AI detector score sparks literary controversy as Swiss experts question the tools

A detector score helped push a bestselling novel out of France’s Goncourt race. Haitian-Canadian writer Thélyson Orélien’s C’était ça ou mourir had won critical and commercial success and was being considered for several of France’s most prestigious literary prizes. Then an anonymous account on X accused him of using AI to write almost the entire book, citing Pangram’s assessment that more than 95% of the novel was AI-generated.

Orélien denied using AI to write the novel. He acknowledged turning to it for historical research during the creative process. The book was removed from the Goncourt selection amid press revelations of plagiarism in his earlier works, for which he apologised. His Canadian publisher later cut ties with him after an external analysis found further evidence of AI use. Orélien continued to deny writing the novel with AI. The episode has left writers and critics divided over what a detector score can establish, and what consequences it should carry.

Swiss filmmaker Manuel Flurin Hendry says a software score cannot justify a writer’s public condemnation. “The idea that you feed a text into Pangram and then socially execute a writer feels unacceptable,” he told Swissinfo at the AI+X Summit in early October. Hendry, a director, screenwriter and lecturer, argues that artists should be free to choose their tools.

His concern reaches beyond the technical result: a detector’s label can affect a writer’s standing and career. Orélien’s case has split the literary world between people defending him and questioning detector reliability, and critics of the author. The dispute also involved plagiarism revelations and an external analysis commissioned by his Canadian publisher, making the consequences more complicated than a single score. Still, the Pangram result became a prominent part of the accusation. Hendry’s intervention puts the focus on how institutions and audiences interpret such assessments when they make judgments about creative work.

AI detection has become harder as language models have grown more sophisticated. Hendrik Strobelt, a senior researcher at IBM in the United States, helped develop GLTR in 2019, when GPT-2 had just been released. GLTR was an experimental tool designed to help identify text that might have been generated by a language model.

At that time, Strobelt explains, AI models tended to produce dull, stylistically similar text because they drew on common patterns. That made it relatively straightforward to look for the words a model was most likely to choose. Language models have since advanced, and Strobelt describes the question of Pangram’s reliability as complex. His account traces a shift in the conditions under which detection operates: early tools could look for predictable wording, while newer models produce more sophisticated language. That history offers context for the controversy around Orélien’s novel, where a numerical classification was treated as significant evidence in a public dispute.