In May, judges for the Commonwealth Short Story Prize praised the Caribbean region’s winning entry, “The Serpent in the Grove,” for its “voice of restraint and quiet authority” and prose that was “sublime” and “precise yet richly evocative.” Weeks later, an AI-detection company ran the same text through its system and reported that it was 100 percent machine-written.
A prize committee’s language, and a detector’s verdict
The story’s author, Trinidadian writer Jamir Nazir, denies using artificial intelligence and says the story came from childhood memory. He has also, oddly, defended AI as a legitimate creative tool, comparing it to typewriters and word processors. Pangram, the detection company, flagged two other regional winners in the same competition as likely AI-assisted; one denied it, the others did not respond publicly. Granta, the magazine that publishes the prize’s winning stories, kept the winners posted on its site pending an investigation. The foundation’s director said she places “complete trust in writers,” which is a fine ideal and not much of a verification process, as the ensuing coverage made clear.
What makes this case sting is the mismatch between the praise and the accusation. Judges rewarded exactly the qualities editors are trained to prize in fiction: restraint, precision, a controlled voice. The same qualities are now being read as symptoms.
A publisher’s confidence, cancelled in real time
In March, Hachette pulled the horror novel “Shy Girl” from its US release list and discontinued the UK edition, after it had already spent months on shelves. Author Mia Ballard self-published the book in February 2025; Hachette picked it up for UK publication that November. The trouble started when readers noticed odd tonal shifts, and it escalated once Max Spero, founder of the detection firm Pangram, ran the text and reported a 78 percent probability of AI generation. The New York Times conducted its own analysis using multiple detection tools and found, in its words, “recurring patterns characteristic of AI generated text, like gaps in logic, excessive use of melodramatic adjectives and an over-reliance on the rule of three.” Ballard denies writing with AI and says an editor she worked with may be responsible; she is now suing Hachette over what she calls damage to her name.
The detail worth sitting with is not the percentage. It is that a major publishing house, with editors, copyeditors and legal review, put a book through its full production pipeline without anyone catching what independent detection tools caught within days of scrutiny. Traditional publishing’s pitch to readers has always included a quality-control promise. This is the case that tested it and found it wanting.
When the disputed byline belongs to the critic, not the novelist
A different kind of accusation landed on The New York Times itself. In January, freelance critic Alex Preston published a review of Jean-Baptiste Andrea’s novel “Watching Over Her.” A reader noticed that passages closely echoed an August 2025 review of the same book in The Guardian, written by Christobel Kent. One passage in the Guardian original described side characters, including “the lazy Machiavellian Stefano” and his brother, “prone to speaking in tongues”; Preston’s version described “the lazy, Machiavellian Stefano” and a character who “speaks in tongues.” The Times investigated, and Preston admitted he had used an AI tool to help draft the piece and failed to notice the overlap before publication. The paper cut ties with him, calling the episode “a serious violation” of its journalistic standards, and appended an editor’s note linking to the original Guardian review.
This case complicates the authorship question rather than simplifying it. Nobody accused Preston of inventing a plot or a character. The AI tool produced ordinary critical prose, and the offense that mattered was closer to inadvertent plagiarism than fabrication. It shows that authorship disputes in this current wave are not only about whether a machine dreamed something up, but also about whether a writer can vouch for language they did not personally generate, verify or trace back to its source.
Disclosure is what separates a scandal from a footnote
None of this is entirely new, and the older cases show what actually determines whether AI use becomes a scandal. In January 2024, Japanese novelist Rie Kudan won the Akutagawa Prize, one of Japan’s most prestigious literary honors, for “Tokyo Sympathy Tower.” In her acceptance speech, she volunteered that roughly 5 percent of the text was drawn directly from ChatGPT, and said she intended to keep using generative AI in her work. The prize committee, which had already called the novel “practically flawless,” was unbothered; committee member Keiichiro Hirano noted that the book itself is about AI and language, so the disclosure fit the work rather than undermining it. Kudan kept the prize.
Compare that with Sports Illustrated in November 2023, when it was reported that the outlet had published product reviews under invented author names, accompanied by AI-generated headshots sold on stock-image marketplaces. There was no disclosure here, only discovery. The publisher blamed a licensed content vendor and ended the relationship; the magazine’s own union called the practice a betrayal of “everything we believe in about journalism,” and CNN’s reporting captured writers’ public disgust. Same technology, opposite reception. The variable was not whether AI touched the text. It was whether anyone admitted it before getting caught.
What editors are actually testing for
In manuscripts and articles, editors already look past grammar for uneven register, claims that do not hold up, and the kind of repetition a writer would never let stand if they reread their own draft carefully. The current dispute over AI-generated prose asks editors to apply that same scrutiny to a harder question: whether the voice on the page is one the credited writer can actually stand behind. Detection tools like Pangram give a probability, not a certainty, and the industry has no agreed threshold for what counts as decisive evidence, which is why accused writers can plausibly deny a verdict that publishers are nonetheless treating as final. The tics that keep turning up in flagged text are specific enough to be useful signals rather than vague suspicion: a reliance on the word “delve,” an excess of em dashes, three-part lists that repeat rather than build, and intensifiers reaching for import the sentence has not earned. None of that proves a machine wrote the passage. But it is exactly the kind of pattern a careful line edit is built to catch, whether the cause is AI, fatigue, or a writer imitating a style that was never their own.
The role editors are being asked to play here is not new so much as newly urgent: verifying that a text’s claims, sources and voice are what they’re presented as, before a reader or a rival publication does that verification instead.
The Commonwealth Prize judges were not wrong that “The Serpent in the Grove” read as controlled and assured. They were wrong to assume that control could only come from a practiced human hand. Editors have spent decades learning to recognize a distinctive voice as a mark of skill. The unsettling part of this year’s disputes is that the same distinctiveness is now treated as equally consistent with a well-prompted machine, and nobody in publishing has agreed yet on how to tell the two apart.
