French publishers grapple with their first AI scandal. There is no proof yet, but the author’s old texts contain sentences taken from elsewhere

On Monday 21 September, the Haitian-Canadian writer Thélyson Orélien received the Fnac Prize for the best novel, awarded by a jury of eight hundred booksellers and readers. On the same day, an anonymous account on X published an analysis claiming that his book had been “almost entirely” written by artificial intelligence. Two days later, Canadian journalists found passages in his 2012–2014 texts that matched works by other authors.

French publishers are therefore dealing with a first case of its kind, and it is not merely about whether the author cheated. It is about what happens when an accusation is made by software — and whether we can confirm or disprove it.

An account with 92 followers and the day the prize arrived

The novel “C'était ça ou mourir” (It was either this or death) was published in March by the Canadian publisher Boréal; in France, Grasset published it in August. It tells in the first person the story of Jonas Dorléon, a history and geography teacher who flees a country torn apart by gang violence and crosses twelve countries before reaching Canada. More than 35,000 copies have been sold in France, and the book made the first selections for the Goncourt, Renaudot, Femina and Médicis prizes. Le Monde ↗

The accusation came from an account called “Balance ton Claude”, which until then had 92 followers. The name is a play on the French #BalanceTonPorc movement and the name of one of the AI assistants. The account describes itself as a collective seeking to expose “promoters of AI slop”, meaning machine-produced filler. As evidence, it published results from the American Pangram detector based on excerpts from the book: the first chapter was supposedly 100 percent AI, and the second 91 percent. BFM TV ↗

Orélien denied this in the daily Libération. He says he began working on the book in autumn 2017 and completed the first version in 2019, before generative tools were publicly available. “I have always loved creating and writing; I see no reason why I would turn to a machine,” he said, adding that he was compiling a dossier with his publishers to document his authorship. Grasset dismissed the accusations as fake news, saying that AI-detection tools are themselves artificial intelligence that makes things up. France 24 ↗

A detector with an almost negligible error rate — and contradictory results

Pangram is not a fringe tool. Max Spero and Bradley Emi founded the company in Brooklyn in 2023, and its detector performs best in independent tests: in a University of Chicago study, it was the only one of four compared tools to meet the strict error-rate threshold without losing the ability to detect machine-generated text, while the research firm Epoch AI found no false positives among 495 human texts in July 2026. Marzena Karpinska of Simon Fraser University reports a false-positive rate of 0.0026 percent for French texts. Le Monde ↗

Radio-Canada therefore did what the anonymous account did not: it tested the entire novel as well as a control sample. Orélien’s book received an AI score of 94 percent, while six other novels from the Goncourt selection and nine French classics from Maupassant to Baudelaire were rated 100 percent human. That control is important — it shows that the detector does not label literature as machine-generated across the board. Radio-Canada ↗

Yet other tools produced different results. AFP had the same excerpts analyzed by InVID-WeVerify, a tool whose development it helped support, and it rated them as “very likely written by a human”. Seven other detectors produced varying results. RTS ↗

Pangram itself has acknowledged a mistake. When journalist Taylor Lorenz faced an accusation based on the tool, Spero reviewed the case and admitted that it was a false positive. Other cases remain disputed, not resolved: the detector rated Mie Ballard’s novel 78 percent machine-generated and the publisher cancelled its release, although the author argues that an external editor may have introduced AI into the text; and in the case of the accused winning stories in the Commonwealth Short Story Prize, the foundation ultimately revoked no prize. Epoch AI also found that the detector misses 10.1 percent of texts that imitate a particular author’s style. El País ↗

There is also a methodological caveat that no score can eliminate: a text written by a human and then edited with a proofreader or language assistant may receive a different score, because the detector sees only the final surface. Le Figaro ↗

A second accusation unrelated to AI

On Wednesday evening, a second wave came from Canada — and this one is factually firmer. The daily La Presse and public broadcaster Radio-Canada found passages taken from other authors in Orélien’s texts from 2012 to 2014. Le Monde ↗

La Presse made the most uncomfortable discovery in its own pages. As a freelancer, Orélien submitted a column on 28 February 2013 titled “Let us not put our heroes on pedestals”, about fallen athletes including Tiger Woods, Lance Armstrong and Oscar Pistorius. The passage is a lightly edited translation of a commentary by Andy Bautista in the Filipino daily The Philippine Star, published five days earlier. The short story “Où va le monde”, which won Radio-Canada’s audience prize in 2012, contains at least four sentences identical to passages in Louis Pauwels and Jacques Bergier’s 1960 book “Le Matin des magiciens”. Radio-Canada speaks of at least five partly copied texts; another case is a November 2014 Mediapart blog post whose paragraph literally copies a commentary by former Colombian minister Rodrigo Botero Montoya. franceinfo ↗

The way this was discovered is significant. “Our doubts about Thélyson Orélien’s use of AI led us to discover plagiarism,” wrote François Cardinal, La Presse’s vice-president for news. “At a time when the public is being flooded with fabrications, our daily diligence in verifying information is our greatest responsibility. This deception is disappointing and inexcusable.” The newspaper added a plagiarism notice to its original article. La Presse ↗

This is the heart of the paradox, and it is subtler than it may initially appear. Pangram helped Canadian journalists search for matches in old texts, but it did not label those same texts as machine-generated — they predate generative AI. A tool designed to answer whether a machine wrote a text thus led to a discovery for which the answer is no — while the authorship is nevertheless in question. The original accusation and the second one therefore do not reinforce each other: each concerns a different period and a different kind of wrongdoing.

Orélien has not yet substantively addressed the plagiarism allegations. At a festival in Manosque, he called the entire affair an attempt to silence him and said he would speak “when everything has calmed down”. Boréal declined an interview, but said in a statement that it takes the allegations “very seriously” — for the author’s sake and for the trust connecting the publisher with its readers. franceinfo ↗

Claudel: why now and why him

Philippe Claudel, chairman of the Goncourt Academy, called for caution and raised four questions that remain unanswered: who is behind the anonymous account, why now, with what aim, and how reliable are the tools used? On the question of timing, he remarked: “You have an author who is Haitian and Canadian, who is Black, who is Francophone, but a foreigner. All themes that may seem provocative at a time of heated passions.” The academy will publish its shortlist on 6 October, its four finalists on 27 October, and award the prize on 3 November. If irrefutable evidence that AI generated the book emerges by the October deadline, the jury would make a decision, Claudel said. BFM TV ↗

The writer Estelle-Sarah Bulle, who writes in French and comes from a Guadeloupean family, recalled an earlier case. In 1968, the Malian author Yambo Ouologuem won the Renaudot Prize for his novel “Le devoir de violence” as the first African ever to do so — and four years later Britain’s Times Literary Supplement accused him of taking passages from Graham Greene and André Schwarz-Bart. Greene went to court, the book disappeared from the shelves, and English editions from 1977 onward carry a publisher’s note about the borrowed passages. Ouologuem returned to Mali and lived modestly as a marabout; in France, the book was not republished until 2018. Bulle is asking something other than whether he was guilty: “Will we have to film ourselves writing with a pen to prove that we are not using AI?” Outre-mer La 1ère ↗

The authors’ reactions split in three directions. Marc Levy called the account “an account run by racists” on the radio. Doctor and novelist Baptiste Beaulieu asks: “So what am I supposed to do now? Change the way I write so that software recognizes me as human?” And Martinican Goncourt Prize winner Patrick Chamoiseau is taking the opposite approach — he considers AI wonderful and says he will write his next books with it. Virginie Despentes dismissed the detector with a vulgar insult and added that Orélien had no need to defend himself. Outre-mer La 1ère ↗

Alexandre Gefen, research director at the CNRS and co-author of a book on the cultural history of AI, points out that a publisher has no way to guarantee a text is 100 percent human — machine-generated writing can be made “human” in a matter of minutes. Robert Laffont is already preparing Sébastien Bailly’s novel for early 2027; its first version was generated by AI. According to publishing director Frédéric Martin, about two percent of it remained in the final version, and the book will carry a notice for readers. BFM TV ↗

Limits of this perspective

This article is based on French, Belgian and Canadian media outlets that are not neutral observers — La Presse and Radio-Canada are also participants in the affair, since both gave Orélien a platform or a prize in the past. Le Figaro, whose commentator spread the accusation on X, also announced on Wednesday that it had discovered who was behind the anonymous account. The article is behind a paywall; HotInfo was unable to verify it, and therefore names no one here.

One more point is crucial: we do not know whether Orélien used AI, and neither does anyone else. A detector score is not proof, no confession exists, and the 2019 manuscript has not been published or independently verified, although Boréal says the text was produced through repeated exchanges with an editor. The discovery of foreign sentences in texts from thirteen years ago is supported by comparison, but says nothing about how the 2026 novel was created — and conversely, the novel’s high score says nothing about those old texts. These are two separate matters that have merged into one in the media narrative.

What hotinfo is watching