A French novel titled C’était ça ou mourir by Thélyson Orélien has sparked controversy after being accused of being largely written by an AI. The book, which has already won prestigious awards such as the Fnac Prize and the Méduse Prize, and is being strongly considered for the Goncourt and Renaudot Prizes, has received widespread critical acclaim. However, since September 21, a post on the social media platform X titled "Balance ton Claude" has accused the novel of being "almost entirely" generated by artificial intelligence. This claim has led to heated discussions online, especially after some passages of the book were flagged as "100% AI" by the text analysis tool Pangram. The controversy stems from the use of Pangram, an AI detection tool that analyzes text to determine whether it was written by a human or an AI. The company claims its tool has a very low false positive rate—only about one in 24,400 cases where a human-written text is incorrectly labeled as AI-generated. However, critics have pointed out certain repetitive structures in the novel, such as the phrase "It is not... it is...", which are commonly seen in AI-generated texts. Some users attempted to calculate the probability of the book being AI-generated by multiplying the false positive rate by the number of flagged passages. This approach, however, was criticized as flawed, as the passages are not independent and could share similar stylistic traits. Pangram’s detection method involves a process called "synthetic mirroring," in which the system starts with human-written texts and uses large language models to generate similar AI texts. It then trains a classifier to distinguish between the two. The company also employs "hard negative mining," a technique that reintroduces previously misclassified human texts into its training data to improve accuracy. However, Pangram explicitly states that its tool provides a statistical probability rather than a definitive answer. It cannot be 100% certain whether a text is AI-generated or human-written, especially with shorter texts or when context is limited. The CEO of Pangram admitted in an interview with Wired that the tool is less effective on short texts, particularly those with fewer than 100 words, and that the same passage can yield different results depending on the context in which it is analyzed. Some studies have also shown that AI detectors may be more likely to misidentify texts written by non-English-speaking authors. While Pangram claims to have trained its latest version on a wide range of languages and text types, including literature, the exact content of its training data remains undisclosed. In France, it seems unlikely that the controversy surrounding C’était ça ou mourir will be fully resolved. The nature of AI-generated text and the limitations of detection tools mean that any accusation will always remain a possibility. Regardless of whether the book was partially, entirely, or not at all written by an AI, it has clearly resonated with French critics and is set to be translated into multiple languages. The debate highlights the growing complexity of distinguishing human creativity from machine-generated content in the literary world.