Perhaps this is where the real issue lies. Not in the question – legitimate, but ultimately rather trite by now – of whether a writer can use artificial intelligence. But in this dizzying reversal of the burden of proof. On 21 September, the Twitter account ‘Balance ton Claude’ claimed that the novel had been produced almost entirely by AI, based on analyses carried out using Pangram. The post went viral. Orélien categorically denied the claim and stated that he had never used artificial intelligence to produce his texts.
In other words, a computer programme generates a probability; that probability becomes a public accusation; and it is then up to the human being to prove that they are indeed human. This is a situation that Kafka might well have found somewhat excessive.
Pangram is by no means a worthless gimmick, however. Detection tools have made considerable progress. A recent investigation by *Nature* highlights the excellent performance achieved by Pangram in several independent tests, and the company itself claims extremely low false-positive rates. But even an academic study published in 2026, which was generally favourable towards the recent performance of these detectors, explicitly recommends against using them as the sole evidence when an important decision is at stake. A detector may provide a clue. It does not, on its own, constitute a forensic assessment of literary creation.
Pangram itself acknowledges that false positives are the most serious error associated with this type of technology, precisely because wrongly accusing someone of not having written their own text can damage their reputation. This simple observation should be enough to introduce into the debate what it sorely lacked from the outset: doubt.
The story took an even more unusual turn on 24 September when the identity of the creator of ‘Balance ton Claude’ was revealed. According to reports published by *Le Figaro* and picked up notably by *Libération*, the creator is Samuel Fitoussi, a columnist at *Le Figaro*, who had himself helped to publicise the account’s accusations at a very early stage. This new development is obviously worth noting, but it does not, on its own, allow us to determine who wrote what. Above all, it serves as a reminder of a basic rule of journalism: a sensational allegation must be treated as an allegation, not instantly turned into a verdict.
For by frantically searching the texts for traces of the ‘machine’, we risk creating a strange sort of stylistic police force. A sentence that’s too regular? Suspicious. A repetition? Suspicious. A structure that’s too polished? Suspicious. A predictable vocabulary? Suspicious. And soon every writer will have to keep their drafts, their dated files, their notebooks, their successive versions – perhaps even film their screen whilst they write – in order to obtain a certificate of human authenticity.
This would be a far more profound revolution than the arrival of ChatGPT in publishing houses. For centuries, we have asked literature to move us, to surprise us, to create a language and a world. We are beginning to demand industrial-scale traceability from it.
The paradox is delightful: whilst we fear that machines are becoming too human, we are beginning to treat humans as machines that need to be authenticated.
And perhaps this is ultimately what the Thélyson Orélien affair reveals. Even before we know exactly what role artificial intelligence will play in literature tomorrow, it has already changed the way we view writers. Suspicion now sometimes precedes reading. Statistics rival judgement. The algorithm is no longer content merely to write: we ask it to tell us who is writing.
At this rate, the next major event of the literary season may no longer be the Goncourt Prize.
It will be the CAPTCHA test.
