In September 2026, France’s most prestigious literary prize dropped a longlisted novel after an AI detector said it was written by a machine. The author says he never used AI. The case raises the question millions of writers, students, and teachers are now asking: is this AI generated — and can any detector actually answer that with certainty? Here is what really happened, what the test scores mean, and what you should do if a detector ever flags your work.
What Happened: A Goncourt Contender Accused of Writing With AI
The short answer: in late September 2026, the Académie Goncourt removed Orélien’s novel C’était ça ou mourir from its longlist after AI detection scores suggested machine authorship — making it the first major literary prize controversy decided by a detector’s verdict.
It started on September 21, when an anonymous account on X calling itself “Balance ton Claude” (“Expose Your Claude,” a reference to Anthropic’s Claude model) published a detailed thread claiming that large portions of the novel had been written by artificial intelligence. The account said it had run excerpts, chapter openings, and full sections through Pangram, an enterprise AI text detector, and that multiple chapters returned scores indicating near-total machine authorship.
The numbers that followed were dramatic. Canadian broadcaster Radio-Canada reported running the entire 63,584-word French manuscript through the detection system, which returned an aggregate probability of 94% machine generation. Pangram’s own analysis, described by founder Max Spero, put some sections at 95% and certain passages at 100% AI likelihood. Comparative tests on other novels on the Goncourt longlist reportedly showed no similar statistical anomalies.
The fallout was swift. The Académie Goncourt announced it was removing the novel from consideration, saying it wanted to preserve the integrity of the prize. Both the French publisher Grasset and the Canadian publisher Les Éditions du Boréal had initially defended the author, but Boréal soon halted promotional activities for the book.
Orélien pushed back directly. He said he had never used AI to produce his writing, that he began work on the book in 2017, and that an early manuscript already existed by 2019. He also argued that the detectors may have misread his style: the repetitions and rhythms characteristic of Haitian and Caribbean literature, he said, may have been mistaken for the hallmarks of machine-generated text.
Why the Same Passage Got Different Scores From 8 Detectors
The short answer: when AFP ran the same passage through eight different AI detectors, the results disagreed with each other — which is exactly why a detection score is a probability estimate, not proof of authorship.
This is the single most important fact in the whole story, and it comes from real test data, not marketing claims. According to news reports on the case, the French news agency AFP tested one identical passage from the disputed novel with eight different AI detection tools — and the tools did not agree. (Data source: AFP’s comparative testing as reported in coverage of the Goncourt case, September 2026.)
That disagreement is not surprising to anyone who studies these tools. Pangram, the detector at the center of the controversy, publicly claims a false positive rate of about 1 in 24,400 documents based on its own English-language benchmarks. Two recent academic evaluations — one from the University of Chicago and one from Belgium — ranked Pangram as the most reliable detector on the market, ahead of GPTZero, Originality.ai, and RoBERTa-based tools. In other words, the tool that flagged the novel is widely considered the best available — and even it produces probabilities that other tools dispute.
The lesson: every detector output is a statistical guess about text, dressed up as a percentage. A “94% AI” score means the text looks like what the model’s training data taught it to associate with machine writing. It says nothing about who actually typed the words. For a deeper look at how these accuracy numbers are measured, see our breakdown of how accurate AI detectors really are.
Why Literary Fiction Is the Hardest Text for Detectors to Judge
The short answer: detectors measure statistical patterns in finished text, not the writing process — and a distinctive literary style can look “statistically perfect” to them.
Most AI detectors work by measuring two things: perplexity (how predictable the word choices are) and burstiness (how much sentence length and structure vary). Machine-generated text tends to be smooth, predictable, and uniform. Human writing is usually messier.
The problem is that “usually” does a lot of work in that sentence. A novelist with a controlled, rhythmic prose style — deliberate repetition, parallel sentence structures, carefully managed cadence — produces exactly the kind of smooth, patterned text that perplexity metrics flag. Experts quoted in the case coverage noted that these statistical patterns can also appear in texts using non-standard forms of English or French, which is precisely the territory of postcolonial and diaspora literatures.
That is why Orélien’s defense matters beyond this one case. If a detector can look at an award-contending novel and return 100% on some passages while other detectors disagree, then a 70% score on a student essay or a freelance article proves even less. The tool is not reading your draft history, your notes, or your intent. It is grading the surface.
What Writers Should Do If a Detector Flags Their Work
The short answer: keep evidence of your process — drafts, notes, and version history will beat any detector score in a dispute.
The Goncourt case is a warning shot for every working writer: if it can happen to a published novelist, it can happen to you. Here is the practical protection, in order of importance:
- Keep your drafts. Save version history, not just the final file. Google Docs, Word, and Scrivener all keep timestamps. A manuscript that evolved over months is the strongest possible evidence of human authorship.
- Keep your notes. Outlines, research links, character sketches, deleted scenes — anything that shows the messy human process behind the polished text.
- Don’t try to “beat” the detector. Paraphrasing tools and “humanizers” that promise to fool detectors are unreliable, and using them makes you look guilty even when you are not.
- Respond calmly and in writing. If you are accused, present your process evidence, ask what corroborating evidence exists beyond the score, and keep a record of every exchange. Our full step-by-step guide covers what to do if you are falsely accused of using AI.
What Teachers and Editors Should Do With a Detection Score
The short answer: treat an AI detection score as the start of a conversation, never as a verdict.
The most dangerous sentence in academic integrity right now is “the detector says 80%, so it’s AI.” No major detector vendor — not Turnitin, not Originality.ai, not GPTZero, not Pangram — claims their score alone proves authorship. They all describe their outputs as indicators that require human review. (Our Turnitin vs Originality.ai comparison breaks down exactly what each tool’s score does and doesn’t mean.)
If a score concerns you, corroborate it before acting: compare the piece against the writer’s previous work, ask for drafts or an explanation of the argument in their own words, and check whether the flagged passages share the stylistic traits — repetition, rhythmic parallelism, formulaic transitions — that are known to trigger false positives. A score is a reason to look closer. It is not, by itself, evidence.
Frequently Asked Questions
Can AI detectors tell if a novel is AI-generated?
Not reliably. In the 2026 Goncourt case, one leading detector scored a published novel at 94–100% AI likelihood, the author denied using AI, and eight different detectors tested on the same passage disagreed with each other. Detectors estimate statistical similarity to machine-written text; they cannot determine who actually wrote a passage.
How accurate is the Pangram AI detector?
Pangram claims a false positive rate of about 1 in 24,400 documents based on its own English-language benchmarks, and two academic studies (University of Chicago and a Belgian evaluation) ranked it the most reliable detector available, ahead of GPTZero and Originality.ai. Even so, it outputs probabilities — not proof — and its scores can conflict with other detectors on the same text.
What should I do if an AI detector flags my writing as AI-generated?
First, stay calm and do not run your text through “humanizer” tools. Gather your process evidence: drafts with timestamps, outlines, notes, and research materials. Present that evidence in writing, ask what the accusation is based on beyond the detector score, and keep records of all communication. See our full guide on what to do when falsely accused of using AI.
Can a detector tell the difference between AI-assisted editing and full AI writing?
No — and the major vendors admit it. Grammar checkers, translation tools, and light AI editing change the surface statistics of text in ways detectors cannot reliably distinguish from full machine generation. Polished human writing and lightly edited AI writing can produce nearly identical scores.
Why did the Goncourt Academy remove the novel from the longlist?
The Académie Goncourt said it removed Orélien’s C’était ça ou mourir to preserve the integrity of the prize after AI detection scores — including a reported 94% aggregate on the full manuscript — suggested machine authorship. The author denies using AI and disputes the detectors’ reading of his literary style.