“It was that or die”: what we know about the AI usage allegations against Thélyson Orélien

The issue surfaced on September 21, just hours after Orélien was awarded the Fnac Novel Prize. The account “Balance ton Claude,” which had very few followers at the time, claimed to have analyzed excerpts from the book using Pangram, a detector for AI-generated texts. According to this account, the results indicated that the writing was “almost entirely” artificial. The post quickly went viral, triggering a wave of checks, comments, and responses.

Le Monde subsequently submitted more than half of the novel to Pangram. The detector classified over 95% of the tested text as likely generated by AI, with a high level of confidence. The newspaper also tested the opening pages of about fifteen other novels released this season, which the tool identified as human-written. These results raise questions but do not, on their own, provide material proof of how the manuscript was produced.

This is precisely the central point of caution. Pangram is among the most highly rated detectors in several recent academic studies. A study by researchers at the University of Chicago credited it with very low error rates on a corpus that mixed human and AI-generated texts. However, the University of Chicago itself emphasizes that no detector is infallible and recommends that such tools should never be used as the sole basis for an accusation.

Orélien denies any writing by artificial intelligence.

Thélyson Orélien has unequivocally rejected the allegations. In an interview with Quebec media, he stated that he has “never used artificial intelligence to produce texts.” He also explained that the genesis of the book dates back to 2017, with an initial version of the manuscript existing in 2019, several years before the widespread availability of generative AI tools to the public.

This timeline is an important element of his defense. However, it does not, on its own, establish what may or may not have been used during later phases of rewriting, editing, or proofreading. The author is currently working with his publishers in France and Quebec to compile a dossier documenting his writing process.

Grasset, which has published the novel in France since August 19, has dismissed the accusations and condemned the malicious comments. Boréal, which published the book in Canada in March, stated that it is aware of the allegations while maintaining its confidence in the author’s work. At this stage, neither publisher has announced any withdrawal, commercial suspension, or official challenge to the authorship of the text.

Why the controversy has gained such traction.

The controversy arises at a time when It Was That or Die is experiencing significant exposure. The novel has sold approximately 35,000 copies in five weeks in France and is included in the selections for several major literary prizes, including the Goncourt, Renaudot, and Femina. Its Fnac Prize, awarded just hours before the controversy began, has further heightened attention on the book and its author.

The novel follows Jonas, a migrant, and is set against a backdrop of displacement, survival, and exile. Orélien, 38, left Haiti for Canada in 2010. The critical success of the book is partly attributed to its rhythm, repetition, and voice, characteristics that the author also cites to explain why a statistical detector might interpret his writing as artificial.

The identity of the account behind the accusation remains unknown. “Balance ton Claude” claims to be run by several individuals and explains that it aims to raise awareness about the growing use of AI in written productions. The account has not published any documents directly establishing the writing process of the novel, such as file histories, successive versions of the manuscript, or evidence of using a generative model. Its argument primarily relies on the results from the Pangram detector.

The issue thus extends beyond just the Orélien case. The controversy raises a new problem for literary publishing: how to reliably establish the use or non-use of AI when the only available clues are probabilistic style analyses? Researchers and specialists in academic integrity are already recommending that these tools be combined with other elements, such as version histories, drafts, metadata, and documented creation processes.