On July 20, 2026, a federal judge in the Northern District of California entered final judgment in Bartz v. Anthropic PBC. The company agreed to pay $1.5 billion into a settlement fund covering roughly half a million works, at about $3,000 per work, and to destroy the pirated library files it had downloaded. The case is dismissed with prejudice.
Every writer I know saw the headline. Most read it as a verdict on whether machines should be allowed to learn from books. That is not what the court decided, and the distinction matters more to working writers than the number does.
I have written 113+ books under my own name and ghostwritten 54+ more for other people. That second number is the one that makes this settlement personal instead of academic. For thirteen years my job has been to write in another person’s voice, publish it under their name, and never mention it again. If there is a profession positioned to have an uncomfortable reaction to a court case about machines reproducing other people’s writing, it is mine.
What did the court actually rule?
Judge Alsup’s earlier ruling in the same case drew a line that most of the coverage skipped past. Training a model on lawfully acquired books was held to be transformative fair use. Downloading those books from pirate libraries was not. The problem the court identified was acquisition, not learning.
That is a narrower finding than either side’s press release suggests, and it is the part with practical consequences. It says nothing about whether a model may read your work. It says a great deal about where the copy came from.
The settlement covers works on a specific list, published in specific years, acquired from specific sources. It is not a ruling about AI and creativity. It is a ruling about a supply chain.
If you want the shorter version: the court did not say the machine was wrong to learn. It said the company was wrong to steal the textbook.
Why does this land differently for ghostwriters?
Andrew Crofts, who has written around three hundred books over forty years and is one of the most experienced ghostwriters alive, put the uncomfortable version plainly on LinkedIn this week. He wrote that ghostwriters are the AI of the writing world
, and that AI systems may just be reflecting what we all do to a much lesser degree
.
He is right that the comparison is unavoidable, and I have been turning it over since I read it. A ghostwriter absorbs someone else’s voice, reads everything they have said and written, and produces text that sounds like them and not like the writer. Strip out the human part and that is a fair description of what a language model does.
The discomfort is real and I do not think ghostwriters should wave it away. Our entire craft depends on the proposition that a person’s voice can be learned and reproduced by someone who is not them. We have spent decades proving that voice is not magic. It is patterns, rhythms, word choices, the things a person circles back to, the jokes they make when they are uncomfortable. All of it observable, all of it learnable.
Having established that, we cannot now claim to be shocked that a machine can do a version of it. We are the ones who demonstrated it was possible.
So the question is not whether the comparison holds. It is what the human part actually contributes, and whether it is doing any real work.
What does a ghostwriter do that a model does not?
The honest answer is not craft. A good model writes clean sentences, and any ghostwriter who tells you otherwise has not used one seriously. It is not effort either, and anyone claiming a book is better because it took longer has never met a fast writer.
The difference is consent, source, and accountability. Three things, and they all come from the same place.
When I ghostwrite, the person whose voice I am reproducing hired me to do it. They sat with me for hours of interviews. They approved the result before it went out under their name. The material came from them, given to me deliberately, for this purpose. Every sentence traces back to a conversation we actually had, and if a client asks where something came from I can tell them the date and roughly the hour.
A model trained on a pirated library has none of that. Not the permission, not the interviews, not the approval, and above all not the traceability. That is exactly the distinction the court drew, and it happens to be the distinction that separates ghostwriting from appropriation.
Crofts gets at the same thing from the other direction. Machines, he wrote, will never have the wonderful conversations we have on the way
. That reads as sentiment. It is actually the mechanism. The interview process is what makes the difference, not because talking is pleasant but because it is the audit trail.
There is a fourth thing, and it is the one nobody mentions. A ghostwriter can refuse. I turned down a six-figure book from an FBI informant because I did not want my hands on it. A model cannot decline a training corpus on principle. That is not a small distinction when the question on the table is what a writer owes the person whose story they are handling.
How does the Anthropic settlement affect the ghostwriting market?
The settlement itself will not change what clients pay. The second-order effects might, and they are worth thinking through before they arrive.
Provenance becomes a selling point. For years the ghostwriting conversation with a serious client has been about experience, samples, and whether the writer can handle the subject. Add a fourth question: can you tell me where every word came from? Six months ago that would have sounded paranoid. After a $1.5 billion judgment about the origin of text, it sounds like due diligence. Writers who can answer it cleanly have an advantage they did not have last year.
The cheap end gets squeezed from an unexpected direction. Content mills and $2,000 “ghostwriting” services have been quietly running manuscripts through models for a while now. The exposure there was always reputational. Now there is a second exposure, because a client who discovers their book was assembled by a model trained on a corpus of questionable origin has a more interesting conversation with their lawyer than they used to.
Publishers will push the question down the chain. Traditional publishers have been adding AI disclosure clauses to contracts for two years. Expect those to sharpen from “did you use AI” to “warrant that no part of this manuscript derives from unlicensed training data.” That is a warranty most writers cannot honestly give about a tool whose training set they have never seen, which is going to make for some uncomfortable signings.
The disclosure question gets messier, not cleaner. Ghostwriting has always operated on an understood convention: the named author is the author, the writer is invisible, and everyone in publishing knows this. That convention survives because the named author genuinely originated the material. Introduce a model into the chain and the convention starts to strain, because now there are two invisible contributors and only one of them sat in the room.
Insurance and indemnity language will follow. Publishing contracts already carry warranties that the work is original and does not infringe. Those clauses were written for plagiarism, and they are about to be tested against a question they were not drafted for. A writer who warrants originality while using tools trained on unknown corpora is making a promise they cannot personally verify. Expect that language to get longer before it gets clearer.
The value of documented process goes up. This is the part I find genuinely interesting. For most of my career, keeping meticulous records of interviews and drafts was a personal habit and nothing more. Clients did not ask. Now a writer who can produce a dated trail from conversation to chapter has something a client’s lawyer can actually use, and the writer who works from vibes and memory does not.
Are ghostwritten books in the AI training data?
This is the question that should worry ghostwriters more than the settlement does.
Ghostwritten books are in these training corpora. Hundreds of thousands of them. And a ghostwritten book is a peculiar artifact from a training perspective, because the voice in it belongs to one person and the prose belongs to another, and nothing in the file distinguishes them.
When a model learns from a memoir I wrote for a client, what has it learned? The client’s voice, which is what the book presents. But also my sentence construction, my paragraph rhythms, my habits of transition, my instinct for where a chapter should break. Those are mine. They appear in fifty-four books under other people’s names.
I have no standing to object. I signed away my rights to that text, as every ghostwriter does, and the person with standing is the named author who does not know my fingerprints are in there. The settlement’s claim process runs on registered copyright ownership. It does not have a field for “the person who actually wrote it.”
This is not a complaint. It is the deal I made and I would make it again. But it is a structural blind spot, and as these cases multiply somebody is going to notice that a large share of the world’s business books, memoirs, and celebrity autobiographies were written by people with no legal relationship to the resulting text.
What should clients ask a ghostwriter about AI?
If you are hiring someone to write your book, the settlement gives you four questions that were optional last year and are not anymore. None of them are hostile. A writer who cannot answer them has told you something useful.
What tools do you use, and where in the process? The answer should be specific. “I use AI for research and never for prose” is a real answer. “I use AI responsibly” is not one. You are entitled to know whether the sentences in your book were written by the person you hired.
Where does my material live while you work? Interview recordings, transcripts, and drafts sit somewhere. On a laptop, in a cloud service, inside a tool with terms of service nobody read. For a memoir containing things you have never told anyone, that answer matters more than the writing samples do.
What happens to it when we finish? Most contracts cover the manuscript and forget everything that produced it. Ask what gets deleted, what gets kept, and for how long.
Can you tell me where any particular passage came from? This is the one that separates a real ghostwriter from a fast one. I can go to any page of any book I have written for a client and tell them which interview it came out of. That is not a party trick. It is the difference between a book that is yours and a book that merely sounds like you.
A writer who bristles at these questions is telling you their process does not survive them.
How has publishing handled automation before?
Publishing has absorbed this kind of shock more than once, and the pattern is consistent enough to be useful.
When desktop publishing arrived, typesetters were told the machines would end their profession. What actually happened was that the ability to set type stopped being scarce and the judgment about how a page should look became more valuable, not less. The people who understood typography survived. The people who only knew how to operate the equipment did not.
Self-publishing did the same thing to the gatekeeper function. When anyone could publish, the scarce thing stopped being access and became discoverability, and an entire industry rearranged itself around that shift.
The through line is that automation removes the scarcity from the mechanical part and moves the value to the judgment part. Machines can now produce competent prose on demand. That makes competent prose worth less. It makes knowing which story to tell, which interview question opens someone up, and which chapter to cut worth considerably more.
The ghostwriters who will struggle are the ones selling words per hour. The ones selling judgment about a person’s story were never in competition with a text generator.
Does this change what a ghostwriting contract should say?
It should sharpen what is already there. A ghostwriting agreement has always specified who owns the work, who is credited, and what happens to the source material. What the settlement adds is a reason to be explicit about a fourth thing: whether any part of the work may be used to train a model, and by whom.
Most contracts written before 2023 are silent on that, because there was no reason not to be. Silence now means the question gets answered by whoever’s platform the manuscript passes through.
Four clauses worth having, in rough order of how often their absence causes a problem:
Tooling disclosure. State what AI tools, if any, are used in the process and at what stage. Research help and drafting are different things, and a client who is fine with one may not be fine with the other. Get it in writing, so the mismatch does not surface at chapter nine.
Training rights. State explicitly whether the manuscript, the interview recordings, or the transcripts may be used to train a model. The default answer should be no, and it should cover both parties, because the client uploading the draft to a consumer chatbot creates the same exposure as the writer doing it.
Platform handling. Name where the manuscript lives during the project and what the terms of service on that platform say about content rights. Most writers have never read those terms for the tools they use daily. The client is entitled to know whose servers their unpublished book is sitting on.
Interview material. Recordings and transcripts are the rawest version of a client’s voice that exists anywhere. Confidentiality clauses usually cover the manuscript and forget the source material. Fix that.
None of this is exotic. It is a page of contract language that prevents a conversation nobody wants to have later.
What about writers who use AI in their process?
Most working writers now use these tools somewhere, and pretending otherwise helps nobody. The line I hold is simple enough to state and easy to check: the tools can help me find things, organize things, and test things. They do not produce the prose that goes in the book.
That is not a moral position about machines. It is a practical one about what I am selling. A client is paying for a specific person’s judgment applied to their story. If a model wrote the chapter, they did not get that, and the fact that it reads well is beside the point.
The settlement gives this line a second justification it did not have before. If I do not put model output in the manuscript, questions about the provenance of that output never touch my client’s book. That is a cleaner answer than any disclosure clause.
Where it gets genuinely hard is research. Asking a model to summarize an industry, find a statistic, or explain a technical process is enormously useful and produces nothing that appears in the book. But it also hallucinates confidently, which means everything it tells you has to be verified against a real source before it informs a sentence you publish under someone’s name.
Does the $3,000 per work figure set a value for a book?
Three thousand dollars per work sounds like a lot until you divide a book’s development time by it. It is a settlement figure, not a valuation, and treating it as the market price of a book would be a mistake in both directions.
The more useful signal is that the case settled at all. A company with substantial resources decided that litigating the acquisition question was worse than paying to end it. That tells you where the legal exposure sits, and it is not in the training.
The second signal is the destruction requirement. Anthropic agreed to destroy the pirated files. That is the part that will shape behavior across the industry, because it establishes that the corpus itself is a liability with a cost attached, more than an input. Every AI company with a data acquisition team read that clause carefully.
What comes next in the AI copyright cases?
This is one case among many, and the others are still running. What this one settles is narrow: how these specific books were obtained by this specific company. What it does not settle is whether a model may learn from lawfully purchased text, whether output resembling a living author’s style infringes anything, or what a writer is owed when their work shapes a system that then competes with them.
Those questions are coming, and they will be decided by people who have never written a book. Which is a reason for writers to be conversant in the details, past the headlines. The commentary is going to be loud and mostly wrong in both directions, because the loudest voices in this argument have an interest in the outcome and very few of them have sat in a room for three hours getting someone to tell them the truth about their own life.
What I would watch for specifically: whether any court addresses output instead of input, whether the destruction remedy shows up in other settlements, and whether anyone raises the question of who wrote the works in these corpora as opposed to who owns them. That last one is where ghostwriting stops being an analogy and becomes a party to the argument.
For anyone writing under their own name or someone else’s, the practical takeaway is unglamorous. Know where your work is. Know who has a license to it. Write the clause. Keep the interview recordings somewhere you control. Be able to say where every word came from. If you want the wider context on how these questions fit together, the Ghostwriting Hub collects the pieces on contracts, ethics, and process in one place, and my ghostwriting service page explains how the interview trail works on an actual project.
The philosophical question about machines and creativity will outlast all of us. The contract question can be settled this afternoon.
The Guides That Get Your Book Written, Published, and Sold
Four short, practical guides on writing, publishing, and selling your book, plus the occasional note when there's something worth your time. No fluff, no daily inbox clutter. Drop your email and they're yours.
We use MailerLite to manage our list and send these emails. Your address is used only to send you what you signed up for. We will not sell it, share it, or use it for anything else, and you can unsubscribe anytime.