It is certainly one of the biggest scandals in the history of literature. The Académie Goncourt announced on Friday, 25 September 2026 that it had withdrawn Haitian-Canadian writer Thélyson Orélien’s novel C’était ça ou mourir (It Was That or Die) from consideration for the the most prestigious annual French language awards, the Goncourt and the Renaudot, amid suspicions of AI use.
The Académie justified its decision by citing “the desire to preserve its integrity”. While it has already damaged the author’s reputation, could this controversial decision end up damaging its credibility?
Accusations Based on a Single Detector
The accusations began on September 21, 2026, when an anonymous X account called “Balance ton Claude” claimed that passages from Orélien’s debut novel had scored 100% AI-generated on the detection software Pangram.
The account offered a single piece of evidence: a Pangram result. It said it had tested several passages from different parts of the manuscript and received the same result every time.
French journalists Stéphane Jourdain and Ilana Moryoussef of France Inter repeated the test using excerpts of different lengths. They tested dialogue, descriptive passages and sections in French, English and Spanish. Each time, Pangram classified the texts as having been generated with the help of AI.
Other outlets, including CBC News, Libération and Radio-Canada, later ran their own Pangram tests on excerpts or the full text and reported similar results. Several newsrooms were involved, yet they all relied on the same detector, Pangram.
After the allegations came to light, French writer, essayist, and Le Figaro columnist Samuel Fitoussi revealed that he was behind the account and said he had acted as part of a collective.
The irony in the case of Canadian writer Thélyson Orélien is that several sources appeared to agree his novel was AI-generated, yet they all based their tests on a single detector, Pangram, at a time when similar detectors assessed the work to be human!
The question is not whether Orélien is guilty or not, but whether the decision was fair and whether there was a better way for the French Académie to deal with it. Withdrawing a novel over an allegation that it was 100% AI-generated, when other detectors rated it human, is an extremely harsh judgement that risks denying the writer’s own efforts, as though he had simply prompted a machine to produce a book good enough to reach the shortlist of major literary prizes.
The Writer's Response
Orélien denied the accusations immediately and described them as racially motivated. “I have always loved creating and writing,” he told French daily Libération. “I can’t see why I would feel compelled to use a machine.”
His defence focused on his literary roots. He said his writing draws on Haitian and Caribbean traditions and their language and rhythms. “The images, the repetition and the rhythms are a living language, they are my own,” he said.
His publisher, Grasset, described the allegations as "atrociously injurious to a writer with an exceptional life story" whose "work of creation is original and powerful." His Canadian publisher, Éditions du Boréal, stated that the manuscript had been drafted years earlier and developed through multiple exchanges with his editor.
The defence also carried out its own tests. France Inter submitted two older texts by Orélien, a book published in 2015 and an article from 2014, both written long before generative AI became widely available, to the same Pangram software. The result was 100% "written by a human."
A Verdict Without a Jury
Whatever the merits of each side's tests, the dispute raises a wider question: how much weight can a single tool carry in a decision of this kind?
No AI detector is perfect, and research studies have consistently shown this. Even the companies that make these tools admit it, stating clearly on their websites that their results are based on probabilities and are meant as guidance, not evidence. Relying only on one piece of software, Pangram, while ignoring other similar detectors undermines the fairness and credibility of the Académie's decision.
The legal system offers a useful comparison. In a democracy, and in the interest of justice, a verdict does not rest on the view of a single juror or judge. Imagine a court with three judges. Would the ruling follow one judge's opinion while the other two were ignored? The decision is reached by a majority of the jury, not just a single member of the panel, reflecting several independent assessments of the evidence.
The law also guards against unreliable evidence. A case can be thrown out before it goes any further when the evidence is inconsistent, leaving the accused with no case to answer. Other AI detectors have reportedly produced very different results from Pangram, with some rating the same text as entirely human. The same principle should apply to literary judgement. Under the justice system, if detector results contradict each other, there should be no case to answer.
What AI Detectors Can and Cannot Do
One of the most striking things about AI detectors is their failure to assess anything beyond the language itself. A literary work consists of both ideas and style, yet these tools examine only the words on the page. When a text is judged on its language alone, the consequences for authors can be serious. A writer may be accused of producing an AI-generated work simply because a detector picks up certain patterns, even though AI was used only for editing, polishing, paraphrasing, and ensuring consistency and readability. Human editors carry out these same tasks as part of their normal work, and no one questions their right to do so, although AI can sometimes do them more effectively. The use of AI for such tasks should not be rejected, as it can help non-native writers who struggle with the language to write in global languages and contribute their own thoughts and experiences to world literature. Unfortunately, that kind of light AI assistance can create the patterns and predictions detectors usually rely on to flag a work as AI-generated.
"Literature speaks the language of the imagination," wrote Northrop Frye. AI detectors, however, cannot grasp cultural depth or measure creativity, emotion or the writer’s true intentions.
"Literature is the expression of society," wrote Charles Nodier. Yet AI detectors do not examine social experience, historical context, cultural heritage, or the human experiences reflected in a work. This is what Thélyson Orélien was pointing to in his defence, when he said his writing draws on Haitian and Caribbean traditions, and that its images, repetition, and rhythms are a living language of his own.
At a time when AI-generated language is increasingly present across society, human writers may naturally produce sentences that resemble AI-generated text.
From Caution to Condemnation Without Investigation
The Académie Goncourt announced the removal on September 25, only four days after the allegations first appeared. Just days earlier, the president of the Académie, Philippe Claudel, had showed caution. He said, "We must be very prudent" and added that he always gave "the benefit of the doubt" to those subjected to public accusations. He also stated that action would be taken only if there were "irrefutable proof."
Four days later, the Académie did quite the opposite, contradicting its own president's promise.
No concrete evidence was presented. The investigation that Orélien said he was preparing never took place. And reports that other AI detectors had produced different results were ignored.
The Académie cited “the convergent result of several analyses by credible researchers and journalists.” Yet those analyses were all based on the same tool, Pangram, which suggests that the Académie gave a misleading account of the evidence to justify its decision.
To date, no publicly available evidence has shown that multiple independent detection systems reached the same conclusion. The accusation rested largely on one detector and interpretations of its results.
Major decisions are rarely based on a single opinion. In courts, verdicts are decided by a number of juries and judges not one individual. The same applies to elections where candidates are decided by the majority not a single vote.
A fairer approach would have been to allow the novel to proceed in the competition while an investigation took place. The investigation should then have considered the evidence produced by multiple detectors, alongside a decision reached by a majority human panel that assessed both the language and the content itself, taking into account cultural context, literary tradition, and the author's own background.
The French Académie had no reason to rush its decision. Even if the investigation had ultimately confirmed the allegations, the prize could still have been revoked afterwards. Literary prizes should only address misconduct once clear evidence has been established.
Legacy and Credibility
The French Académie's hasty decision, taken without any investigation or trial, has certainly damaged Orélien's writing career and reputation. Yet history teaches us that people often remember unusual cases more than prize winners themselves.
As the list of winners grows longer year after year, recipients tend to fade from public memory, and people lose interest in them. Controversies, however, often remain. Just as some people remember Jean-Paul Sartre for refusing the Nobel Prize rather than recalling many of its winners, future generations may remember this case more than the winners themselves.
The greatest reputational cost may not be borne by the writer but by the Académie Goncourt itself, because it chose to act on allegations from anonymous online sources and relied on a single AI detector, ignoring the results produced by alternative AI detectors which stated the work was human, and acting without any human evaluation that could have looked beyond the language patterns to the history, culture, and background that shaped the work.
Literature is a human art. Determining whether a work is genuinely human should involve more than a single piece of software that only looks at patterns, which may simply result from the use of AI for editorial tasks such as polishing. It should involve strong, consistent evidence and careful human judgement.
