A polygraph isn’t admissible in an American courtroom.
The reason is simple enough: the machine measures how a body behaves under stress, and a body under stress behaves that way for more than one reason. An innocent person who’s frightened produces the same trace as a guilty person who’s calm about the wrong question. Courts decided a long time ago that a test which can’t separate those two cases has no business deciding whether somebody goes to prison.
Publishing has spent this year adopting a polygraph.
The Wall Street Journal counted the wreckage in August.
A literary agency withdrew support for a crime novel it had already sold, telling the publishers it could no longer verify that its client had written the book. Hachette canceled the American release of a horror novel in March after similar allegations, which the author denied, saying an acquaintance had used AI while editing. The Atlantic reported that a self-published romance picked up by Simon & Schuster carried hallmarks of AI writing, citing a Stony Brook researcher who had run the text through a detection tool.
That paper hasn’t been peer reviewed. Simon & Schuster is standing behind the author, and said plainly that it doesn’t think conclusions about a writer’s work should come from tools known to produce false positives.
I’ve been a ghostwriter for more than a decade.
My name is on none of the books I write for clients. That means I’ve spent my whole career inside the exact question publishing has decided to panic about. Somebody other than the author put the words on the page. Everybody in the room knew. Nobody canceled anything.
The ugliest part is who pays for it. A writer with no platform gets a detection score attached to her name, and she has no way to prove a negative. Publishers who won’t write a rule are content to let a tool they don’t trust do their accusing for them, and I think that’s cowardice.
What happened in the 2026 publishing AI cancellations?
Read the cases in a row and a pattern shows up.
In almost every one, the accusation is about how the prose feels, backed by a score from a tool, and the author denies it. No manuscript history is produced. No draft trail is examined. No editor testifies about what arrived and when. A book gets pulled, a career takes damage, and the evidence never rises above a machine reading and a reader’s instinct.
There’s one case in the pile that works differently, and it’s the one that should be getting the attention.
A nonfiction author published a book containing quotes that a language model invented. The New York Times found them. He’d disclosed his use of the tools in his acknowledgments, so nobody was deceived about the method. The harm was concrete and checkable: the book said people said things they never said.
That’s a real failure with a real victim, and it needed no detection tool at all. Somebody read the book and checked a quote. I wrote about why this failure mode is the dangerous one in the survival guide for people who publish under their own name, and nothing about this year has changed my view.
Why do AI detectors fail on published books?
A detector doesn’t read for meaning. It measures how predictable a sentence is against everything the model has seen, then produces a probability. Text that scores high has low surprise: even sentence lengths, familiar transitions, vocabulary drawn from the middle of the distribution, structure that resolves the way a reader expects.
Now describe a manuscript that’s been through a professional line edit. The sentences got evened out. The odd transitions got smoothed. The word a copy editor flagged as too strange got replaced with the ordinary one. A house style got applied across three hundred pages. Every one of those moves lowers surprise. Surprise is the property the detector measures.
The tool is doing what it was built to do. It’s measuring polish, and polish is what a publisher pays for. A debut novelist who wrote every word alone and then survived a good editor will score closer to a machine than a rougher writer nobody touched. One agent quoted in the Journal said that badly written submissions have started to feel like a relief. She meant it as a joke about the flood. It’s also a description of what happens when an industry starts treating competence as evidence of fraud.
The false-positive problem is the thing itself. When the accusation can’t be disproved, the accusation becomes the punishment, and the people who take the damage are the ones who can’t afford a public fight.
Why has publishing not written a rule about AI?
Last year seventy authors, Margaret Atwood and Jonathan Franzen among them, signed an open letter asking publishers to pledge that they’d never release a book created by a machine. The pledge didn’t arrive. The Big Five have stayed quiet on a technology that’s both expensive to refuse and potentially lucrative to adopt, and an executive at one of them told the Journal the house had decided against adding specific AI language to its contracts, on the grounds that the ground keeps moving.
Think about the shape of that. The same companies are suing technology firms for training models on their catalogs. That argument says the books have value the machines took without paying. They’ll make that argument in court and decline to make it in a contract, because a contract clause has a cost and a lawsuit has an upside.
So no standard exists. What exists instead is a reflex, and the reflex fires when a story gets loud enough to hurt. A book that sells quietly with a detection score attached keeps selling. A book with the same score and a viral thread attached gets pulled. Decisions that swing on how loud the internet gets are weather, and I find it infuriating that writers’ careers now ride on them.
Why does a ghostwritten book not cause the same panic?
Somebody whose name isn’t on the cover wrote two or three books out of every ten on a nonfiction bestseller list.
Agents know. Editors know. Publishers staff for it. I’ve written more than fifty of these, and not one has caused a scandal, because the arrangement answers the only question that matters before anybody thinks to ask it.
The material is the client’s. The experience is the client’s. The judgment, the structures, the decisions made under pressure, the stories nobody else can tell: all of it comes from hours of recorded interviews with the person whose name goes on the cover. What I contribute is craft. I know how a chapter opens, how an argument builds across two hundred pages, what to cut. The author supplies the thing no writer can invent, and I supply the thing most experts never learned.
Nobody is deceived about anything they had a right to know.
That’s the whole ethical test, and I worked through it at length in the piece on whether ghostwriting is dishonest. A machine drafting the substance fails that test for one reason: there’s no author underneath. The model has no career, no failures, no Tuesday afternoon when the biggest client called to cancel. It’s the average of everything already written. Readers pick up a book by a particular person for the opposite reason.
Are literary agents now policing AI use?
Enforcement landed on the people with the least power to handle it. Agents sit between the author and the publisher, and they’re now expected to verify something no reasonable process can verify, using tools their own trade association is still writing guidance about. One agent said her inbox has never held more queries, which she puts down to writers using models to spam submissions at scale, so the volume went up at the same moment the vetting burden did.
The responses vary because nobody has told them what the rule is. Some agents ask nothing and hope the question stays away.
Some refuse to represent a writer who touches the tools at all, including for research. Some run manuscripts through detection software whose accuracy is the thing under dispute. One agent turned down a nonfiction writer whose work she liked after he said he used a model for inspiration, and she expects him to sign elsewhere without the subject ever coming up, because he already had several offers.
An industry that can’t state its rule has outsourced the rule to individual judgment, then attached career-ending consequences to getting it wrong. The Association of American Literary Agents has a committee on this, and the person chairing it says the concerns are different this month from last month. That’s an honest description of the situation and a poor foundation for anybody’s livelihood.
Agents deserve better than this, and so do the writers they turn away. Publishers who won’t put a sentence about AI in a contract have left the enforcing to the people with the least cover, and I think that’s a disgrace.
Are the cancellations a distraction from the flood of AI books?
While the industry litigates whether one novelist smoothed her sentences with a machine, the machines are producing books at a rate no cancellation touches. The Stony Brook researcher whose work triggered one of the scandals also found that a fifth of the Amazon ebooks in his data set carried heavy AI assistance. The founder of Bookshop.org puts the number of those books written to trick a buyer at nearly all of them, and describes the product as plagiarized, wrong, or too thin to be worth the file it sits in.
Barnes & Noble removes AI-generated titles from its online store when it finds them. Correct policy, and an admission that finding them is the hard part. Amazon says it removes content that breaks its guidelines. Both statements are true and neither describes a system that keeps pace with the supply.
Pulling one debut novel produces a news cycle. It removes one book from a catalog holding millions and does nothing to the flood underneath, which I put numbers to in the piece on who is counting the slop. The visible enforcement is aimed at the authors easiest to reach, and the actual damage is being done by operations with no name, no agent, and nothing to cancel.
Is publishing testing for the wrong thing?
Publishing has always run on trust instead of verification. Editors take an author’s word that the memoir is true and the research is sound. That system has failed before, in every plagiarism scandal and every fabricated memoir, and it failed the same way each time: somebody lied, and nobody had a mechanism to catch it early.
AI applied pressure to an old problem, and the industry responded by buying a machine that answers a question nobody needed answered. Whether a sentence has low surprise tells you nothing about whether the book is honest.
Three questions would do more work than every detector on the market. Are the ideas the author’s own, and can they talk about them without the manuscript in front of them? Do the facts hold when somebody checks a quote, a date, a citation? Was anybody deceived about something they had a right to know?
Those questions have answers a person can find. A five-minute call with an author about chapter seven settles more than a probability score ever will.
I think the detectors are the wrong fight entirely. AI is a tool, and a writer who treats it as one and keeps the ideas and the facts their own hasn’t cheated anybody. The question worth asking is whether the reader got a real book from a real person, and a probability score can’t answer it.
What should an author do about AI disclosure now?
Keep a record. Save the interview recordings, the dated drafts, the notes. A manuscript history is the one form of evidence that survives an accusation, and it costs nothing to keep while the work is happening. Publishers have started asking for it, and a writer who can hand it over will spend an afternoon on the question instead of a year.
Be honest in the acknowledgments about what you used and where. The nonfiction author with the invented quotes disclosed, and the disclosure is the reason that story is about accuracy instead of deception. A line saying a language model helped organize research and clean up transcripts costs an author nothing and closes a door that stays open otherwise.
Check every fact the machine touched. Every statistic, every quote, every citation, every date. A model produces a confident wrong sentence as easily as a right one, and a reader who finds one invented figure stops believing the rest of the page.
Then write your own sentences. Not for a detector, which can’t tell the difference, and not for a publisher’s policy that will be rewritten twice before your book comes out. Write them because the way you see a thing is the only asset in this that nothing else can produce, and a book assembled from the average of the internet has no reason to exist next to the tool that produced it.
A polygraph never made anybody honest. It made people afraid of the machine. Different outcome, and a worse one. Publishing can keep testing prose for how it sounds and canceling the writers whose editors did their jobs, or it can ask the three questions that have always separated an honest book from a fraudulent one. More of my thinking on this sits in the AI and writing hub, and if you want the book itself handled by somebody who will keep the record and answer for every page, that’s what my ghostwriting work is for.
Every cancellation built on a detector score punishes a writer for something nobody proved, and I think the houses doing it owe those writers an apology. Ask for the draft history, talk to the author, and check the quotes. Pulling a book on a probability score is lazy, and it’s cruel to the person whose name is on the cover.
