France's Goncourt dropped a likely-AI novel once plagiarism reports gave it a second reason

According to Morning Brew, the account fed random passages of the novel into Pangram, and the tool scored the selected passages as 100% AI-generated ^[5]. The claim the account made public covered the whole book. AI had written it "almost entirely" ^[4]. Le Figaro identified one of its own columnists, the essayist Samuel Fitoussi, as the person behind the account. Fitoussi later said he ran it with other people, Reuters reported ^[17].

That first user ran one tool on a sample and published a verdict on a novel. The academy described a more careful process, citing "the convergent result of several analyses by trustworthy researchers and journalists" ^[3]. Detector results did not all converge. Academics running similar programs found high percentages of AI-generated text, Morning Brew reported ^[20]. French media found other detectors that rated the novel as very probably human-written, according to the Guardian ^[21].

Pangram's founder, Max Spero, claims the program has a false-positive rate of 1 in 24,000 ^[19]. Detection programs have been criticized for casting too wide a net and for being trained primarily on European text, Morning Brew noted ^[9]. Orélien's defence addresses that criticism directly: he says the novel's rhythms come from Haitian and Caribbean language traditions ^[13]. "The imagery, repetition and rhythm are a living language, they are mine," he told Libération ^[8].

The reporting does not show a prize removing a book on a likely-AI finding alone. The academy's statement names two grounds, the AI analyses and plagiarism the author had committed over time ^[2]. In the Guardian's account, Claudel's call for caution came while the case rested on detector scores, and the academy's view shifted after the plagiarism reports ^[18]. Those reports are evidence a person can check without a model. The 2012 story can be set beside the Louis Pauwels preface it is said to copy ^[6]. Orélien has not answered the plagiarism claims ^[14].

The academy stated its rule in absolute terms. "We neither endorse nor give our blessing in any way to the creation of texts generated with the help of AI," it said, in TNW's translation ^[10]. It addressed that message to its student and prisoner juries and to more than 50 Choix Goncourt juries abroad ^[11].

For a team whose product generates or certifies text, the case sorts onto a simple grid. One axis is the AI signal, running from a single tool's score to several independent analyses that agree. The other is whether a second finding exists that a person can verify without a model, such as copied passages. A single score with nothing beside it is the position from which Claudel urged caution ^[18]. Agreeing scores with nothing beside them is the box this case never tested on its own, and here other detectors disputed the scores anyway ^[21]. A score of either kind next to a documented copy is where the Goncourt acted ^[2]. I would set a removal line at that last row and write it into policy before a case arrives. The tradeoff is that machine-written work by an author with a clean record stays listed, since a probability is all a reviewer will have.