// Contents

“If you want, I’ll invent a cliché for you.” That was ChatGPT’s reply when the French novelist Philippe Jaenada asked it for a worn-out expression. Jaenada laughed, said yes, and the machine’s made-up cliché went into his novel L’Inconnue du quai de Javel (The Unknown Woman of Quai de Javel). On November 3, 2026, that novel won the Prix Goncourt, France’s most prestigious literary award.

Jaenada himself let slip that the line was there, and he dared readers to find it. Seventy answers came in. Every one was wrong. “These are sentences I wrote myself, which stings a little,” he said on November 6, speaking live on RTL radio from the book fair in Brive. The machine didn’t hide because it wrote like a genius. It hid because it had been asked for the most ordinary thing language has. Six weeks earlier, a different novel had been pulled from the Goncourt race after an AI-text detector flagged it. Put the two stories side by side and a pattern appears: people struggle to spot the machine, while a software score gets believed too quickly.

Half a sentence, hidden on purpose

Jaenada, 62, doesn’t even own a cell phone. He spends years in archives reconstructing old unsolved murders, and in this book he hunts for whoever killed Louise Cansot, a 29-year-old model found on the bank of the Seine in September 1949. Of all French novelists, it was this paper-and-archives man who slipped a ChatGPT line into his book.

Philippe Jaenada at the Le Livre sur la Place book festival in Nancy, 2024
Philippe Jaenada at the Le Livre sur la Place book festival in Nancy, 2024Photo: Gérald Garitan / Wikimedia Commons · CC BY-SA 4.0

The morning after the prize, he went on France Inter radio and admitted he had turned to ChatGPT three times. Once, the chatbot genuinely helped him dig up technical information. Another time he set out to mock it: on one technical detail it gave what he called “completely absurd” answers. The third time was the hidden line, planted “to play with the reader,” Euronews reported. Only the nine jurors knew where it was. He had told them on Tuesday, the day of the award, and joked that if the secret got out, he would know who had talked.

Three days later he gave it away himself. In the book, he studies a photograph in which he thinks he recognizes Louise Cansot, and writes “Croire, ça ne mange pas de pain.” The French saying literally means “believing doesn’t eat any bread”; roughly, there’s no harm in believing, since it costs you nothing. Jaenada usually runs from clichés, but here he wanted to follow one stock phrase with another meaning “but it gets you nowhere.” Nothing came to him, so he asked ChatGPT. The first suggestions didn’t work. Then the bot offered to make a cliché up.

“You don’t invent a cliché!” the writer protested on air. He took the offer anyway, “sniggering.” The machine supplied the ending, also in French: “mais ça n’en met pas non plus sur la table.” The whole sentence in the book reads like an old country proverb: “Believing doesn’t eat any bread, but it doesn’t put any on the table either.” The first half is human. The second half is the chatbot’s, though to the ear both sound equally ancient.

Seventy readers went looking for the machine and kept finding the author.

Why nobody found it

In hindsight it makes sense. A cliché is a phrase thousands of people have written before you. A language model learns from exactly that: text written by thousands of people. Ask it to invent a new cliché and it does what it does best, assembling the most likely continuation out of familiar parts. The result sounds as if it has been repeated in village cafés for a century, even though software came up with it.

People hunting for machine text tend to look for something else: a line that is too smooth, too correct, slightly off in tone. Jaenada’s readers flagged seventy suspicious passages, and he had written all seventy. This was no scientific test. Nobody knows who those readers were or how many pages they read. Still, the lesson is hard to miss. In the work of a writer with a distinctive voice, half a machine-made sentence vanishes into the prose. It vanishes faster still when the machine is imitating the most worn-out phrasing there is.

Half the loaf is bread, half is paper: one half of the sentence is human, the other is the machine’s
Half the loaf is bread, half is paper: one half of the sentence is human, the other is the machine’s

The book a program caught

At the other end of this story stands Thélyson Orélien, a Haitian-Canadian novelist based in Quebec. His novel C’était ça ou mourir (It Was That or Die) topped the French bestseller list, sold around 35,000 copies, and won the prize run by the Fnac bookstore chain. Around the same time, an anonymous X account called Balance ton Claude (“Rat out your Claude”) ran passages through the detector Pangram and announced that the book was almost entirely machine-written, Dexerto reported. Things moved fast after that.

// Timeline

  1. The Balance ton Claude account posts Pangram results on Orélien’s novel.
  2. The Académie Goncourt drops the book from its first list of 16 novels, citing analyses it considered credible and borrowings in the author’s earlier writing.
  3. Orélien admits to old plagiarism but denies that a machine wrote the novel.
  4. His publisher, Boréal, cuts ties with him and refuses to reprint the book.
  5. On the day of the award, Académie president Philippe Claudel promises to write into the rules, before the year is out, that the prize goes to a book written by a human being.

The detector’s percentages were not what settled it. Boréal gathered every working draft of the manuscript and ordered an outside review. It found “traces of AI tags in the metadata of certain documents,” meaning the hidden information software writes into a file alongside the text. The publisher admitted it does not know how much AI was used and called the real issue one of trust between authors and publishers. Orélien stands firm: “The book was written by me.” He says AI only helped with historical research and with translating from other languages.

What a detector actually counts

Press coverage of the case gives no percentage for Orélien’s book. It repeats only the account’s conclusion that the novel was almost entirely machine-written, and the Académie has cited no figure. Even when Pangram does show a number, it is easy to misread. When it says “99%,” that is the share of words it flagged, not how sure it is of the verdict, Selectra points out. Confidence is reported separately, passage by passage: high, medium, or low. The detector itself, trained on roughly a million human- and machine-written documents, labels each chunk as human, AI-assisted, or AI-generated, the company explains.

The company says false positives are close to zero, which is not the same as zero. Pangram itself lists what can set off an alarm: text polished by the Grammarly proofreader, a translation tool, the “Help me write” button in Google Docs, or a single sentence pasted from a ChatGPT conversation. Jaenada’s insert is exactly that case. But across a whole novel its share is tiny, and share of words is what a detector measures. Going by Pangram’s own description of the method, the software would most likely have missed Jaenada, just as his seventy readers did.

With Orélien it went the other way: the software was believed before the evidence arrived. His book was dropped from the country’s top prize over a detector alarm and his old borrowings, before anyone examined the manuscript files. The review later backed up the suspicion, but the order was backwards.

A detector counts suspicious words, not the author’s intent
A detector counts suspicious words, not the author’s intent

A rule about humans

The Académie Goncourt has drawn its own conclusions. “We will change the rules and adopt them before the end of the year, to make it clear once more that the Prix Goncourt rewards a work written by a human being,” its president, Philippe Claudel, said on November 3, Radio-Canada reported. Antoine Gallimard, head of the publishing house Gallimard, thinks an author who used AI should simply say so. His company, according to BFMTV, has decided to run manuscripts through an AI detector before publication.

Three French entrepreneurs took the opposite approach. In October they launched the Prix Pascal for books written together with AI. Along with the manuscript, entrants submit a technical dossier saying which model helped and how. The winner will be announced in mid-December.

Jaenada himself describes his use of AI as “ironic” and very limited. Pretending it doesn’t exist is pointless, he says; you have to “tame it, not let yourself be replaced.” His novel contains half a sentence of machine text, and he disclosed it himself, before any detector could. Will a book like that still count as “written by a human being” once the Académie finishes its rules? The detectors, meanwhile, have earned a line from the same novel: “Believing doesn’t eat any bread, but it doesn’t put any on the table either.”