The literary world is embroiled in a scandal that has ignited a fierce debate over the trustworthiness of artificial intelligence detection tools. At the center of the storm is Thélyson Orélien, a 38-year-old Haitian author who was awarded the prestigious Prix du roman Fnac on September 21 for his novel C’était ça ou mourir, published by Grasset. The same day, an anonymous X (formerly Twitter) account, Balance ton Claude, alleged that the work was “almost entirely” written by an AI.
The Accusation and the Tool at Its Heart
The account, which has around 5,000 followers, claims to have run several passages from the novel through Pangram, an AI-writing detection tool, and consistently received results indicating the text was 100% AI-generated. The post quickly went viral, amassing over 3.2 million views by the afternoon of September 22. French public radio station France Inter conducted its own test, submitting excerpts from the beginning, middle, and end of the book to Pangram. Each time, the tool returned a result between 95% and 100% in favor of AI generation.
The controversy has thrust Pangram, a New York-based startup founded by two former Stanford students, into the spotlight. Launched in October 2023, the company claims its tool is “recognized as the most reliable and accurate AI detector on the market by independent researchers, notably from the University of Maryland and the University of Chicago.”
Pangram’s Rise and Claims of Precision
While Pangram remained relatively obscure for over two and a half years, its profile exploded in July when Substack integrated it directly into its platform. The startup subsequently raised $9 million, bringing its total funding to $13 million. According to The New York Times, its monthly user base surged to 120,000 by June, up from just 2,700 a year earlier.
Pangram’s key selling point is its accuracy. The company asserts that its latest model, Pangram 4, can determine with near-perfect precision—9,999 times out of 10,000—whether a text is fully AI-generated, AI-assisted, or 100% human-written. To achieve this, it uses a unique training method: instead of a generative model, it ingested two million human-written documents, each paired with a “synthetic mirror” created by an AI using the same length, topic, and tone. When a human text statistically resembles an AI-generated one, the detector flags it.
A 2025 study from the University of Chicago found that Pangram maintained near-perfect accuracy on texts of around 500 words or more, surpassing competitors like Originality.ai and GPTZero. Pangram claims a false positive rate of just 0.0041% (one in 24,000 documents) and a missed AI detection rate of 0.34%.
Cracks in the Algorithm: Known Flaws and Edge Cases
Despite these impressive figures, the tool’s reliability has been challenged. Users have found that Pangram can systematically flag content as AI-generated even after substantial human rewriting. Freddie deBoer, a Substack author, found that adding a few AI-written sentences to a human-written paragraph caused Pangram to flag the entire text as AI-generated.
A significant caveat is that Pangram’s training data consists of two million documents dating from before 2022, largely in English. As a blog post on the website of Write Human, an AI paraphrasing tool, notes, “the human style is increasingly moving towards automated phrasing as people read more, making the boundary between [AI and human writing] blurrier every year.”
While Pangram claims to detect AI content in about twenty languages, including French with a stated 100% accuracy rate, questions arise when an author’s style falls outside the norm. François Negroni, an author who co-writes aviation books with AI, offered a widely shared perspective on X. He explained that Orélien, “born in Haiti (and) arrived in Quebec after the 2010 earthquake… writes in a French that critics unanimously describe as ‘creolized,’ mixed with Haitian Creole, Spanish, Portuguese, infused with orality, aphorisms, and rhythms from elsewhere.” Negroni concluded, “It is precisely this type of prose that AI detectors handle the worst.” Orélien himself told Libération, “I think the software is used to French literary traditions and classical authors.”
A History of False Positives
Doubts about these tools are not new. In April, author Bernard Minier recounted a troubling experience to HuffPost. “I copy-pasted a paragraph from ‘Ruptures’ (his latest book), and it replied: ‘100% written by an AI, this text will be refused by a publishing house or a university.’ I tried again, same thing. I thought, wow, I wrote like an AI. So I took a good old ‘The Black Echo’ by Michael Connelly, which no one could suppose was written by an AI unless he had a time machine. ChatGPT also told me it was ‘100% written by AI.’ So even this simple thing is not reliable at all.”
Pangram faced a similar embarrassment earlier this summer when it flagged Jamir Nazir’s short story The Serpent in the Grove, which won the 2026 Commonwealth Prize for Best Short Story, as 100% AI-generated. The author was eventually cleared after presenting drafts, plans, and evidence of his artistic process to the organizing committee.
The Balance ton Claude account maintains its position, stating it tested four pre-2022 texts by Orélien on Pangram, all scoring 0% AI, but several post-2023 texts (excluding his novel) scored between 76% and 100% AI. The controversy continues to raise critical questions about the role and reliability of AI detection in creative fields.
