Mia Ballard’s horror novel did not lose a trial. It disappeared in an afternoon.
As Alexandra Alter reported in the New York Times on March 19, 2026, Hachette Book Group’s Orbit imprint decided not to publish Shy Girl in the United States, where it had been due that spring, and to discontinue the British edition. The Times had approached Hachette citing evidence that the novel appeared to be written with AI. The decision was announced the next day. That afternoon the book was gone from Amazon and from Hachette’s site.
Hachette described the review behind the pull as thorough and lengthy. That is the publisher’s account. What anyone outside the building can check is narrower. A forthcoming book left the market before any public finding about who wrote it.
Ballard denied using AI to write the book. She blamed an acquaintance she had hired to edit a previously self-published version, and she said she was pursuing legal action, which is why she would not walk through those edits. She described the damage as landing on her name and her mental health for something she said she had not done herself. That is what she told the Times. It has not been adjudicated in public.
This was not a blockbuster recall. The Times, citing NielsenIQ BookData, put UK print sales at 1, 800 copies after a fall 2025 publication. The scare was the speed, and what it implied for every other manuscript in a season of launches.
Two fights, running in the same month
Publishing is stuck between two AI problems that do not point the same way. Writers have spent years challenging the unauthorized copying of books for model development. Now writers are being asked to prove they did not use a model themselves.
Andrea Bartz sits in both. She is the first-named plaintiff in Bartz v. Anthropic, alongside Charles Graeber and Kirk Wallace Johnson, and the author of the thriller We Were Never Here. She told WIRED what the new climate felt like from the author’s side.
I kind of immediately felt like this was going to lead to a potential witch hunt.
She also said there is no agreed-upon standard for proving a person wrote the work, and that the practical effect would be simple. Anyone could accuse anyone. That stops being a philosophical quibble when a publisher can take a book off sale in a day.
The climate WIRED described is already commercial. Since Shy Girl, the magazine reported, other fiction writers have faced similar suspicions. One drew suspicion after winning a short story prize. Another drew it after publishers bid into the millions for a crime novel. WIRED did not name them. Both have denied the allegations. Bartz has heard of Goodreads review-bombing after a stranger decided a book sounded like AI, and she has seen readers say they will not read anything published after 2022 so they can be sure it contains no generative AI. Those are things she has encountered, not a counted boycott. They are still enough to make a publisher flinch.
The same month had a deadline in the other fight. The Authors Guild’s settlement explainer lists March 30, 2026 as the claim deadline in the Anthropic case. Authors were still inside a major copying dispute while a separate allegation of AI writing had already cost one novelist a Hachette publication.
The $1.5 billion figure is real. The story usually told about it does not match the docket. On the Authors Guild’s account, the settlement pays for a piracy class, not for a court finding that training on books is infringement. The Guild says Anthropic downloaded 7 million copies from the pirate libraries LibGen and PiLiMi, and that about 500, 000 titles qualified once duplicates and ineligible works were stripped out.
The June 2025 order and the later class are different events. Judge William Alsup held that training on legally acquired books was fair use. The Guild says that ruling applied only to the three named plaintiffs, not to the class, and that he left the piracy claims standing. In July 2025 the court certified a class built around those pirated copies. The settlement resolved that piracy class. It did not decide that training itself is theft. Anyone still collapsing those three steps into one moral is selling a cleaner story than the docket supports.
Registration rules did most of the shrinking from 7 million downloaded copies to about 500, 000 titles. Paperwork, not rhetoric, decided who could make a claim. The dates and the per-title math are below. They matter less than the distinction. Pirated inputs are one risk. A suspicious manuscript is another.
What the receipt actually records
Into that gap, a small industry is selling what gets called a paper trail. Log the writing. Fingerprint it. Store a declaration that a human did the work. If a stranger, a prize judge, or a publisher later cries AI, the author points at the receipt.
A handful of startups are selling some version of that receipt. The two worth understanding, because their own makers have described the limits, are Pangram and ObraVera.
Brady Ridgway, technical founder of ObraVera, was unusually plain about his. He told WIRED the platform can cryptographically fingerprint patterns in writing and in voice recordings, store them for a registered user, and use a time stamp and a crypto hash to say these words were held under this name at this hour. If the text later appears where it should not, or someone else claims it, the database can answer with a when and a who. The registered person must also declare whether AI was used. The software cannot check that declaration.
In the same interview he pointed to what he called a kerfuffle in May (the May before WIRED’s November 9, 2026 story) over a writer accused of winning a Commonwealth Short Story Prize with an AI-generated submission. Everybody, he said, seemed to agree the accused person was the one who submitted the story. The software would only have recorded that, on a given day, somebody declared the work human-only or not AI-assisted. Compliance teams already know this object. It is an attestation. A hash can prove a file existed under an account. It cannot prove a model did not draft the sentences an hour earlier.
Max Spero, cofounder and CEO of Pangram, sells detection and a product that tracks writing in Google Docs. He told WIRED that many users find the tracker helpful, and that paper-trail tools are still not a catch-all. Someone can hand-type AI text, or run an auto-typer, and the trail looks human. The average writer does not sit down and type out a book in one file. Work moves across documents. Drafts go back and forth. Voice notes get taken on the go. Not all of that is captured.
Mary Rasenberger, CEO of the Authors Guild, does not want any of this turned into a requirement. She told WIRED that asking authors to keep receipts is insulting and a burden, and that the guild does not feel it should be required. The guild will not endorse one tool. Umair Kazi, its director of policy and advocacy, told WIRED that many of the companies have already come pitching. Rasenberger still wants writers to know the tools exist, as a security net they can choose.
Her sharper line is about incentives. If what you are selling is your words, your voice, your imagination, your creativity, using AI to supply those things does not make sense. She has also said pride in craft makes low-quality AI output less common in publishing than the panic suggests. That is her assessment, not a census. Editors she has quoted are more confident still. They never see AI-generated submissions, and they would know one if they did. Shy Girl is a rough week for that kind of certainty.
The detector is a witness, not a judge
Pangram is the name most often attached to the Shy Girl numbers, and the attachment is thinner than it looks. WIRED reported that Pangram estimated some 78 percent of the novel’s text was AI-generated, and called the detector robust but by no means infallible. Alter’s Times account of the cancellation did not name Pangram and did not publish a percentage. Treat 78 percent as the company’s finding, relayed later, not as an independent measurement of the book.
Pangram also claims an industry-leading false-positive rate of 0.01 percent on creative writing. A false positive, here, means human writing labeled as machine writing. Rasenberger told WIRED the Authors Guild has heard of lots of instances in which people submitted their own work and were told it was AI-generated. That is the guild’s account of what it has heard. It is not a published error study. Both figures can sit on the table. Neither closes the case.
Jane Friedman, who writes the industry newsletter The Bottom Line and whom Digital Book World named Publishing Commentator of the Year in 2023, is the skeptic worth hearing on adoption. WIRED reported that she had argued in the newsletter the scandal era is likely to be short: detectors such as Pangram are improving, and authors have seen what happens when AI-generated writing is passed off as their own. Speaking to WIRED, she was colder about a shared fix. Getting agents and publishers onto one verification standard, she said, will be like herding cats. Publishing is slow to adopt new technologies, especially technologies built by outsiders.
If any of this costs extra money or extra time, forget it.
She also flagged the obvious next fraud. A certificate can be faked or manipulated.
What to copy, and what not to
The publisher practice actually described in the Times account is lighter than a crypto receipt, and more plausible for that reason. Hachette told the Times it requires submissions to be original to the author and asks authors to disclose AI use during writing. That is a contract clause and a question. It is not a process log. It is also one publisher’s stated practice in one controversy, not an industry rule.
Taken at her word, Ballard’s account is the pattern that clause cannot see. The person who signs the disclosure may not be the person who introduced the disputed sentences. A distributed edit breaks a personal pledge. Most company writing is already a distributed edit. The clause never answered the question the cancellation turned on: who put those sentences in the file, and how anyone would prove it.
Bartz’s next book, as she described it to WIRED, is a stress test the detectors are not built for. Delete Your Account, which WIRED reported is due in 2027, is about generative AI. Its blocked protagonist asks a chatbot to imitate her favorite thriller and then finds herself trapped in that story, facing death. Bartz said she wrote portions meant to sound like large language models, including invented transcripts and marketing copy that mimics machine prose. She does not know what she would do if a detector flagged them.
Satire, pastiche, and a lot of brand voice work by intentional imitation. A company that trains a house style on its own old copy is walking into the same trap with better fonts.
Carry that out of the book business and the liability changes shape. The pitch for these tools inside a company is usually speed: faster drafts, a thinner staff on the content line. The quieter problem is an authorship story you cannot reconstruct when a customer, a regulator, or a journalist asks who wrote the thing. Procurement people do not need a 0.01 percent error rate to walk. They need a feeling. When proof is expensive, refusal is cheap.
A few operating conclusions follow from what is known, not from the brochure.
Disclosure will spread faster than fingerprints. Friedman is probably right that a shared technical standard dies the moment it adds cost or time. A sentence in a contract is cheaper than a platform, which is why Hachette’s ask is the response most likely to be copied. It is also a weak control once a freelancer, a vendor, or an acquaintance hired to edit can touch the file after the pledge.
Do not outsource a cancellation to a detector score. Shy Girl shows what an allegation can do before anyone has adjudicated it. The false-positive stories Rasenberger relayed to WIRED show the other error. Pulling a book before anyone outside the company can inspect the evidence protects the imprint and can wreck a person. Waiting can look like negligence. There is no clean third option. There is a decision you should be able to explain without hiding behind a vendor’s percentage. Rasenberger’s advice to writers applies in the office too. An accusation you cannot support does not make a healthier market.
Keep the training fight and the authorship fight in separate folders. On the Guild’s account, the Anthropic settlement is a warning about pirated inputs and about copyright formalities. It is not a ruling that learning from lawfully acquired books is theft. Suspicion about a finished manuscript is a different risk, aimed at brand, retailer, prize, and employer. Mixing the two produces bad policy and worse press releases.
Decide, in advance, what “human” means in your shop. Editing, research, translation, a model pass on a headline, a chatbot used as a thesaurus: the tools on offer mostly record whatever the user is willing to declare. If you have not defined the line, a receipt cannot defend it.
Questions worth settling before the next cancellation
Did the $1.5 billion Anthropic settlement make it illegal to train on books?
No. The settlement did not do that. According to the Authors Guild, the money resolves a piracy class covering copies downloaded from LibGen and PiLiMi, not the act of training. In June 2025, Judge William Alsup held that training on legally acquired books was fair use. The Guild says that holding applied only to the three named plaintiffs, not to the class, that he left the piracy claims standing, and that he set a piracy trial for December 1, 2025. Mediation produced a settlement before that trial. The July 2025 class, and the settlement that followed, are a third thing again: pirated acquisition, plus eligibility paperwork. They are not a holding that training on lawfully acquired books is infringement.
What would an eligible author actually receive, and as of when?
On the Authors Guild’s timeline, Judge Alsup preliminarily approved the settlement on September 25, 2025. Claims were due March 30, 2026. Final approval came on July 20, 2026. In late October 2026 the Guild said initial payments would go out the week of November 23, which was still ahead when WIRED published on November 9. A 2025 agreement was not a closed, paid case.
The Authors Guild states two figures a calculator will set against each other. Fees and expenses come off the top of the $1.5 billion. Each included title was expected to receive an equal share of at least $3, 000, possibly more depending on valid claims and interest. Divide $1.5 billion by about 500, 000 titles and you get $3, 000 before any deduction. The Guild explainer does not publish a fee total that would show how “at least $3, 000” survives that deduction. Treat the number as the Guild’s stated expectation for an included title, not as a sum you can verify net of counsel, and not as $1.5 billion handed to authors. Trade and university-press books then default to a 50/50 split between the author side and the publisher. Sole rightsholders, including many self-published authors and authors whose rights have reverted, can claim the full per-title amount. The July 2025 class required an ISBN or ASIN and U.S. Copyright Office registration within five years of publication, and either before the download or within three months of publication. That is why millions of copied books never entered the class.
Can a paper trail prove a person wrote every sentence?
No. As Ridgway described ObraVera to WIRED, the record is a time, an account, a stored version, and a declaration the user chose to make. Spero’s limit is practical: model text can be typed in by hand or by an auto-typer, and real drafts, files, and voice notes often never enter the log.
Are detectors solid enough to police a catalog?
Not on the public evidence. The 78 percent figure on Shy Girl and the 0.01 percent false-positive claim are Pangram’s numbers. WIRED relayed the first. The Times account of the cancellation did not. The Authors Guild told WIRED it has heard of human work flagged as AI. That is anecdote, not an error study.
Will publishers require cryptographic receipts?
Unlikely, if Friedman is right that anything costing extra money or time will be ignored. The practice the Times actually documented is Hachette’s statement that submissions must be original and that authors are asked to disclose AI use. That clause was never built to stop a publisher from pulling a book. It also never answered who introduced the disputed sentences, or how anyone would prove it.
What should a company that publishes words do now?
Separate pirated-input liability from trust in a finished text. Define what assistance you allow before a customer or a journalist asks. If you keep a record, record the human decision, and do not let a detector score become an accusation you cannot explain in your own words.
Bartz warned that refusing every book after 2022 is a fine way to tank an industry at once. The accusation is already cheaper than the proof, and a receipt still only shows that a file was stored, not that a mind made it.