Publishers Are Moving AI Detection From After the Scandal to Before the Deal
By Dusan Boljevic · AI/ML Engineer at TheChecker.AI
Quick answer
For most of 2026, book-publishing AI-detection stories followed the same shape: a book came out, someone ran it through a detector, and a scandal followed. Hachette pulled Mia Ballard's "Shy Girl" after an online accusation. The Académie Goncourt dropped Thélyson Orélien's novel from its prize longlist over an anonymous detector claim. Both happened after the work was already public. Industry reporting now shows a second, quieter pattern replacing it: publishers and agents checking manuscripts for AI before they sign anything. Wiley already screens every submission in-house. Penguin Random House has confirmed it uses detection as one input during editorial evaluation. And a $2 million crime-novel deal collapsed in July 2026 for the opposite reason, an agency that skipped a pre-acquisition check and later couldn't back up the author's word. The shift matters because it moves detection from a reactive weapon used after public pressure to a routine step done quietly, with more room for context and less for a single score to end a career.
The old pattern: detection after the book is already out
The two highest-profile publishing AI-detection cases of 2026 both happened post-publication. Pangram Labs scored "Shy Girl" 78.3% AI-generated after online readers had already spread the accusation, and Hachette's response came after the controversy was public, not before the book went to print (TheChecker.AI coverage). The Goncourt case followed the same order: an anonymous account ran Orélien's already-bestselling, already-longlisted novel through Pangram, posted the result, and the prize académie withdrew the book four days later, even though other detectors disagreed with the first result (TheChecker.AI coverage). In both cases, the detector result became public first and the publisher reacted to the reaction, not to its own screening.
Publishing industry journalist Jane Friedman, who has tracked AI and publishing closely since 2023, argues this sequence is why the scandals look so damaging: nobody checked anything before the deal was done, so the accusation arrives as a surprise with the book already on shelves and the publisher's reputation already attached to it (Jane Friedman, "Why the AI Detection Scandals Will End—Probably Quite Soon").
What's actually changing: real screening, not just promises
Friedman's reporting documents what institutional AI screening in publishing looks like right now, and it's more limited than the scandal headlines suggest. As of 2026, Penguin Random House is the only major publisher to have publicly confirmed it uses AI-detection software institutionally, and even there, detection functions as one input into editorial evaluation, not an automatic pass-or-fail gate (Jane Friedman, "AI and Publishing: FAQ for Writers"). Academic and professional publisher Wiley goes further: it screens all submissions using proprietary in-house tools that include AI detection, a practice Friedman describes as among the clearest in the industry. Nonfiction editors, she notes, are more likely than fiction editors to use detection at all, and even then it's typically a starting point for a conversation with the author, not a verdict delivered without one.
That's a narrow set of named adopters against the volume of manuscripts moving through the industry every year. A trade-focused writing-industry guide covering the same period found the same picture from the other side: most publishers still have no public detection policy at all, adoption is uneven across the Big Five, and independent presses range from active screening to explicit anti-AI philosophy as a publishing stance (Writers & Artists, "How AI Detection Tools Are Changing the Ghostwriting Industry in 2026"). What's changing isn't that everyone screens now. It's that the publishers who do are increasingly doing it before a contract, not after a controversy, and that's a meaningfully different use of the same tool.
The case that shows what skipping the check actually costs
A July 2026 deal collapse makes the cost of the old approach concrete. Debut author Jerry Falade's crime novel "Call Me, I'll Hide the Body" drew a 14-way auction and an offer over $2 million from Minotaur. His agents' own statement to publishers, reported by the Guardian, explains exactly what happened: "During the submission process we vetted the manuscript with Jerry for AI and felt trusting in his reassurance that it was written without utilising AI... We can no longer substantiate that, and we are withdrawing the book" (The Guardian, "$2m crime novel deal collapses amid questions over AI use").
The agency had explicitly chosen not to run the manuscript through an AI detector, relying instead on the author's own assurance. That's the exact gap pre-acquisition screening is meant to close: not proving guilt or innocence, but giving an agent or editor an actual data point to weigh alongside the author's account, instead of nothing at all. Falade denies using AI and has said the accusation reflects racial bias in publishing, pointing out that his manuscript had already passed through "first readers, second readers, editors, and acquisitions teams at numerous publishing houses" without anyone flagging it as machine-written. Whichever account of the underlying facts turns out to be right, the agency's own words show the actual failure: a choice to trust a verbal assurance over any independent check, on a deal large enough that the absence of one became the story.
What pre-acquisition screening actually looks like done right
None of the publishers doing this well are treating a detector score as a verdict. Penguin Random House's own description, per Friedman's reporting, frames detection as "one component of editorial evaluation," used alongside the editor's own read of the manuscript, not instead of it. Wiley's in-house system works the same way for nonfiction: a screening result opens a conversation with the author about sourcing and drafting process, rather than ending one. An industry framework published by the Gotham Ghostwriters AI Guidelines Working Group in April 2026 pushes the same principle into contract language, recommending specific, itemized disclosure of what AI was used for at each stage of a project rather than a blanket "no AI" clause that nobody can actually verify after the fact (Writers & Artists).
That mirrors a point TheChecker.AI has made about a different, parallel approach to the same problem. A new French literary prize for AI-assisted fiction that launched this fall lets authors declare their AI use upfront, which solves part of the trust gap but not all of it, since disclosure alone still can't verify whether the declared use matches the actual manuscript (TheChecker.AI coverage). Screening before acquisition and honest disclosure at submission are two different tools aimed at the same underlying problem: publishers currently have almost no way to confirm an author's account of how a manuscript was written until something goes wrong in public.
What this means if you're on either side of a submission
If you're an agent or acquiring editor, the practical lesson from 2026 isn't "run every manuscript through a detector and reject anything flagged." Pangram's own score on "Shy Girl" and the detector disagreement in the Goncourt case both show a single number isn't dispositive. The lesson is closer to what Wiley and Penguin Random House are actually doing: screen early, treat a flag as a reason to ask the author specific questions about drafting and editing process, and don't let the first real check happen after a deal has already gone to auction.
If you're a writer submitting work, the same logic cuts the other way. A pre-acquisition check that comes back clean is a genuine asset during submission, not just a risk to survive. Running your own manuscript through a detector before you query, and keeping your draft history and revision trail as supporting evidence, gives you something concrete to point to if a question ever comes up later, instead of relying on an agent simply taking your word for it the way Falade's did, right up until it didn't.
FAQ
Does a clean AI-detection score prove a manuscript is human-written? No. A detector estimates a probability based on statistical writing patterns; it isn't a certificate of authorship. The Goncourt and Shy Girl cases both show detectors can disagree with each other, and publishers who use detection well treat a result as one input into a human editorial decision, not a final answer.
Are most publishers currently screening manuscripts for AI before acquisition? Not yet, as of late 2026. Penguin Random House is the only major publisher that has publicly confirmed institutional use of AI detection, and Wiley is the clearest example on the academic and professional side. Reporting on the wider industry describes adoption as limited and uneven across the rest of the Big Five and independent presses.
Why did Jerry Falade's publishing deal collapse? His agents stated they had relied on his own assurance that he hadn't used AI, rather than an independent check, and later said they could no longer substantiate that account after the manuscript's authenticity was questioned during a multi-publisher auction. Falade denies using AI and disputes the handling of the situation.
What should an author do to protect themselves before submitting a manuscript? Keep your draft and revision history as a record of how the work developed, and consider running your own manuscript through a detector before you query, so you have an independent data point available if the question comes up rather than relying solely on your own account later.
You can run the same kind of check on your own writing before it goes anywhere near a submission. Our free demo scores a passage the same way editors and in-house tools now do, sentence by sentence, with the same honest caveat publishers themselves have learned this year: a score is a starting point for a conversation, never the end of one.
Dusan Boljevic
AI/ML Engineer at TheChecker.AI
Dusan Boljevic writes at TheChecker.AI, covering how AI-text detection works and how students, writers and teams can use it responsibly.
Related posts
When a Lawmaker's Quote Scores 100% AI, What Does That Actually Prove?
A 100% AI score on a politician's quote doesn't prove they didn't say it. Here's the narrower question that score actually answers.
Read more
Why UK Universities Are Dropping Turnitin Over Data, Not Accuracy
UK universities are dropping Turnitin over a data-rights EULA dispute, not AI-detection accuracy. Here's the actual mechanism.
Read more
France's New AI-Literature Prize Still Needs an AI Detector
A new French AI-literature prize lets authors declare AI use upfront. But disclosure without verification has the same gap as detection without context.
Read more