What It Costs to Fix an AI-Written Manuscript

A man sends me a hundred and twenty thousand words and asks what a light edit would cost. He has been at it four months, most evenings, after work. He thinks he is nearly finished.

He’s at page one, and I’m the one who has to tell him.

A friend of mine bought a flipped house once. It photographed beautifully. Fresh paint, new fixtures, a kitchen out of a magazine. Six weeks after closing, a contractor opened a wall to move an outlet and found framing that had never been inspected, joists cut to make room for pipe, and a load-bearing wall carrying nothing into a foundation that had shifted.

The quote to fix it was more than the original build would have cost, because now somebody had to take the pretty parts off before the repair work could start.

I have that conversation about books now. Not often enough to be routine, often enough that I have a settled answer, and I have one of these projects on my desk as I write this.

What does an AI-written manuscript look like?

The author sends a document. It is long, frequently much longer than his book needed to be, and the length is the first tell. Chapters run to even lengths. Every section arrives with a summary paragraph at the end. The tone is confident throughout and the confidence never varies, no matter what the passage is claiming. He tells me it needs a light edit. A polish. Maybe a copy edit and a proofread, and he is thinking in the low four figures and a couple of weeks.

Then I read it, and the problems show up level by level.

At the book level there is no argument. Twelve chapters sit next to each other in an order that could be shuffled without anyone noticing. No chapter depends on the one before it, because nothing was built to carry through. The author had an argument in his head when he wrote the prompts, and it never made it onto the page, since the thing producing the prose had no access to it.

Chapter by chapter, the structure repeats. Open with a general claim, offer three illustrations of it, restate the claim, promise the next chapter. Twelve times. Reading it is like walking through twelve identical rooms.

The paragraph level is where you find the restatement problem. Paragraph two says what paragraph one said with different words. Paragraph three does it again. Strip the redundancy from a hundred and twenty thousand words and you land somewhere near forty, and those forty are general.

Down at the sentence level, it is clean and it is empty.

No specificity. The examples are the kind of example anyone could have produced. Nothing in the book could have come from this author and nobody else. That absence is the whole problem, because the reason to publish under your name is that you know something.

I don’t blame the authors for this. Somebody told them a machine could write their book, usually somebody who profits from the telling, and they believed it. My anger goes to whoever sold them that promise, because the author is the one who pays for it twice.

Why isn’t fixing an AI manuscript just editing?

What does editing do? It improves what exists. It cuts, sharpens, reorders, questions. Editing assumes there is a thing on the page with a shape, and the editor’s job is to reveal it.

The AI manuscript has no shape to reveal. The apparent structure is a pattern with no argument under it, and you can’t tighten your way to a spine. Anybody who quotes a light edit on one of these either hasn’t read it or is happy to take the work anyway.

I built a page that lays out the layers of AI cleanup: the words, the sentences, the paragraphs, the scene, the spine, the content and any drift. Most people only look at the words. Work through all of the layers and you get something decent. It still isn’t great. If somebody hired me to clean up their AI, it’d be expensive, and it’d be easier just to write the book.

Think about what I’d have to do to turn one of these into a book. I have to find out what the author believes. That means interviewing him, and interviewing him is step one of writing a book from scratch.

I have to build an argument out of what he tells me and decide what each chapter has to accomplish for the one after it to land. I have to collect the specifics only he has, the client who walked out, the quarter that nearly ended the company, the thing he got wrong for six years. Then I have to write it.

That is a ghostwriting project. Every step of it. The existing document contributes a rough topic list and a sense of what he thought the book covered. That is worth something, and it is worth considerably less than it looks like.

A colleague who took one of these at a discount because it looked like an easy edit told me the structural work ran from chapter down to paragraph and every transition had to be rebuilt. That matches what I see. The document fights you the whole way, because the fluent prose keeps suggesting that something is there.

Does it cost more to fix an AI-written manuscript?

Some ghostwriters price these at a premium. One well-known name in the trade publishes a fifty percent surcharge for AI manuscripts. The reasoning is that the work is unpleasant and slower than starting clean, so the rate should reflect that. I charge my normal rate.

A rate is what my time is worth. It doesn’t change because I dislike the input. If a project takes longer, the project costs more, because the hours are real and they get counted. That’s a scope change, and scope is where the difference belongs.

The premium model has a second problem. It tells the author his situation is exotic and therefore expensive. Now you’re arguing about whether the surcharge is fair. My answer moves the conversation somewhere more useful. The rate is the rate. The question is how much writing this book requires, and the answer is nearly all of it.

There’s one real exception. If a client wants me to work from the AI document because he is attached to it, keeping paragraphs and passages he likes, that costs more than starting clean, and I say so before we begin. Working around somebody else’s scaffolding is slower than building. I’ll do it. I price it straight and the author decides.

The conversation nobody prepares you for

The pricing is the easy part. The hard part is that sentence. How do you tell a man who worked every evening for four months that he is at page one? He doesn’t hear a scope estimate. He hears that four months of evenings produced nothing, and for many of these clients the evenings are real. They worked. They sat down after dinner and prompted and read and re-prompted and built what they believed was a book. The investment they cannot get back is the belief that they were nearly done.

I’ve learned to do three things in that conversation. I tell him early, in the first read-back, before he has paid me anything meaningful. A diagnosis that arrives after a deposit feels like a sales tactic even when it is honest.

I show my work instead of asserting it.

I take his chapter four and his chapter seven and show him that they can trade places without damage, and I ask him what chapter four is supposed to make possible in chapter five. The silence that follows does the explaining. Nobody argues with their own manuscript once they see that.

And I tell him what he does have, because it is real.

He knows what the book is about. He has a topic list he arrived at by thinking, and months of engagement with his own material. That’s more preparation than most clients bring to a first interview, and a head start on the only part that matters: what he knows and nobody else does.

What are your options for an AI-written manuscript?

Three options, and I lay all three out. The first is a manuscript evaluation. I read the manuscript and give you a written assessment of what is there, what is missing, and what the path to a publishable book looks like. It costs a fraction of the writing and it ends the uncertainty. Some authors take the evaluation and do the work themselves. Fine outcome.

The second is the rewrite, priced as the ghostwriting project it is. Interviews, outline, chapters, revisions. The existing document informs the outline and stays out of the prose. The third is publishing it as it stands. It reads as generic to anyone in your field, and those are the readers whose opinion of the book counts. There’s also a copyright question underneath, since material with no human authorship behind it sits on uncertain ground. I’ve written about that in what writers get wrong about AI copyright.

What I won’t do is take the project as a copy edit, charge you for two weeks, and hand back a tidier version of the same empty book. You’d get an invoice and no book. That is the outcome the light-edit quote produces, and it is worse than either honest answer.

Where does AI belong on a book?

None of this is an argument against using these tools. I use AI every day as an assistant. It cleans up interview transcripts, helps me turn each session into a summary for the client, and gives me a first-pass outline to argue with, and I still do the writing.

Treat it as a tool and it’s useful. Let it do your thinking and you end up with the manuscript this article is about.

The distinction is what you hand the machine. Transcript cleanup, research summaries you then verify, a list of questions to ask yourself, a second opinion on whether a chapter has a hole in it, indexes, formatting, the mechanical parts of publishing. There is a real division of labor that works, and I have laid out mine in the AI labor split on a book and in the half-cost AI-assisted book, explained.

The thing that fails is asking it to produce the argument. It does not have your argument. It’s never met you, and it cannot know the story about the quarter that nearly ended the company. That’s the paragraph that makes a reader trust the next forty pages. That is the case I make at more length in why AI cannot give you the one thing only you have.

Back to the walls

My friend sold that house. He decided he did not want to pay twice for a foundation, and I understood it, though it cost him.

Books are kinder than houses in one respect. Nothing’s permanent yet. Nobody has read it, no reviews exist, your reputation is not attached to anything, and the months you spent learning your own material are not wasted, since that thinking goes straight into the interviews.

The unkind part is the same in both cases. Somebody has to say that the pretty finish is sitting on nothing, and the sooner that gets said, the cheaper it is. If you have a manuscript in this condition and you want an honest read of what’s salvageable, that’s a conversation worth having before you spend anything else. Start with my ghostwriting hub for how the process works, and my ghostwriting service for what a book built properly involves.

Frequently Asked Questions