A Novel Got Pulled From France's Top Literary Prize Over One Anonymous AI Scan

Quick answer

On September 21, 2026, an anonymous X account called Balance ton Claude claimed it had fed Thélyson Orélien's bestselling novel, "C'était ça ou mourir" ("It Was Either That or Die"), into the AI-detection tool Pangram and found the book was "almost entirely" written by artificial intelligence. Four days later, the Académie Goncourt, which runs France's most prestigious literary prize, withdrew the novel from its longlist. The book had sold roughly 35,000 copies and already won the Prix du roman Fnac. Orélien denies using AI and says he started writing the book in 2017, years before generative AI was commercially available. Separately, French media outlets ran the same text through other AI detectors and got the opposite result: very probably human-written. No single detector score decided this case. What actually moved the Académie was a combination of the AI claim and a second, unrelated plagiarism allegation, and the whole episode is a clean, high-profile illustration of what happens when an unverified detector run gets treated as a finding before anyone checks the methodology, the sample, or whether a second tool agrees.

What actually happened, in order

Orélien's novel follows a Haitian teacher fleeing gang violence on a journey across a dozen countries to Canada. Published in France by Grasset in August 2026, it became one of the season's biggest literary successes: a bestseller, a prize winner, translation rights sold in more than 20 countries, and a spot on the Goncourt longlist, the shortlist for which was due October 6.

On September 21, the same day the book won the Fnac prize, the anonymous account posted its claim. It said it had run the novel through Pangram, an AI-detection company, and that the tool found the text "almost entirely" AI-generated. The Académie Goncourt's president initially urged caution, according to the Guardian's reporting. That changed within days, as a second and separate allegation surfaced: Libération reported that an early Orélien short story, which won an award in 2012, shared passages with a 1960 book by French journalist Louis Pauwels. That's a plagiarism claim, not an AI claim, and it rests on text-matching, not a detector score, but it arrived at the same moment and changed the calculus.

On September 25, the Académie Goncourt withdrew the novel, citing both the plagiarism finding and what it called "the convergent result of several analyses by trustworthy researchers and journalists" pointing to AI involvement. Orélien has denied both allegations. He told Libération that the novel's style reflects "traditional language and rhythms that were Haitian and Caribbean," not AI patterns, and has said he believes the accusation was racially motivated.

The detail that gets lost in the headlines: the detectors didn't agree

This is the part worth sitting with if you actually care about what a detector score means rather than just who it was pointed at. The Guardian's own reporting on the story notes, in a single sentence easy to miss: "French media outlets have found other AI detection programmes that identified It Was That or Die as very probably written by a human." One anonymous account ran one tool once and announced a verdict. Other people ran other tools on the same text and got the opposite answer.

A French AI commentator, Benoît Raphaël, went further and tried to pressure-test the claim directly rather than just repeat it. He ran several of Orélien's own pre-2023 texts, written before commercial generative AI existed, through the same detector. He also tested excerpts from other published Haitian novels. According to his own published test results, those pre-AI texts and the comparison novels came back as human-written, while Orélien's new novel scored differently. That's a more careful test than the original anonymous claim, but it still isn't the same thing as an independently replicated, methodologically disclosed finding, and it illustrates exactly why one score from one unnamed source shouldn't have been enough to move a national literary institution in the first place.

We've written before about why a single AI-detection score should never be the thing that decides a case on its own, using a very different setting: a US college lawsuit where a student failed a class over one flagged score with no follow-up conversation. The mechanism here is the same, just moved from a classroom to a literary prize. A score is a starting point for a human review, not a conclusion a committee or a prize jury can cite as settled fact, especially when it arrived from an anonymous account with no disclosed sample size, no disclosed prompt or test conditions, and no public methodology at all.

Why style alone is a weak signal here, and why that's not new

Orélien's defense centers on something researchers studying detector bias have flagged for years: his prose draws on Haitian Creole-influenced French rhythm, repetition, and oral storytelling traditions. Detectors trained predominantly on standard, formal written English or French can misread unusual but entirely human stylistic choices, including repetition and rhetorical patterning common in oral-tradition-influenced writing, as the kind of repetitive patterning associated with machine output. That's a documented failure mode, not a new excuse invented for this case. We've covered the research on exactly this kind of false positive against non-standard or non-native writing styles before, and it applies here regardless of which specific tool was used or what the eventual truth about this manuscript turns out to be.

None of this proves Orélien didn't use AI. It proves that an accusation resting on one undisclosed detector run, contradicted by other tools on the same text, isn't proof either. Both things can be true at once, and the honest position is that the manuscript's actual authorship remains genuinely disputed, not settled by a screenshot of a percentage.

What this means if you're a writer, editor, or publisher

If an unverified social-media claim can get a bestselling, prize-winning novel pulled from consideration days before a shortlist announcement, the lesson for anyone submitting work for publication, review, or a prize isn't "detectors are useless." It's that the way a score gets produced and disclosed matters as much as the number itself.

1. A detector score without a visible methodology is not evidence. No sample size, no disclosed test conditions, no statement of the tool's own error rate on this type of prose. Treat an unverified screenshot the same way you'd treat an unsigned anonymous tip: worth checking, not worth acting on alone.
2. Run your own check before someone else runs one on you. If your manuscript, article, or submission is heading somewhere that cares about this, know what a transparent, sentence-level breakdown actually shows before an anonymous account decides to publish its own version of that answer first. Our demo gives you that view instead of a single bare number.
3. Different detectors can and do disagree on the same text, as they did here. A single tool's output, especially one you can't independently re-run under the same conditions, is not a consensus finding.
4. Stylistic unfamiliarity isn't proof of machine authorship. Writing shaped by oral tradition, translation, non-native structure, or genre convention can trip pattern-based detection without a single word of it being AI-generated.

FAQ

Was Pangram's detector wrong about Orélien's novel? That's genuinely unresolved. The Académie Goncourt cited multiple analyses pointing toward AI involvement, but other detectors run on the same text by French media and independent commentators returned the opposite finding, and the original claim came from an anonymous account with no disclosed methodology. The accurate statement is that the result is contested, not that any single detector has been proven right or wrong.

Is the plagiarism allegation related to the AI allegation? No. They're separate claims resting on different evidence. The plagiarism finding involves specific overlapping passages with a 1960 book, identified through text comparison. The AI claim rests on a detector's pattern analysis. Both arrived around the same time and both factored into the Académie's decision, but one doesn't prove the other.

Can style associated with a specific cultural or linguistic tradition get flagged as AI-generated? Yes, this is a documented limitation. Detectors trained mostly on standard formal prose can misread repetition, rhythm, and oral-storytelling patterns common in some literary traditions as signs of machine generation. It's one reason a raw score should never stand alone as proof.

Check your own writing before someone else's unverified claim does

If you're submitting work anywhere that a detector result could follow you, whether that's a publisher, a journal, or an employer, run it through TheChecker.AI's demo first. You'll get a transparent, sentence-level view instead of a single disputed percentage with no methodology attached to it.