The book C'était ça ou mourir has received the Fnac novel prize, the Medusa and some others, has sold more than 35,000 copies and is among the candidates for the great literary autumn awards. As of this Wednesday, the book has another merit: it is the first major French scandal of literature allegedly made with artificial intelligence (AI). Its author, the Canadian of Haitian origin Thélyson Orélien, denies this. He says that he finished the first draft in 2019 (ChatGPT came out at the end of 2022) and that his “Haitian and Caribbean” literary tradition may have caused confusion because “the images, repetitions and rhythm transmit a living word,” he explained in French media. His publisher, Grasset, is currently supporting him. The complaint against Orélien arose from an anonymous account on X.
The problem with Orélien is that the time when no AI-generated text detector was truly efficient at finding culprits is long gone. Because now there is one that gets a lot right: Pangram, launched in 2024 in the US. And it is also multilingual. “Pangram works well in 20 other major languages apart from English, and that includes Spanish and French,” explains Max Spero, co-founder of Pangram, to EL PAÍS.
Pangram's success rates are more than 99%: only 1 in every 24,000 samples goes wrong, according to the company. The scientific community has also proven that Pangram works better than the alternatives. “There is no reason to think that these models do not work, because they are trained with an enormous amount of data, both written by humans and generated by AI,” says Tuhin Chakrabarty, professor of computer science at Stony Brook University (USA) and a specialist in text generated with AI.
Pangram is also AI. It is a model trained to detect text generated by machines. You are shown huge quantities of human and similar AI-generated articles, novels or essays until you grasp the differences in style. Since 2024, the company has been releasing new versions of its software. You can do a free trial, but it works with a subscription. In August 2025 it had only 7,000 users; Last month there were already 222,000.
Professor Marzena Karpinska, from Simon Fraser University (Canada), verified the successes of a previous version in a 2025 scientific article about how the people who use ChatGPT the most are the best at later detecting their writing style: "The current version of Pangram, version 4, is much more precise. It is like the difference between GPT-4 and GPT-6. They have made an effort to ensure that their tool does not have biases that are seen in others, such as labeling as IA the texts of people whose English is not their native language,” he explains.
The doubts with Orélien are not just because of a writer. The French literary world has showered the novel with praise and awards. How can it be that Pangram categorically says that Orélien's novel has been generated with AI and no one has noticed? Because they don't use AI and don't know how it writes, Chakrabarty believes. "Critics hate AI. They have never read what AI prose is like. How are they going to recognize it? One reason why our institutions fail to detect AI prose is that they are not familiar with it. People like me, who dedicate themselves to this 24 hours a day, are not part of award juries. I read this French novel and it has the same patterns as bad fiction in English: negative parallelisms, mixed metaphors, clichés. All very loaded," he explains. In the US there have also been several controversies with books and articles detected by Pangram.
Since the popularization of these tools began, there has been a race to avoid their detection: adding typos, “humanizing” phrases with other software and looking for prompts that fool the detector, such as asking it to imitate the style of a famous writer. This newspaper has tested examples of texts obtained from Claude, but with the style of Javier Marías or David Trueba: Pangram has declared 100% AI. “I've done this experiment tons of times,” says Chakrabarty. EL PAÍS has also tested the first two chapters of Santos Cerdán's book, The Fall, and Pangram gave between 70% and 100% confidence that it is AI. “These are enough numbers to say that there has been intensive use of AI,” says Spero. Cerdán has not responded to messages from this newspaper.
The most sophisticated tricks require computer skills that are unusual among writers. "If you fine-tune a model with the books of a famous writer, which is something I have done, it is still possible to cheat. But no one does that because it is very expensive and requires a lot of technical knowledge. These writers are not computer scientists, but rather they go into ChatGPT, give it a prompt and ask it to write each chapter," says Chakrabarty.
A big problem with this tool is that the authors can forever deny the use of AI and it is impossible to prove. Without confession, everything remains suspicious. And it can also lead to a witch hunt if there are any false positives. These days, the networks in France have been filled with the now famous Pangram chart with people saying that texts prior to the launch of ChatGPT have been generated with AI. In one case, Spero offered $100 to the author of one of those captures, a famous journalist, to give him the text he had used. The journalist, at least in public, did not order it.
Pangram is worse at detecting short texts, just one or a few sentences. The more sample paragraphs, the more reliable. Massive use in all chapters of a book is a huge red flag, almost indefensible: "Pangram works better with long texts. Below 50 words it is not reliable, and the more text there is, the better. Likewise, it works better with creative writing than with recipes, for example," says Karpinska, who did an extra check with Orélien, whose defense is that his style is different, far from the dominant traditions. "Some have suggested that the problem is the author's specific Haitian French. I found an article of his from 2015 and analyzed it. The result was that he was human, so it does not seem, at least with this example, that Pangram is detecting any particularity of the dialect," adds Karpinska.
