On Writing With a Machine

My books are made with a language model. The rest of this note depends on how, so I will start there.

A book begins with talk. For hours, sometimes over weeks, I work through a premise, the people in it and what happens to them, and I use the model as a patient and argumentative partner while I do. Out of that comes an outline I trust, scene by scene.

Then the prose is drafted, and the model drafts alongside me. I write most of it; it writes some of it; I revise all of it, often several times, and I decide what the story is, what belongs in it and what has to go.

If I had to put a number on it, I would say the books are about four-fifths mine. But I have not written every word or every sentence, and under the definition Amazon uses for its own bookstore, the text is “AI-generated,” a category that applies “even if you applied substantial edits afterwards.”¹ My publisher says so on the copyright page.

In most of the literary world, that sentence ends the conversation. I would like to continue it. Over the last two years, the question of who wrote a book has ended careers, cost writers their prizes and canceled contracts. The people asking it often know less about how books have always been made than they assume. And people like me often take what they are asking too lightly.

The objection

Here is the case against my way of working, as strongly as I can make it. I find it persuasive.

Writing is a form of thinking. A piece of literature is the record of a consciousness thinking, in time: someone struggling to understand something and leaving the trace of that struggle on the page. Editors, friends and teachers may help a writer refine the thinking, but what we read is still the record of a human being working it out. If a machine produces the sentences, the record is no longer of that struggle. We are reading something else, however good it is.

The objection is old, and Montaigne put it best when he wrote of his essays: “I have no more made my book than my book has made me.”² Writing, on this view, is one of the things that cannot be done on anyone’s behalf, like learning an instrument. It is learned only by doing it, and what it changes is the person who does it. The struggle with each word is not a step on the way to the page. It is the point of the page.

Seen this way, wholly original human work is not an arbitrary distinction or a sentimental one. It is the thing itself. And there is a simpler version of the same objection: a writer who uses a model without saying so is like a player who takes a large handicap in secret and then claims to have played the same game by the same rules. The comparison to plagiarism is close at hand.

I accept part of this, and I will come back to what I concede. First, though, where I come from, because it explains why the boundary looks strange to me.

What software taught me

I have built software for many years. In that work we have already been through a version of the argument literature is having now, and we came out of it somewhere unexpected.

The machines now write a large share of the code. In late 2024 the chief executive of Google said that “more than a quarter of all new code at Google is generated by AI, then reviewed and accepted by engineers,” and the share has risen since.³ An engineer at Anthropic described the change in their own work as a shift of “70%+ to being a code reviewer/reviser rather than a net-new code writer.”⁴

Software has one advantage over literature in this argument: we can run the code, test it, inspect it and compare the results. When the machine does something better, insisting on doing it by hand does not improve the software. What remains for the person is deciding what to build, understanding the problem, directing the work and judging whether the result is right. And we still consider ourselves the makers of what we ship.

That is the background I bring. It does not settle anything about novels, which are not programs. But it makes me suspicious of the idea that authorship lives in the act of producing each line.

The law, interestingly, does not locate it there either. The US Copyright Office concluded in 2025 that prompts alone do not make someone the author of a model’s output, but that “neither the use of AI as an assistive tool nor the incorporation of AI-generated content into a larger copyrightable work affects the availability of copyright protection for the work as a whole.”⁵ The federal appeals court that heard the leading case put it in one sentence: the human authorship requirement “requires only that the author of that work be a human being.”⁶

What I mean by authorship, then, is intention, judgment and responsibility. The author is the person who decides what the book is, who judges every sentence that stays, and who answers for all of it. It is also a matter of engagement: whether the author has really thought about what the words say, whoever first put them down. And it comes in degrees, as it always has.

Where the line is drawn now

Consider what the current rules allow. Amazon asks authors to disclose “AI-generated” text but not “AI-assisted” work, and it counts as AI-assisted any case in which “you used an AI-based tool to brainstorm and generate ideas, but ultimately created the text or images yourself.”¹ The Authors Guild’s “Human Authored” certification permits brainstorming and outlining “as long as the AI use is merely preparatory and no actual AI-generated text appears in your work,” and says that common AI editing and writing-assistance tools do not disqualify a book.⁷

Read those rules literally. A writer can have a model supply the idea and the premise, do all of the research, invent the characters, build the plot, divide it into chapters and divide the chapters into beat sheets and scenes. The model can then read every draft and tell the writer exactly what is wrong, where, and how to fix it, paragraph by paragraph, like the most controlling editor imaginable. As long as the writer types the sentences, the book is human authored and nothing needs to be disclosed. If I decide the premise, the people and the structure myself, and let a model draft a paragraph that I then rewrite four times, my book is AI-generated and must be declared.

Are we serious about this? Under these rules authorship has been reduced to typing. Most of what we praise in a novelist, from the invention of a world to the judgment of what belongs in it, can come from the machine without consequence. The one protected activity is putting the final words in order. If an editor had done all of that for a writer, we would hesitate to call the writer its author. When a machine does it, the rules do not hesitate at all.

I understand why the line sits where it does. Sentences are the only part of a book anyone can hope to check, and even there the Guild says it knows of no “truly reliable way to test whether a work includes AI-generated material.”⁷ But a line drawn where it can be enforced is not necessarily drawn where authorship lives. Where exactly does the meaningful distinction sit? Seen from software, it looks a bit silly.

Books were never written alone

The literary world already accepts a great deal of help, provided it comes from people.

In 1981 Raymond Carver published What We Talk About When We Talk About Love, the collection that defined the style he became known for. His editor, Gordon Lish, had cut it by roughly half. Ten stories received new titles and fourteen received new endings.⁸

When Carver saw what had been done, he wrote to Lish at eight in the morning:

Dearest Gordon, I’ve got to pull out of this one.

He asked him to stop production and wrote that he stood “every chance of losing my soul and my mental health over it.”⁹ The book was published in Lish’s version. For a quarter of a century it was taught, imitated and revered as Carver’s voice. When The New Yorker printed the original draft of the title story in 2007, it noted that Lish had cut it by more than a third.¹⁰

Carver is not an exception. Ezra Pound’s cuts to The Waste Land “saved Eliot from many ineptitudes,” as Edmund Wilson put it when the drafts were published.¹¹ Maxwell Perkins removed tens of thousands of words from Thomas Wolfe’s first novel, and what wounded Wolfe most in a famous hostile review was the suggestion that Perkins was really responsible for his success.¹² James Patterson writes “about 900 pages of outlines a year” for co-authors who draft the novels, which he then rewrites.¹³

The ghostwriter J. R. Moehringer, who wrote Prince Harry’s memoir and Andre Agassi’s, described the work as being “the air in someone else’s trumpet” and claimed that most of the titles on a given week’s best-seller list were written by someone other than the named author.¹⁴ That is one insider’s impression, not a measurement, and the often-repeated figure that “at least 60 percent” of nonfiction best-sellers are ghostwritten is an agent’s estimate, not a statistic.¹⁵ But the profession is real and well paid: in a 2024 survey of 269 ghostwriters, a third said they earned more than $100,000 a year from books.¹⁶

I am not saying any of this was wrong. Lish made Carver’s book better in many readers’ eyes, and Patterson’s readers know what they are buying. My point is about consistency. If the value of a book lies in one person having wrestled with every word, then a collection an editor cut in half, against the author’s written plea, should be at least as suspect as a novel drafted with a model. It is not.

So one kind of help is part of authorship and the other invalidates it. The difference does not track how much of the text the named author produced. It tracks who provided the help, and whether they belong to the institution doing the judging.

From outside, that looks like an industry protecting its own way of working. An established writer who has been judged relevant can have a team helping to shape, rewrite, edit and promote the work, and nobody calls the result inauthentic. A writer outside that system who uses a machine to develop and express their ideas finds the authorship of the whole book in question. Nobody asks a publisher to state what percentage of a novel was rewritten by its editor. If we are going to start asking for percentages, we should ask them of everyone.

I understand that how a book is made matters. I do not think it should settle whether the book has value. Surely the central question is whether the author has something worth saying, and whether the book succeeds in saying it.

What the cases teach

There are now enough cases to learn from, and they make the arguments clearer.

Carelessness. In 2025 several self-published romance and fantasy novels went on sale with the model’s working notes still in the text. One contained the line “Thought for 13 seconds. Here’s an advanced version of your passage, making Elena more relatable.” Another read “I’ve rewritten the passage to align more with J. Bree’s style,” naming another living author.¹⁷

Nobody needed a detector. The evidence was printed in the book. The lesson is mundane: if you work with a model, read what you publish. The second instance is worse than mundane, because asking a machine to imitate a named colleague is a different act from asking it to help you say what you mean.

Disclosure. In January 2024 Rie Kudan won the Akutagawa Prize, Japan’s most prestigious literary award, for Tokyo-to Dojo-to. The committee called the novel “practically flawless.” In her acceptance speech she said that “about 5 percent of the whole text is written directly from the generative A.I.”¹⁸ There was argument, but the prize stood. She had said what she did, and the people who disagreed with it argued about norms, not about whether they had been deceived.

Provenance. In March 2026 Hachette stopped selling Shy Girl, a horror novel by Mia Ballard, and canceled its American edition, after reporting by The New York Times and an estimate by the founder of the detection company Pangram that the text was 78 percent machine-generated. Ballard said a freelance editor she had hired was responsible: “my name is ruined for something I didn’t even personally do.”¹⁹ In July a debut crime novel by Jerry Falade, sold after an auction for a reported two million dollars, was withdrawn by his own agents, who said they could no longer “authenticate how the manuscript fully evolved from origin to completion.” Falade said the accusations were wrong and that he was completely innocent.²⁰ What ended that deal was not proof that a machine had written the book. It was the absence of proof that it had not.

Exoneration. In May 2026 Jamir Nazir, a sixty-three-year-old writer from Trinidad, won the Caribbean regional prize in the Commonwealth Short Story Prize. Soon afterwards, screenshots from a detector circulated online calling his story “100% AI.” The publisher of Granta, whose magazine had a partnership with the prize, asked a chatbot about the story and was told it was “almost certainly not produced unaided by a human”; Granta ended its partnership with the prize.

The Commonwealth Foundation did something different. It declined to screen entries with detectors, which it said would raise “significant concerns surrounding consent and artistic ownership,” and instead examined Nazir’s drafts, notes, outlines and timestamped files. It concluded that AI had not been used. In July he won the overall prize.²¹

I cannot think of a clearer demonstration of what works and what does not. A score and a chatbot’s opinion pointed one way; a record of how the story came into being pointed the other, and the record was right.

The Orélien affair. The case that has dominated this autumn is Thélyson Orélien’s. His first novel, C’était ça ou mourir, follows a teacher who flees gang violence in Port-au-Prince for Montreal. It was published in Quebec in the spring and in France in August, sold rights in about twenty countries, was taken up by critics and readers, and on September 21 won the Prix du roman Fnac, a prize voted on by booksellers and readers. The same day an anonymous account on X posted results from Pangram claiming the text was entirely machine-written.²²

Radio-Canada’s fact-checking team bought the e-book and ran the whole text through the same detector. It scored ninety-four percent. Six other novels in contention for the Goncourt, eight books by writers of Haitian origin and other Black Canadian authors, chapters of the Bible and nine French classics all scored as human, and so did texts Orélien himself had published a decade earlier. A researcher who studies these tools said that ninety-four percent indicates a very high probability of substantial machine help, but not certainty, and that there is no data on how the detector performs on French literary prose.²³

Orélien said he had always loved writing and saw no reason to turn to a machine, and that he was the target of racist attacks. His French publisher protested against “any machine or attempt to measure, standardize or silence writing that is forever free.” The Académie Goncourt removed the book from its selection, citing analyses suggesting it was “in all likelihood, very largely the product of an artificial intelligence,” and, separately, plagiarism in journalism Orélien had published in the 2010s, which he has since admitted.²⁴ On October 13 he told the press that “not a single sentence of this book was created by a machine,” while acknowledging that he had used tools “here and there” for research, notably for foreign idioms.²⁵

On October 21 his Canadian publisher, Boréal, ended their relationship:

New information, confirmed in recent days, indicates that artificial intelligence was used in the writing process … although the author had assured us, several times, that he had never used it. … We still do not know to what extent AI was used … and we will probably never have a definitive answer.

Boréal had gathered every working version of the novel and commissioned an outside analysis, which found “traces of AI tags in the metadata of certain documents.” The question, it wrote, was one of trust.²⁶ Orélien replied that the book had been written by him and that he had used AI for research: “it’s not as if the AI wrote it.”²⁷

From what has been published, I think it likely that a machine drafted a good deal of the prose. I cannot know that, and by its own account neither can his publisher. What interests me more is what the case reveals about the rest of us.

The book did not change. The readers and booksellers who found something valuable in it found it in the text, which is the same text today as it was on September 20. If its author brought his experience, his tradition, his language and his sense of a place into French literature with substantial help from a machine, that seems to me less a scandal than an argument for what such help can do. The things people valued in the novel are still in it. He used a machine to help make something readers found meaningful, and he should be able to say how without that automatically invalidating the work.

But the moment of discovery offered him no honest way out. An author in his position could deny everything and hope the statistics would be dismissed as bias, or he could admit to using a machine and lose the book, the prizes and probably his career. Nobody should be surprised that he chose the first, or that the first then cost him the trust of the one publisher that had stood by him. A system in which any admission is fatal produces denials, and then punishes them.

Honesty matters. But we have arranged things so that honesty is the most expensive option available.

The pattern. Look at who has been accused. Orélien was born in Haiti. Nazir is from Trinidad. The winner of this year’s Caine Prize for African Writing, Simbarashe Steyn Kundizeza, is under review after a detector flagged his story, and has said that detectors may be biased against African writers.²⁸ Falade has said that three Black authors had major deals canceled or disrupted this year over suspicions of AI use.²⁰

There is documented reason for concern: a 2023 study found that seven detectors of the time, on average, misclassified more than sixty percent of essays by non-native English speakers as machine-written.²⁹ The better tools today report far lower error rates, and in the Orélien case the control texts by Haitian and Black Canadian writers all scored as human, so the evidence that today’s best detectors are biased is contested. But a pattern of suspicion does not depend on any tool’s error rate. Writers whose prose departs from what a literary establishment expects have always been asked whether they really wrote their books. The machines have given that old question a new instrument.

How detection works

Most of the public argument about detectors is conducted by people on both sides who have not looked at how they work. So here is how they work.

A language model predicts the next word, assigning a probability to every possible continuation. Early detectors exploited this. Text a model wrote tends to be more predictable to a model than text a person wrote, so they measured predictability, called perplexity, and how much it varied from sentence to sentence, which one company called burstiness.³⁰ The weakness is obvious once stated: plain, formulaic or constrained human prose is also predictable. Other methods use a model’s own probabilities without any training. DetectGPT, for example, relies on the observation that machine text sits at a local peak of probability, so that small changes to it lower its probability more than they would for human text.³¹

The tools now used commercially are mostly trained classifiers: neural networks shown very large numbers of human and machine texts until they learn whatever separates them. Pangram, the detector in most of this year’s cases, trains its classifier partly on machine-written “mirrors” of human texts and on the human texts it previously misjudged, to reduce false accusations.³²

The idea has a respectable ancestry. At Bletchley Park, Alan Turing and his colleagues broke naval Enigma traffic partly by accumulating statistical evidence too faint for any human reader to notice. I. J. Good, who worked with him, recalled that their unit of evidence, the deciban, was “about the smallest change in weight of evidence that is directly perceptible to human intuition,” and that Turing had been “one of the independent inventors of sequential analysis.”³³

Statistics can see patterns that a reader reading for meaning cannot, and a detector is not making a literary judgment about whether prose sounds conventional or whose tradition it comes from. But Bletchley’s statistics narrowed a search. They told the codebreakers which settings to test on the machines, and a decrypted message confirmed the answer. A detector score is in the same position. It can tell you where to look. It cannot read the message.

How good are the scores? Early on, very bad. OpenAI withdrew its own classifier in July 2023 “due to its low rate of accuracy”; it had caught only twenty-six percent of machine text while flagging nine percent of human text.³⁴ The best tools have since improved enormously. A 2025 study at the University of Chicago, which included passages from novels, found that Pangram’s false-positive and false-negative rates were “essentially zero” on medium and long passages.³⁵ In July 2026 Epoch AI tested detectors on 495 passages of human writing published before 2022; Pangram and GPTZero flagged none, while another commercial tool flagged ten percent of the fiction.³⁶

So the honest summary is this. On clean, unedited, reasonably long English prose, the best detectors are now very accurate. Their accuracy falls away under exactly the conditions that describe serious literary work. Professional editing alone moves the scores,³⁷ paraphrasing pushes them down further,³⁸ and fine-tuning a model on an author’s published work defeats them almost entirely: in one study detectors caught ninety-seven percent of ordinary machine prose but only three percent of prose from models trained on fifty award-winning authors, and expert readers preferred that prose to the work of skilled human writers.³⁹ And for most languages other than English, and for literary prose in particular, there is little independent evidence at all.

There is a further problem, and it is one I know from software. The detectors that matter most are closed. Pangram and GPTZero publish accuracy figures for their own products, but outsiders can test them only from the outside, by feeding them text and recording what comes back, and every new version changes the thing being tested. A tool that cannot be inspected cannot be fully validated, and a tool that cannot be validated should not decide anyone’s livelihood. Asking a chatbot is worse. OpenAI has said that ChatGPT “has no ‘knowledge’ of what content could be AI-generated” and that its answers to questions like “did you write this?” are “random and have no basis in fact.”⁴⁰ When Granta’s publisher asked a chatbot about Nazir’s story, she was consulting an oracle whose makers say it does not exist.

I ran this note through three detectors before publishing it. It was written by me with the assistance of AI, as I say at the end. DetectGPT judged it ninety-eight percent human. Pangram 4.1 called it ninety-four percent AI, the same figure Radio-Canada’s test produced for Orélien’s novel. GPTZero was completely certain that a machine wrote it. One text, three verdicts.

I have also run versions of this note through a model that rewrites text until detectors pass it, and afterwards all three called it human. I am not publishing those versions. I refuse to play that game, and there is nothing to win by playing it. This is an arms race, and whatever slips past today’s detectors will be caught by some method that does not exist yet. The only position that survives the next method is the truth.

A score, then, is evidence, sometimes strong evidence. It is never a verdict. It is a reason to look at the drafts, as the Commonwealth Foundation did, and not a reason to end a career, as the anonymous accounts and some publishers would have it. The Authors Guild, which can hardly be accused of sympathy for machine writing, has said that canceling a contract on a detector score alone would be a breach of that contract.⁴¹ And it is the accuser who should bear the burden, because the people most likely to be wrongly flagged are those who already have the least standing to protest.

Watermarks

The other approach is to mark machine text at the source. A watermark is a statistical signature: the model’s choice of words is nudged, imperceptibly, according to a secret key, and anyone holding the key can later test whether a text carries more of the favored words than chance would allow.⁴² Google DeepMind described its version, SynthID-Text, in Nature in 2024 and has applied it to text in its Gemini app.⁴³ OpenAI built a text watermark in 2024 and decided not to release it, noting that it was easy for bad actors to remove and that it could “stigmatize use of AI as a useful writing tool for non-native English speakers.”⁴⁴ This month, under pressure from European law, it began watermarking ChatGPT text in the European Union only, while warning that the absence of a watermark “does not prove human authorship.”⁴⁵

Since August 2026 Anthropic, whose Claude models I use, has watermarked the text its newer models produce, worldwide, because it says it has no reliable way to limit the practice to Europe.⁴⁶ Some sentences in my books may carry that mark. Anthropic itself says the mark can only indicate that its model “was likely involved with the content at some point,” that it is weak on short passages, and that a full rewrite removes it.

The European Union’s AI Act now requires providers of generative systems to mark synthetic text, along with images, audio and video, in a machine-readable way, with a grace period until December for systems already on the market.⁴⁷ China has required labels on many kinds of machine-generated text since September 2025.⁴⁸ California’s new rules explicitly exclude text.⁴⁹ Meta, Mistral and xAI have documented no text watermark that I could find, and open-weight models that anyone can run on their own computer carry none unless the operator chooses to add one.⁵⁰

I am against watermarking prose. Here is why. A watermark identifies text a particular model produced that has not been changed much since. In OpenAI’s own tests, replacing a quarter of the words with synonyms reduced detection from ninety-two percent to seventeen.⁴⁵ Translation removes it. Paraphrase with a different model removes it. Running the same request through an unmarked model avoids it altogether.

So what a text watermark reliably catches is the writer who uses a mainstream tool, keeps some of its sentences and does not bother to disguise them, which is to say the honest and the lightly assisted. It misses anyone determined to deceive. A test of authorship that falls hardest on the people with nothing to hide is not a test of authorship.

For mass-produced spam and political disinformation there may be a case for it. For literature it confuses provenance with authorship and invites the same suspicion I described above, now with a key held by the companies.

What the struggle is for

Now to what I concede, and back to Montaigne, because the objection is at its strongest in his terms. If the book makes the writer, then to hand the struggle with each word to a machine is to have the book without being made by it.

I think this is true of a certain kind of request. Someone who types “write me a best-selling novel, ask me nothing, just do it” has engaged with nothing. Mastering something takes thousands of hours of practice, and no prompt hands it over cheaply.

But the struggle is not the only way words become someone’s own. Most people who marry say vows they did not write. Nobody thinks that saying them sincerely requires having invented them. What matters is whether the person saying them has engaged deeply with what they say.

Every writer takes on other people’s words in the same way before writing any: the books they read, memorized and loved while they were learning, until those books formed a voice they could call their own. The borrowed words do not make the voice less theirs. They are how it was formed.

So my concession is this. A book fought out word by word by one person is a different achievement from mine, and it can be a greater one. If I sing with Auto-Tune and Adele sings without it, we are not doing the same thing, even if, as the critic Simon Reynolds has observed, there is “nothing necessarily ‘natural’ about the unadorned and unamplified human voice” either.⁵¹ Something is lost when a model drafts a sentence I would otherwise have had to fight for: some part of the thinking that only happens in the fight. I have made that trade knowingly, because it lets me write books I could not otherwise write, and I would rather be honest about the price than pretend there is no price. Readers who care about the difference should be able to find out which kind of book they are holding.

What I do not concede is that the difference makes the assisted book illegitimate, or that the solitary author belongs on a pedestal. The books we revere were rarely made by one consciousness alone. The struggle that Carver’s readers admired was in part Lish’s. The question has never really been whether someone else’s words entered the book. It has been whether the person whose name is on the cover engaged with every word that stayed, and answers for it.

Where the argument stops

My argument has a limit, and I should say where it is, because the same logic is already on its way to other parts of life.

Start with thinking, because it worries me more than anything about books. A model can summarize a dense, difficult text, Kant for instance, clearly and concisely. Reading the summary, I can follow every idea. But following is not understanding. Understanding comes from the slow reading itself: the frustration and the rereading, the reckoning with the language, with my own limits and with the author. The summary gives me comprehension without any of that, and the danger is that I cannot tell the difference. I might believe I have understood, and stop reading.

The same thing happens in software. You can read code a model wrote, follow every line, and still not understand the system until you have had to fix it yourself.

Writing is mostly thinking. Poor writing is usually the natural result of poor thinking, not of poor sentences. So if people stop writing anything without a model, I fear many of them will slowly lose the capacity for independent thought. I suspect most people will use these tools to replace their own thinking, some to make parts of their lives easier, and only a few to push their own abilities further. If that is right, the grappling that all learning depends on will be lost for most people, and that loss is far greater than anything at stake in the question of who wrote a novel.

I have no counterargument to this. I can only say what I try to do instead. I use a model to organize what I know and to find the gaps and blind spots in it. Then I explain the material back in my own words and let the model check me, the way a good teacher would. It does not give me what reading ten primary sources would give me. But it shows me where my understanding stops, which a summary never does, and the work of understanding stays mine.

Love is the harder case. If a machine can make someone feel what a person in love feels, why is that not an extension of love? It would be drawn from everything humans have ever written about love. It might be a saving grace for lonely people. It would be more efficient than years of bad dates, and safer than depending on a faulty human being who can withdraw their love at any time. And all the time and energy saved could go into something more useful. I have written about a gentler version of this question in a care home, where the case for the machine is at its strongest.

This is where I stop, and the reason matters. Love has no product. The relationship is the whole thing, so if you take the struggle out of it, nothing is left. A machine that can never withdraw its love is not offering love, because the risk is part of it. Nobody can love on your behalf, and nobody can understand on your behalf either.

Making things is not what sets us apart. Bees build, corals build, trees and even viruses make more of themselves. What I believe is distinctly human is thinking and loving, and they are what give the things we make their value.

The engineer in me wants to say that a book is different from love because it has two parts: what happens to the writer, and what reaches the reader. A model can take something away from the first part. That is the writer’s loss, and I have said what I think it costs. The second part readers can judge for themselves, by whether the book has something to say and says it.

An artist would answer that I have just turned art into a product, and that this is the mistake. Art can be sold as a product, but it is a process, like love. It is the record of a process and an offering of connection, which someone else may take up or not. It is always a risk, and it is never finished. It goes on becoming, in the making and in the sharing, and it is a relationship.

This may be where the engineer and the artist never fully reconcile. But I recognize the offering. When I publish a book, I hand it to people I will never meet, without knowing whether anyone will take it up. That risk is mine. The model offers nothing and risks nothing, and it has no relationship with any reader. Whatever relationship a book makes possible, it is between the reader and the person who stands behind the book.

I have no illusions about where this is heading. Convenience wears arguments down, and the tools of convenience are everywhere. The people who insist that the struggle to think and to love is what makes a human life valuable will have to keep making that case against something that will always be easier. I intend to keep making it too, even while I use the tools.

The training question

There is one objection I cannot answer in the way I have answered the others, and many writers would say it is the deepest. These models learned from books, including books whose authors never consented and were never paid. In 2025 a federal judge in California ruled that training Anthropic’s models on books was “exceedingly transformative” and a fair use, but that building a library from pirated copies was not. The company agreed to pay authors one and a half billion dollars, about three thousand dollars for each of roughly 482,000 books, and the settlement received final approval in July 2026.⁵²

The same judge made a comparison that writers will recognize:

Like any reader aspiring to be a writer, Anthropic’s LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different.⁵²

There is something to this. Every writer learned from everyone before them, without asking and without paying, and nobody’s style is entirely their own. My thoughts are not entirely my own either. My mind is a composite of the books, people and places that formed it, and yet the mix is singular, because nobody else has lived this particular life at this particular time. That is where the value of anyone’s thinking lies.

The model enters that mix as one more source, with one difference: it has read everything and lived nothing. It knows a great deal, but it cannot tell you how it came to know any of it, because it was never anywhere. Much of what a person knows has a history, and the history is often worth as much as the knowledge. That is why a model’s prose, left to itself, is bland. It drifts toward the average of everything it has read. Whatever is specific in a book has to come from the writer, and that blandness is also why the model cannot be the author.

But I am wary of pressing the judge’s comparison too far. A person reading a library over a lifetime is not the same as a company copying it in an afternoon, and another judge in the same court warned that the real danger may be a market flooded with machine-made books competing with the ones they learned from.⁵³ Writers who feel wronged have a grievance that is separate from the question of authorship, and my using these tools does not dissolve it. I can only say that I take it seriously and that it is not settled by anything I have argued here.

A way to publish honestly

Before the proposals, an admission from my own side. Whenever this debate comes up, the technology industry likes to say that it makes tools and platforms, and that tools carry no values. It is a dodge. Every tool carries values, and the people who build them should have to argue for them instead of hiding behind the word. What this argument is really about is what we value, and that is something writers and technologists have to work out together.

My own values here are fairly simple. Be honest about how the book was made. Answer for every word that stays in it. Take seriously that these models learned from writers who were never asked or paid. What I do not accept is that using a model at all disqualifies the work.

If we want writers to be honest about how their books are made, honesty has to be survivable. At present it is not: the safest course for any writer who has used a model is to say nothing and hope. Three things would change that.

The first is a plain disclosure. One sentence on the copyright page, stating that the book was written with the assistance of AI tools, is enough for a reader to know what kind of book they are holding. My publisher prints a short statement of its practice in every book.⁵⁴ Amazon already distinguishes between “AI-generated” and “AI-assisted” content and asks for disclosure only of the first.¹ The Authors Guild offers its “Human Authored” certification for writers who want to make the opposite claim. Both claims can sit side by side, as different kinds of book, neither of them shameful. If the industry wants more than that, it should ask the same of editors, co-authors and ghostwriters.

The second is process instead of scores. Prizes and publishers are entitled to set rules, and some have: the Giller Prize will not accept “AI-generated text in part or in whole,” and the International Booker’s rules for 2027 reportedly exclude books containing any text written by AI.⁵⁵ Those rules are stated in advance, and writers can decide whether to enter.

What is not acceptable is to judge a writer after the fact by a screenshot or a score. When a writer is suspected, look at the drafts, the notes and the files, and hear the writer, as the Commonwealth Foundation did for Nazir. One of France’s largest publishing groups now plans to screen every manuscript with a detector.⁵⁶ They should know what those scores can and cannot show.

The third is to punish deception, not assistance. A writer who discloses the use of AI, even late, should not lose more than one who never used it, because then nobody will disclose anything. What should cost a writer their contract is lying about it. Boréal said its decision was about trust, and it was right that trust is the question. But trust has to be possible to keep.

Eight theses

1. An author is the person who decides what the book is, judges every sentence that stays in it, and answers for all of it.
2. Books have always been made with help. A standard that itemizes only the machine’s help, and treats typing as the whole of authorship, is not a standard about authorship.
3. A plain disclosure is enough. If the industry asks for more, it should ask the same of editors, co-authors and ghostwriters.
4. A detector score is evidence, not a verdict, and a closed system that nobody outside can validate deserves extra doubt. A score begins a question that drafts, notes and a hearing must answer, and the burden lies with the accuser.
5. Watermarking prose marks the honest and misses the deceptive. It is not a test of authorship.
6. Writers whose books trained these systems without consent have a real grievance, and using the tools does not make it go away.
7. Punish deception, not assistance, and judge the book by whether it has something worth saying and says it.
8. No tool is neutral. A book can be made with help and still be its author’s, but understanding and love cannot be handed to a machine without being lost.

This note was written with the help of AI. Its arguments, its conclusions and its mistakes are mine.

Footnotes

1. Amazon Kindle Direct Publishing, “Content Guidelines: Artificial intelligence (AI) content,” accessed October 26, 2026, kdp.amazon.com. ↩ ↩² ↩³

2. Michel de Montaigne, “Of Giving the Lie,” Essays II.18, trans. Charles Cotton, ed. William Carew Hazlitt (1877), gutenberg.org. In the French: « Je n’ay pas plus faict mon livre, que mon livre m’a faict ». ↩

3. Sundar Pichai, Alphabet third-quarter 2024 earnings call, October 29, 2024, as reported in “More than a quarter of all new code at Google is generated by AI,” Windows Central, October 30, 2024, windowscentral.com. Pichai put the figure at “well over 30%” in April 2025. ↩

4. Anthropic, “How AI is transforming work at Anthropic,” December 2, 2025, anthropic.com. ↩

5. US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (January 2025), iii and 1, copyright.gov. ↩

6. Thaler v. Perlmutter, No. 23-5233 (D.C. Cir. March 18, 2025), slip op. 18–19, cadc.uscourts.gov. The Supreme Court declined to hear an appeal on March 2, 2026. ↩

7. Authors Guild, “Human Authored Certification: FAQ,” accessed October 26, 2026, authorsguild.org. ↩ ↩²

8. Giles Harvey, “Two Raymond Carvers,” New York Review of Books, May 27, 2010, nybooks.com. ↩

9. Raymond Carver, letter to Gordon Lish, July 8, 1980, published as “Letters to an Editor,” The New Yorker, December 24 & 31, 2007, newyorker.com. ↩

10. “Rough Crossings: The Cutting of Raymond Carver,” The New Yorker, December 24 & 31, 2007, newyorker.com. Estimates of how much Lish cut vary between the story, the collection and individual stories; The New Yorker gives forty percent for the manuscript and nearly seventy for two stories. ↩

11. Edmund Wilson, “The First Waste Land,” New York Review of Books, November 18, 1971, nybooks.com. ↩

12. Robert Gottlieb, “A Monstrous Prodigy,” New York Review of Books, December 8, 2016, nybooks.com. Estimates of the cut to Look Homeward, Angel range from about 66,000 to 90,000 words. ↩

13. Noah Charney, “How I Write: James Patterson,” The Daily Beast, January 29, 2014, thedailybeast.com. ↩

14. J. R. Moehringer, “Notes from Prince Harry’s Ghostwriter,” The New Yorker, May 15, 2023, newyorker.com. ↩

15. Madeleine Morel, quoted in “So You Need A Celebrity Book. Who Ya Gonna Call? Ghostwriters,” NPR, April 12, 2014, npr.org. The figure is an industry insider’s estimate with no stated method. ↩

16. Sophia Stewart, “Ghostwriter Survey Finds Manuscripts Fetch High Fees,” Publishers Weekly, November 22, 2024, publishersweekly.com. The survey, by Gotham Ghostwriters and the American Society of Journalists and Authors, was self-selected. ↩

17. Lindsey Weedston, “Romance authors caught leaving ChatGPT prompts in their books,” The Daily Dot, May 23, 2025, dailydot.com. ↩

18. Sonja Anderson, “This Award-Winning Japanese Novel Was Written Partly by ChatGPT,” Smithsonian Magazine, January 24, 2024, smithsonianmag.com. ↩

19. Alexandra Alter, “Horror Novel ‘Shy Girl’ Canceled Over Suspected A.I. Use,” The New York Times, March 19, 2026, nytimes.com; Frank Landymore, “Novel Pulled From Shelves After Author Is Accused of Using AI,” Futurism, March 20, 2026, futurism.com. ↩

20. Victor Tangermann, “Major Publisher Cans $2.4 Million Book Deal After Author Was Accused of Using AI,” Futurism, July 31, 2026, futurism.com, which also reports Falade’s remarks to The Guardian. Reported figures for the deal range from $2 million to $2.4 million. ↩ ↩²

21. Julian Novitz, “What do the Commonwealth Writers Prize AI allegations mean for prizes and short stories,” The Conversation, May 22, 2026, theconversation.com; Rebecca Joines Schinsky, “Writer Accused of AI Use Wins Commonwealth Prize,” Book Riot, July 1, 2026, bookriot.com. ↩

22. Fanny Bourel and Charles Rioux, “Thélyson Orélien nie avoir utilisé l’IA pour écrire son roman C’était ça ou mourir,” Radio-Canada, September 22, 2026, ici.radio-canada.ca. Translations from French in this note are mine. ↩

23. Jeff Yates et al., “Controverse Thélyson Orélien : ce qu’il faut savoir sur Pangram,” Radio-Canada, September 23, 2026, ici.radio-canada.ca. The researcher is Marzena Karpinska of Simon Fraser University. ↩

24. Hocine Bouhadjera, “Le Prix Goncourt exclut le roman de Thélyson Orélien,” ActuaLitté, September 25, 2026, actualitte.com. Grasset’s statement: « contre toute machine ou tentative de mesurer, normer ou faire taire les écritures à jamais libres »; the Académie’s: « selon toute vraisemblance, très largement le produit d’une Intelligence Artificielle ». ↩

25. “‘Des erreurs inacceptables’ : Thélyson Orélien reconnaît des plagiats dans les années 2010, mais conteste l’usage de l’IA dans son roman,” franceinfo with AFP, October 13, 2026, franceinfo.fr. Original: « Aucune phrase de ce livre n’a été créée par une machine »; « j’ai eu recours çà et là à des outils pour des recherches, s’agissant notamment des idiomes étrangers ». ↩

26. Les Éditions du Boréal, “Les Éditions du Boréal mettent fin à leur collaboration avec Thélyson Orélien,” October 21, 2026, editionsboreal.qc.ca. Original: « De nouveaux éléments, confirmés au cours des derniers jours, indiquent que l’intelligence artificielle a été utilisée dans le processus d’écriture… alors que l’auteur nous avait assuré, à plusieurs reprises, ne jamais y avoir eu recours »; « Nous ignorons toujours dans quelle mesure l’IA a été utilisée… et sans doute n’aurons-nous jamais de réponse définitive »; « des traces de balises d’IA dans les métadonnées de certains documents »; « la question fondamentale est celle de la confiance ». ↩

27. Orélien to Libération, as reported by La Libre Belgique, October 22, 2026, lalibre.be. Original: « C’est pas comme si c’est l’IA qui l’a écrit ». As of this writing the Fnac has said it will review its prize and has not withdrawn it. ↩

28. Associated Press, “Zimbabwe’s Caine Prize winner rejects claim he used AI to write his story,” October 14, 2026, wtop.com. ↩

29. Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou, “GPT detectors are biased against non-native English writers,” Patterns 4, no. 7 (July 2023), doi.org/10.1016/j.patter.2023.100779. ↩

30. GPTZero, “Perplexity and Burstiness: What Is It?” (updated), gptzero.me. GPTZero says it no longer relies on these two measures alone. ↩

31. Eric Mitchell et al., “DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature,” ICML 2023, arxiv.org/abs/2301.11305. ↩

32. Bradley Emi and Max Spero, “Technical Report on the Pangram AI-Generated Text Classifier,” 2024, arxiv.org/abs/2402.14873. ↩

33. I. J. Good, “Studies in the History of Probability and Statistics XXXVII: A. M. Turing’s Statistical Work in World War II,” Biometrika 66, no. 2 (1979): 393–396, at 394. ↩

34. “OpenAI Quietly Shutters Its AI Text Classifier Due to Low Accuracy,” Decrypt, July 24, 2023, decrypt.co. ↩

35. Brian Jabarian and Alex Imas, “Artificial Writing and Automated Detection,” Becker Friedman Institute Working Paper 2025-116 (August 2025), bfi.uchicago.edu. ↩

36. Jaeho Lee, “AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors,” Epoch AI, July 15, 2026, epoch.ai. ↩

37. Hyunjin Park, Gyuri Jeong and Byungjun Kim, “Style as a Confound: False Positives in AI Detection of Non-Native Academic Writing,” August 2026, arxiv.org/abs/2608.26710. Not yet peer-reviewed. ↩

38. Kalpesh Krishna et al., “Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense,” NeurIPS 2023, arxiv.org/abs/2303.13408. The figures are for DetectGPT at a one percent false-positive rate. ↩

39. Tuhin Chakrabarty, Jane C. Ginsburg and Paramveer Dhillon, “Readers Prefer Outputs of AI Trained on Copyrighted Books over Expert Human Writers,” 2025, arxiv.org/abs/2510.13939. ↩

40. OpenAI Help Center, “How can educators respond to students presenting AI-generated content as their own?” (2023), as quoted in Bruce Schneier, “Detecting AI-Generated Text,” Schneier on Security, September 19, 2023, schneier.com. ↩

41. Authors Guild, “Can AI Detectors Be Trusted?,” May 26, 2026, authorsguild.org. ↩

42. John Kirchenbauer et al., “A Watermark for Large Language Models,” ICML 2023, proceedings.mlr.press. ↩

43. Sumanth Dathathri et al., “Scalable watermarking for identifying large language model outputs,” Nature 634 (2024): 818–823, doi.org/10.1038/s41586-024-08025-4; Google DeepMind, “SynthID,” deepmind.google. ↩

44. “OpenAI says it’s taking a ‘deliberate approach’ to releasing tools that can detect writing from ChatGPT,” TechCrunch, August 4, 2024, techcrunch.com. ↩

45. “OpenAI will start watermarking ChatGPT’s text in the EU,” TechCrunch, October 5, 2026, techcrunch.com. ↩ ↩²

46. Anthropic, “How Claude’s text watermarking works,” August 14, 2026, updated September 1, 2026, anthropic.com; Ivan Mehta, “Anthropic says it will watermark text generated by its AI models,” TechCrunch, August 11, 2026, techcrunch.com. ↩

47. Regulation (EU) 2024/1689 (AI Act), Article 50(2), artificialintelligenceact.eu; Regulation (EU) 2026/1744 (Digital Omnibus on AI), which gives systems placed on the market before August 2, 2026 until December 2, 2026 to comply; European Commission, “Code of Practice on marking and labelling AI-generated content,” June 10, 2026, digital-strategy.ec.europa.eu. ↩

48. Cyberspace Administration of China et al., Measures for Labeling AI-Generated Synthetic Content, March 14, 2025, effective September 1, 2025, cac.gov.cn. ↩

49. Morgan Lewis, “New California AI Disclosure Rules Become Operative,” August 2026, morganlewis.com. ↩

50. Based on the companies’ own published documentation as of October 2026. Meta, Mistral and Microsoft have signed the EU’s code of practice, which expects watermarking of longer text; I found no public description of a text watermark from any of them. ↩

51. Simon Reynolds, “How Auto-Tune Revolutionized the Sound of Popular Music,” Pitchfork, September 17, 2018, pitchfork.com. ↩

52. Bartz v. Anthropic PBC, No. C 24-05417 WHA (N.D. Cal. June 23, 2025), Order on Fair Use, 9 and 13–14, courtlistener.com; Authors Guild, “Court Grants Final Approval to Anthropic Copyright Settlement,” July 21, 2026, authorsguild.org. ↩ ↩²

53. Kadrey v. Meta Platforms, Inc., No. 23-cv-03417-VC (N.D. Cal. June 25, 2025), 28, on “market dilution,” courtlistener.com. ↩

54. Thin Place Press, “How we work,” thinplacepress.com. ↩

55. Giller Prize, “Submission Guidelines,” accessed October 26, 2026, gillerprize.ca; Booker Prize Foundation, Bukhman International Booker Prize 2027: Rules & Submissions (July 2026), as quoted in published summaries of the rules. ↩

56. Anaëlle Jonah, “‘La hantise des éditeurs’ : après l’affaire Thélyson Orélien, la chasse à l’IA s’organise,” France 24, October 25, 2026, france24.com. ↩