A new novelist won a big French book prize on September 21, 2026. That same day, an anonymous account on X said his book was written by AI. Five days later, one of the most famous literary prizes in the world, the Prix Goncourt, dropped him from its list. The main evidence against him was a score from an AI detector. So here is the question millions of students, writers and workers now want answered: can an AI detector actually prove that someone used AI?
The short answer is no. A detector can raise a flag. It cannot prove anything on its own. Here is why, in plain words.
What happened with the prize-winning novel?
The author is Thélyson Orélien, a 38-year-old Haitian-Canadian writer. His debut novel, C’était ça ou mourir (“It was either that or die”), follows a man who flees gang violence in Haiti and travels across the Americas. It won the Prix du roman Fnac, a well-known French book award.
Then an anonymous X account called “Balance ton Claude” ran passages through an AI detector called Pangram. It said the book was almost entirely AI-written. According to Le Monde, Pangram rated more than 95% of the book as AI-generated. The post spread fast.
But other tests did not agree. When the news agency AFP ran parts of the novel through several detectors, the results went all the way from “very likely” written by a human to “100 percent” AI. Orélien denies using AI. He says his repeated phrases and rolling rhythm come from Haitian and Caribbean writing traditions, not from a chatbot.
On September 26, the Goncourt academy removed the book from its longlist. It pointed to “converging analyses” about AI and to separate plagiarism claims about his older journalism. His French publisher, Grasset, said it still has confidence in him. The Fnac prize has stood by its winner.
Can an AI detector prove a book was written by AI?
No. An AI detector gives a guess, not proof. It looks at the words on the page and says how much they look like machine writing. It cannot see how the text was made, who typed it, or which tool was used. Even the maker of Pangram says its score should be a starting point for a closer look, not the final word.
That warning came from Max Spero, the chief executive of Pangram, speaking to AFP. It is worth taking seriously, because Pangram is one of the best tools in tests.
Simple takeaway: an AI detector tells you text looks like AI writing. It never tells you a person used AI.
How do AI detectors work?
AI detectors are programs trained on huge piles of human writing and AI writing. They learn the small habits that show up more often in machine text. Then they score new text by how closely it matches those habits.
Thierry Poibeau, a language-technology researcher at France’s national research centre CNRS, explained some of the habits these tools look for:
- Lists of three things in a row (“fast, cheap and easy”)
- Lots of long dashes (—)
- Sentences that all have the same shape
- The same words and phrases coming back again and again
Here is the problem. Real people write like this too. Poets repeat lines on purpose. Some cultures tell stories with rhythm and repetition. A detector may mistake a writer’s personal style for a machine’s fingerprint.
How accurate are AI detectors really?
The best ones are very accurate in lab tests. In a 2025 study, University of Chicago Booth researchers Brian Jabarian and Alex Imas tested Pangram, GPTZero and Originality.ai. You can read a summary in the Chicago Booth Review. They found:
- All three paid tools wrongly flagged human writing less than 1% of the time.
- Pangram did best, with almost zero false alarms on medium and long passages.
- All three got worse on short passages of under 50 words.
- A free, open-source detector they tested did not do much better than guessing.
That sounds great. But a lab test is not the same as judging one real book. In a lab, researchers know the true answer for every sample. In real life, a single unusual writer may be exactly the kind of text the tool rarely saw during training.
Older research shows how badly this can go wrong. A 2023 Stanford study, published in the journal Patterns, tested seven detectors on 91 essays written by non-native English speakers for the TOEFL English exam. On average, the detectors wrongly called 61.3% of those human essays AI-written. The same tools judged essays by US eighth-graders correctly. Simpler word choices made real people look like machines.
Simple takeaway: a tool can be right 99% of the time and still be wrong about you.
Why do different detectors give different answers?
Each detector is trained on different examples and uses its own rules. So each one has its own blind spots. One may be very strict about repetition. Another may care more about word choice. That is why the same pages from Orélien’s novel got scores ranging from “human” to “100% AI.”
There is also a moving target. New AI models write differently from older ones. A detector trained last year may not know this year’s style. And “humanizer” tools, which rewrite AI text to hide it, keep getting better. It is a constant race.
What counts as real proof that someone used AI?
Real proof is about how the text was made, not just how it reads. A score alone is weak. Stronger evidence includes:
- Drafts and version history. Word and Google Docs save older versions. Real writing usually grows slowly, with edits and mistakes.
- Notes, outlines and research. Books leave a paper trail.
- People who saw the work grow. Editors, teachers or writing groups.
- Copied mistakes. Made-up facts or fake sources that chatbots are known to produce.
- A chat history showing the text being generated.
If you teach, hire or judge writing, treat a detector score like a smoke alarm. It tells you to go and look. It does not tell you where the fire is. We saw a similar problem when looking at how accurate AI essay grading is: the machine is useful, but it should not be the only judge.
What should you do if you’re falsely accused of using AI?
Stay calm and show your process. Share your drafts, file history and notes. Ask which tool was used and what score it gave. Point out that detector makers themselves say a score is not proof. If you can, write a short sample in front of the person judging you.
It also helps to know that you are not alone. Students, job applicants and freelance writers face these claims every week. And as we covered in why fans get angry when creators use AI, even a rumour can hurt a reputation fast. Keeping your drafts is now simple protection.
The bigger lesson from the Orélien case
We may never know for sure how this novel was written. That is the point. A score from a single tool, posted by an anonymous account, was enough to shake a prize and a career within days. Some experts think the tool may be right. Others think a writer’s style was misread. Either way, the decision rested on evidence that its own makers say is not final.
AI detectors are useful tools. They are not lie detectors. Use them to ask better questions, not to hand out verdicts.
Frequently Asked Questions
Can AI detectors be wrong?
Yes. Even the best AI detectors sometimes flag human writing as AI, especially short texts, simple writing or unusual styles. A 2023 Stanford study found detectors wrongly flagged 61.3% of essays by non-native English writers.
Which AI detector is the most accurate?
In a 2025 University of Chicago Booth study, Pangram had the fewest false alarms, followed closely by GPTZero and Originality.ai. But no tool is perfect, and accuracy drops on passages under 50 words.
Why was Thélyson Orélien removed from the Goncourt list?
The Goncourt academy cited “converging analyses” suggesting his novel was largely made with AI, plus separate plagiarism claims about his older articles. Orélien denies using AI, and other detectors rated parts of his book as human-written.
How can I prove I didn’t use AI?
Keep your drafts, notes and document version history. These show your writing growing over time, which a chatbot cannot fake easily. Offering to write a new sample in person can also help.
