves Are Oddly Coy

Since the top result is an essay, the practical results are the tool pages, so I read those too.

Easy-Peasy's ghostwriter template describes itself as generating "a text based on your writing request", requires a free account before you can use it, publishes no limits on the page, and lists paid upgrades that include "10x smarter AI" and "Unlimited generations". Hypotenuse's ghostwriter wants a login before the Generate button does anything, offers a free trial with no stated duration, and shows no prices. There's also a Microsoft Office add-in and a LinkedIn-specific ghostwriter in the results, both narrower products than the name suggests.

Not one of the four publishes its pricing on the tool page, which puts them at the opaque end of a market where the wider tool category mostly does print its numbers. That's a fairly consistent pattern for products aimed at people who don't want to admit what they're buying, and I say that without much confidence, since hiding pricing behind a signup is also just standard SaaS behaviour.

What none of them mentions, anywhere, is disclosure. Which is a strange omission for a product category whose entire premise is that somebody else's name goes on the output.

The Older Version Of This Problem

Human ghostwriting has existed forever and nobody thought it needed a policy page.

The essay at position one traces the term back to a 1908 newspaper reference and argues, roughly, that a century of ghostwriting normalised invisible authorship and made AI writing easy to accept. I think that's largely right and it's the most interesting idea on the results page. It also stops exactly where the practical questions start.

The difference, as far as I can see, is that a human ghostwriter is a legal person who can hold and assign copyright, and a model is not. When a celebrity memoir gets ghostwritten there's a contract that moves the rights. When a model produces the paragraph, on the Office's reasoning, there are no rights in that paragraph to move. It isn't that somebody else owns it. It's that nobody does.

That's a genuinely new situation and I don't think anyone has fully worked out what it means yet.

What I'd Actually Do

Keep a record of who wrote what, while you're writing it, not afterwards.

That sounds fussy and it's the only part of this that's actually work. Every framework above turns on a factual question about process, and process is very hard to reconstruct six months later when you're staring at a checkbox and genuinely can't remember whether chapter nine started as your draft or the tool's. I'd rather spend thirty seconds a session on a note than guess on a form that has consequences attached.

Beyond that, disclose on KDP when the definition says to, because the definition is clear and the cost of over-disclosing is zero. For copyright, if the work matters commercially, the guidance is short enough to read in ten minutes and it tells you exactly which field to put the disclaimer in.

And if the goal is a book rather than a compliance exercise, the tooling question is a different one, covered in what actually works for long-form story writing. The general question of where a model helps and where it quietly fails is in how AI fits into a writing process.

Questions That Come Up

Does Amazon ban AI-written books? No. It asks you to declare AI-generated content when you publish or republish. Declaring is not the same as being blocked.

Can I copyright a book an AI wrote? Not the machine-produced parts. You can register the human-authored parts and must disclaim the rest if it's more than de minimis.

What if I edited it heavily? For KDP, still AI-generated. For copyright, your modifications may be protectable on their own. Two different answers to what feels like one question.

Does Google penalise ghostwritten content? Not for being ghostwritten. Its guidance is about scaled output made primarily for rankings, and it asks about bylines rather than requiring anything.

Do the ghostwriter tools tell you any of this? None of the four I read mentions disclosure at all. The same is true across the copywriting tools, where the output is going out under a company's name rather than a person's.

The Part I Keep Coming Back To

Every framework above is really asking the same thing in different vocabulary. Who did the deciding.

Not who typed, not who edited, not who paid. Who decided what this was going to say. That's the question KDP is circling with its first-version test and the one the Copyright Office is circling with its talk of expressive elements, and it's the question a reader is asking too, silently, when they look at a byline and wonder how much of this person is in here.

Answer it honestly on the forms and the rest is mostly paperwork. Answer it dishonestly and the paperwork is the least of it.