In the space of a few days in September 2026, Haitian-Canadian writer Thélyson Orélien went from literary sensation of the French rentrée to the center of a storm that mixes questions of authorship, technology, race, and trust in detection tools. His debut novel, C’était ça ou mourir (“It Was Either That or Die”), had just won the Prix du Roman Fnac and sat on the longlists for the Goncourt, Renaudot, Femina, and Médicis. Then an anonymous X account claimed an AI detector had scored large portions of the book as 100 percent machine-generated.
Orélien immediately and forcefully denied it. The resulting uproar has exposed deep anxieties about generative AI in literature and the limits of the tools used to police it.
The Book and the Breakout
Orélien was born in 1988 in Gonaïves, Haiti. He left after the 2010 earthquake, settled in Montreal, later moved to Ottawa, and worked as a journalist and poet before turning to fiction. C’était ça ou mourir, published in Quebec by Éditions du Boréal in March 2026 and in France by Grasset in August, follows Jonas Dorléon, a young Haitian history-and-geography teacher forced to flee gang violence in Port-au-Prince. The narrative traces a perilous northward journey through Latin America, the Darién Gap, Mexico, and the United States toward Canada. Critics praised its poetic intensity, rhythmic repetitions, and emotional force. It became Boréal’s biggest commercial success in six decades, sold more than 35,000 copies in France within weeks, and secured translation deals in roughly twenty languages.
On September 21, the same day Orélien received the Prix du Roman Fnac award voted by hundreds of booksellers and readers, the little-known X account “Balance ton Claude” a play on Anthropic’s Claude AI posted screenshots from the detector Pangram. Multiple passages, the account said, returned a near-identical result: 100 percent AI-generated. The post went viral, drawing millions of views. Several French and Canadian outlets independently ran excerpts and larger sections through the same tool and obtained similarly high scores often 95 percent or above.
Orélien rejected the charge outright. He told interviewers he began the novel in the fall of 2017 and completed a first draft in 2019 well before widely available generative models such as ChatGPT. “I have always loved creation and literature. I can’t see why I would feel compelled to use a machine,” he said. He attributed the stylistic features that detectors might flag including repetitions, strong imagery, rhythmic patterns, to the Haitian and Caribbean literary traditions that shaped him, invoking the living, oral quality of the language rather than algorithmic smoothness. He described the attacks as part of a pattern of racist hostility he had faced since arriving in France and said he was compiling a dossier of drafts and editorial correspondence with his publishers to document the book’s human origins.
Both Grasset and Boréal initially issued strong statements of support. Grasset condemned “malicious comments” that claimed the book was written or nourished by AI and reaffirmed “absolute confidence” in Orélien. Boréal pointed to the long editorial process of successive manuscript versions and the absence of any AI involvement.
Pangram, developed by a New York startup, is widely regarded as one of the stronger commercial AI detectors. Independent evaluations, including work associated with the University of Chicago and Vrije Universiteit Brussel, have reported very low false-positive rates on certain benchmarks sometimes claimed in the range of one in tens of thousands for fully human text and strong performance against both pure and “humanized” AI output. Media tests of Orélien’s novel with Pangram consistently produced high AI probabilities, while control passages from other recent novels and from Orélien’s own pre-2022 writing often scored as fully human.
Yet experts uniformly caution that no detector is definitive. Results can vary by language, genre, length of text, and the presence of non-standard or highly patterned prose.
Creative writing that leans on repetition, oral rhythms, or non-native stylistic habits has historically triggered false positives in earlier tools; studies of non-native English writing, for instance, once showed alarming error rates.
Other detectors applied to the same Orélien excerpts produced mixed or contradictory scores. Pangram itself does not claim infallibility, and researchers stress that a high score should prompt investigation, not automatic condemnation. In short, the tool raised a red flag that demanded scrutiny; it did not, by itself, prove authorship.
As the AI debate intensified, a separate investigation by the Montreal newspaper La Presse examined older journalistic pieces and short texts attributed to Orélien between roughly 2012 and 2014. It reported substantial overlaps with previously published work by other authors—sometimes near-verbatim translations or lightly adapted passages from sources in the Philippines, Peru, India, and elsewhere. Boréal responded that it takes plagiarism allegations “very seriously,” suspended promotional activities around the novel “to ensure the author’s well-being,” and said it was continuing to examine both the AI claims and the older material.
The book itself has remained on sale. These older accusations do not directly prove or disprove the use of AI in the 2026 novel, but they have complicated the public narrative and intensified pressure on the publishers.
The Orélien affair crystallizes several tensions that the literary world has so far treated mainly as abstract. Generative AI can produce fluent, polished prose at scale; detection tools lag behind evolving models and “humanizers.” Accusations of machine writing can be weaponized, especially against writers whose cultural or linguistic background differs from the statistical norms on which detectors were trained. At the same time, the ease of generating text raises legitimate concerns about originality, labor, and the integrity of prizes that still rest on the assumption of individual human authorship.
Publishers, prize juries, and readers now face practical problems: How much documentary evidence of process (drafts, timestamps, editorial notes) should be required? Should detectors be used at all in literary evaluation, and if so under what protocols? What constitutes “use” of AI—brainstorming, stylistic polishing, full drafting? The French literary establishment, already navigating questions of diversity and gatekeeping, has found itself arguing these points in real time while the Goncourt and other awards approach their final stages.
Orélien continues to maintain that the novel is his own work, rooted in lived experience and a specific literary heritage. His publishers have moved from unqualified defense to cautious investigation. The detectors have spoken, but they have not settled the matter. What remains is a human process that includes examination of manuscripts, editorial records, and the author’s own account against a technological backdrop that is only growing more ambiguous.
