Goncourt drops the prize novel accused of being AI-written
France's literary season just produced its first casualty of AI, and nobody involved can actually prove authorship either way.
The Académie Goncourt withdrew Thélyson Orélien's novel C'était ça ou mourir from its selection on Friday 25 September, four days after the book won the Prix du Roman Fnac. The academy said the decision was driven by "the plagiarism of which its author has been guilty over time" and by "the convergent result of several analyses by credible researchers and journalists" concluding the novel is "in all likelihood very largely the product of an artificial intelligence." It framed the move as sending "a clear message" that it does not endorse text generated with AI's help. Orélien was on the first selection for the Goncourt, whose winner is announced on 4 November; the academy, chaired by Philippe Claudel, publishes its second list on 6 October. The book remains in the running for the Femina and Médicis prizes.
The accusation arrived on 21 September, the day the Fnac prize landed, from an anonymous X account called Balance ton Claude — "Denounce your Claude." Created in June 2026 and with roughly 92 followers at the time, the account said it had tested passages with an AI detector called Pangram and found "100 percent AI." It cited a false-positive rate of about one in 24,400; Pangram's own site currently states one in 10,000. CBC News repeated the test on different excerpts and reported the same result, and ActuaLitté's round-up of the coverage records France 24 running the full 63,425-word novel through Pangram, which flagged 92% of it as AI-written or assisted, plus the French firm Lucide scoring all 35 units of the 59,488-word Quebec digital edition as machine-written.
Detectors cut both ways, which is the part that should worry everyone. As ActuaLitté records, the same round of testing sent two older Orélien texts from 2014 and 2015 through Pangram and both scored 100% human — evidence that at least the tool can clear him when the text is genuinely his. Four other 2026 novels came back 99–100% human, while franceinfo's GPTZero run on the same excerpts returned 87% where Pangram returned 100%. No detector has seen the manuscript or the drafts, and the accusers' probability argument — multiplying a one-in-24,400 error rate across 20 passages — treats one author's paragraphs as independent trials, which they are not.
The plagiarism findings are the harder part, because they don't depend on a model's opinion. La Presse reported that Orélien's 2012 Radio-Canada prize-winning short story shares at least four sentences with the 1960 book Le matin des magiciens, and that a 2013 column he wrote for the paper was a lightly reworked translation of a Philippine Star column published five days earlier, with further 2012–2014 cases surfacing in Médiapart and HuffPost Québec texts. Radio-Canada's verification unit confirmed the findings, and La Presse appended a plagiarism note to its own column. ActuaLitté separately found a stray imperative sentence — "Discuss the role of Quebec as guardian of French linguistic heritage in America" — in a September 2023 article on a site Orélien ran, the kind of line that reads as a forgotten writing prompt; the page is now offline and the outlet cautioned that on its own it does not prove AI authorship.
Orélien denies using AI throughout. "I have always loved creation and literature. I can't see why I would feel compelled to use a machine," he told Libération, adding that the images and repetitions in the book "are the living word, they are my own," and that his first draft was written in 2019, before text generators became widely available. He has described a hate campaign against him and told La Presse: "Maybe I have to stop writing." His French publisher Grasset called the claims "atrociously injurious," said it had absolute confidence in him, and dismissed the detector evidence as unreliable; his Quebec publisher Boréal suspended his promotional tour on 24 September, saying it was impossible at this stage to confirm or deny the allegations but that it took them very seriously. The Salon du livre de l'Estrie dropped him as guest of honour and the retailer Renaud-Bray removed the book's staff-pick label. Around 35,000 copies had sold, with rights in roughly 20 languages.
Strip the personalities out and the precedent is the story: a national prize withdrew a novel because detector output, plus a documented plagiarism history, together satisfied its gatekeepers. That is a decision made on tools whose error rates no one can pin down, in a case where the same tool cleared the author's earlier work. It also mirrors an earlier episode this year, when the Commonwealth Short Story Prize was withdrawn over AI allegations and then re-awarded once the organisers were satisfied the writing was genuine. French publishing now has a de facto standard — plausible detection plus any evidence of past copying — and no agreed way to contest it.
What to watch: the second Goncourt list on 6 October, whether Femina or Médicis follow the academy out, and whether Orélien releases the dossier he says he is compiling with his publisher.
Should a literary prize be able to strip a book on detector output alone? Tell us in the comments.
Sources: BBC News · Agence France-Presse via La Presse · CBC News · La Presse — a plagiarist's past · ActuaLitté — the leftover prompt · ActuaLitté — the detector results · Nice-Matin — the case summarised
