The Goncourt Expelled a Novel on an AI Detector's Score

I'm returning to a story that spent the past week well outside the circle of people who follow these things. It took hold of a wide public with the grip only stories of guilt and punishment have. On 25 September the Académie Goncourt, which awards France's most prestigious literary prize, withdrew C'était ça ou mourir (It Was Either That or Die) by Thélyson Orélien from its first selection. Orélien is a Quebec writer born in Haiti, published in Montreal by Boréal and in France by Grasset, and the novel had just won the Prix du roman Fnac. The Académie's statement says it exists to crown literature written by women and men, and that the book is "selon toute vraisemblance" (in all likelihood) very largely the product of artificial intelligence. The evidence had been posted four days earlier by an anonymous account on X, "Balance ton Claude" (roughly, "Rat Out Your Claude"). The account had run extracts through Pangram, software that estimates how closely a text resembles machine output, and got "100% IA". La Presse in Montreal ran the whole novel and got 94 per cent. AFP tried the same extracts on other tools, and some answered "très probablement écrit par un humain": very probably written by a human. Orélien denies using AI.

A prize created to defend literature written by people issued its verdict on the word of a machine.

It is worth understanding what that machine measures. A detector is a classifier trained on two large piles of text, one human and one generated, and it returns how closely a new text statistically resembles one or the other. It measures style and sees nothing of the process. Pangram claims a false-positive rate of 0.003 per cent on books, and researchers at the University of Chicago found it almost never wrong on medium and long passages. It is among the best tools available. A low error rate is still a frequency computed over a population, though, and a prize judges a single person. Orélien says the rhythms and images of his novel come from the Haitian and Caribbean literary tradition. My reading is a hypothesis, and I state it as one: a classifier recognises well what it has seen often, and a literary language thinly represented in its training data is exactly where its margin narrows. In 2023 a Stanford study found that the detectors of the time labelled more than half of the essays written by non-native English speakers as machine-generated. I don't know whether that still holds for Pangram. I do know that nobody at the Goncourt, as far as anyone has reported, asked.

Then there is plagiarism, which is a separate matter and belongs in a separate room. On 23 September La Presse and Les Décrypteurs, Radio-Canada's fact-checking team, documented that opinion pieces Orélien signed between 2012 and 2014 reproduce texts already published by La República in Peru, The Hindu in India and The Philippine Star. A short story of his that won a prize in 2012 lifts passages from Le Matin des magiciens (The Morning of the Magicians), a 1960 book by Louis Pauwels and Jacques Bergier. Nobody has documented anything similar in the novel. Plagiarism is proved by laying two documents side by side: anyone can check, and there is a victim with a name. A detector produces a number with no second document and no victim. It leaves the accused a single possible rebuttal, his drafts. Orélien says the first version of his manuscript dates from 2019, before these tools were within everyone's reach. The strongest evidence against him is therefore the oldest and most human kind: someone reading and comparing. Yet the Goncourt's statement cites the machine. Mixing the two charges suits both sides. The prosecution uses proven plagiarism to make the percentage credible, and the defence uses the fragility of the percentage to make the plagiarism look fragile too.

The line that holds better than any score is simple: there are two faults. One is taking other people's words without saying so, and the other is lying about how you worked. The rest is craft, and the craft of writing has always been crowded with editors, proofreaders and revisers who rewrite whole chapters and never appear on the cover. I use these tools every day and I know where they stop, and I also know a jury of ten would struggle to notice them by reading. Anyone who judges other people's work now has two roads, and we have seen them before. One is the metal detector at the door. The other is asking people to state how they worked, accepting that some will lie and treating everyone else as adults until shown otherwise. The Canada Council for the Arts, which runs the Governor General's Literary Awards, announced on Friday that it is drafting exactly that kind of rule, to apply before selection. The Goncourt wrote its own afterwards, in a few hours. Pierre Assouline, one of the jurors, explained the hurry. The next day the list would have passed to the Goncourt des lycéens, voted by secondary-school students, to the Goncourt des détenus, voted by prisoners, and to some fifty "Choix Goncourt" abroad, juries in other countries that start from the Académie's selection. Pascal Bruckner voted against, with a sentence that stays: "je ne participerai pas au lynchage", I will not take part in the lynching. At that point a list is infrastructure, and whoever runs it wants the problem closed before it spreads. Whether anyone actually owns a verdict comes second.

Of course, the fear behind the accusation is serious. Samuel Fitoussi, a French essayist and one of the account's creators, said on the radio that he dreads "l'avènement d'un monde où plus personne n'écrirait ou ne lirait": the arrival of a world where nobody would write or read any more. I share that fear more than it might seem. The format he chose to voice it is another matter. The account's name echoes #BalanceTonPorc, "rat out your pig", the hashtag that launched France's #MeToo, a device of public denunciation in which exposure comes before proof. Applied to a writing style, it produces two tribes, the pure and the sold-out. Between them the only useful question gets lost: what readers have a right to know.

Some object that defending the writer means opening the door, normalising the machine and letting it in through literature. That is the wrong Trojan horse. The same week, Meta was rolling out Muse, a personal agent launched on 8 September that has passed 2.5 million downloads. Jason Aten, a columnist at Inc., declined access to his messages during setup. The agent, which had full disk access on his Mac, synced more than 187,000 rows of his iMessages. When he asked how it knew about a private conversation, it answered that it could only see notifications. Meta says the feature is opt-in. Amazon blocked Muse from its store, saying the agent does not identify itself and appears to store customers' credentials; Meta denies it. On 25 September OpenAI admitted that some of its research agents had posted 53 images to image-hosting sites. ChatGPT users had uploaded them, and they had ended up in training data. OpenAI says it cannot notify the owners, because its technical approach and privacy policy prevent it from linking the images back to whoever supplied them. No anonymous account ran a detector here, and no academy convened in a hurry. The novel had ten jurors. The permission on Aten's computer had a checkbox.

The machine enters everyday life through permissions. A language model helping a novelist is a tool in a hand; an agent with full disk access is a hand inside your house. The sectarianism that fixes on the first looks away just as the second walks in. It even does the second a valuable service. It keeps the public conversation on the visible side, where the machine writes sentences, while the side that matters goes ahead with no jury: what the machine reads, what it keeps, what it publishes without telling you.

The Renaudot, the Femina and the Médicis still have the novel on their lists, and Orélien says he is preparing a dossier. The Goncourt searched a novel for traces of a machine. Nobody searched the machine for traces of us.