How Do AI Text Detectors Work? UncovAI's Florian Barbaro Explains the Thélyson Orélien Case
One AI detector scored passages of a bestselling, prize-winning novel at up to 100% likely AI-generated. Another judged the same kind of text very likely written by a human. Florian Barbaro, founder of French AI-detection platform UncovAI, breaks down why that gap exists, and what it means for how detector scores should actually be read.
AI text detectors score writing on statistical patterns, not certainty, so different tools trained on different data can reach very different conclusions about the same text. That's exactly what happened when Québécois-Haitian author Thélyson Orélien was accused of using AI to write his novel C'était ça ou mourir: the detector Pangram flagged some passages near 100%, while AFP's own verification tool and several other detectors it tested judged similar passages likely human-written. No detector result, however extreme, should be treated as proof on its own. UncovAI's Text Detector is built to be used the same responsible way: as one signal to check, not a final verdict.
What Happened With Thélyson Orélien's Novel
C'était ça ou mourir, a novel following a Haitian migrant's journey to Canada, became one of the breakout successes of France's autumn literary season after its August release with Grasset, selling roughly 35,000 copies in five weeks and winning the Prix Fnac. It was also under consideration for major prizes including the Goncourt, Renaudot, Médicis and Femina.
On the evening the book won the Fnac prize, an anonymous X account called "Balance ton Claude" accused Orélien of writing the novel almost entirely with AI. The account, run by a small group presenting themselves as ordinary readers, submitted passages to Pangram, a detector the group described as one of the most reliable available, and reported scores between roughly 95% and 100% likely AI-generated. Some outlets that re-ran the full novel through Pangram themselves got a similar result.
Orélien rejected the accusation, describing it as part of a pattern of racist attacks since his arrival in France and saying he is compiling documentation with his publishers, Boréal and Grasset, to establish his authorship. Grasset called the claims rumors tied to a possible hate campaign. When AFP tested the same selected passages through other tools, including its own co-developed InVID-WeVerify, the results diverged sharply, with that tool judging the passages very likely human-written, and seven further detectors AFP tried producing a spread of different conclusions. As of this writing, AI use in the novel has not been publicly established.
Le Parisien, covering how these detectors actually work, quoted UncovAI founder Florian Barbaro on Pangram's standing in the field: he described it as a well-regarded tool in this young industry, already adopted by major platforms for content filtering.
How AI Text Detectors Actually Work
AI text detectors don't "know" whether a human or a machine wrote something. They compare statistical properties of a passage, patterns like how predictable each word is given the ones before it, how much sentence length and rhythm vary, and how repetitive the vocabulary is, against patterns learned from large sets of text known to be human-written or AI-generated. The output is a probability score, not a fact.
Mila's scientific director Hugo Larochelle, commenting on the Orélien case, noted there's no theoretical guarantee that these tools are perfect. That's a structural property of the approach, not a flaw specific to any one detector.
Why Detectors Disagree on the Same Text
- 1Moving target. AI models are updated constantly, and detectors are trained on the models that existed before them, so they're always playing catch-up to the latest writing style AI can produce.
- 2Mixed authorship. Text that blends human writing with AI editing, suggestions, or light assistance is much harder to score cleanly than text that's purely one or the other.
- 3Distinctive human style. A repetitive, formulaic, or otherwise atypical personal writing style can look statistically similar to AI output and trigger a false positive.
- 4Different training data. Each detector is built on its own model and reference dataset, so two tools can weigh the same stylistic signals very differently and land on opposite conclusions.
Even Pangram's own co-founder, Max Spero, has said publicly that the tool shouldn't serve as a final arbiter, but as a starting point for further verification. That's a reasonable standard to hold any detector to, including UncovAI's.
How to Read a Detector Score Responsibly
Treat a high or low score as a lead to investigate, not a verdict to publish or accuse someone over · check whether multiple detectors agree, since a single tool's outlier result is common · consider whether the writer has an unusual or distinctive personal style that could trigger a false positive · remember that partial AI assistance (editing, brainstorming) scores very differently from a fully AI-generated text, and detectors can't always tell the two apart.
This matters most exactly when the stakes are highest, like a public accusation against a named person's professional work. A detector score is evidence to weigh, not proof to publish.
How UncovAI Can Help
UncovAI builds the same class of detection tools discussed in the Orélien case, designed to be used as a responsible first signal rather than a final verdict:
| Tool | What it checks |
|---|---|
| Text Detector | Manuscripts, articles, essays and messages, for a confidence score on AI-generated writing |
| Image Detector | Cover art, author photos and graphics for AI generation or manipulation |
| Video Detector | Recorded interviews or promotional video for deepfake or face-swap signs |
| Audio Detector | Audiobook narration or interview audio for synthetic or cloned voices |
| Meetings Protection | Live video calls, such as author interviews on Zoom or Teams, checked for AI manipulation in real time |
| URL Phishing Detector | Links shared alongside claims or "evidence," checked before you click |
| Browser Extension | Scans text, images and links on any page as you read, without switching tabs |
| WhatsApp & Telegram bots | Instant checks on a passage, image or clip sent directly in chat |
Publishers, newsrooms, and educators dealing with authorship questions can also reach for UncovAI's full product suite, and organizations that need to keep sensitive manuscripts or unpublished work entirely in-house can ask about UncovAI's private, zero-data-retention deployments, run on-premises within an organization's own systems. Get started free to run your first check.
Frequently Asked Questions
What happened with Thélyson Orélien's novel?
In September 2026, an anonymous X account called Balance ton Claude accused Québécois-Haitian author Thélyson Orélien of writing his bestselling novel C'était ça ou mourir largely with AI, citing a near-100% score from the detector Pangram on some passages. Orélien and his publishers, Boréal and Grasset, denied the accusation and called it a hate campaign. When outlets including AFP ran the same passages through other detectors, results diverged sharply, including one tool that judged the text very likely human-written.
How do AI text detectors actually work?
AI text detectors analyze statistical patterns in writing, such as word predictability, sentence rhythm and vocabulary variety, that tend to differ between human and AI-generated text, and output a probability score rather than a yes-or-no verdict. They are trained by comparing large sets of human and AI-written text, which means their accuracy depends heavily on how closely a given piece of writing resembles the examples they were trained on.
Are AI text detectors reliable?
AI text detectors are useful as a signal but not as a final verdict. AI models change constantly, so detectors are chasing a moving target, and text that mixes human writing with AI assistance is especially hard to score accurately. A distinctive or repetitive personal writing style can also trigger a false positive. Even Pangram's own co-founder has said the tool should be a starting point for further checks, not a final judgment.
Why did different AI detectors give different results for the same text?
Different detectors are built on different models and training data, so they weigh stylistic signals differently. In the Orélien case, Pangram scored some passages between 94% and 100% likely AI-generated, while AFP's InVID-WeVerify tool judged the same kind of passages very likely human-written, and seven other detectors AFP tested produced a range of divergent results. This spread is common and is one reason experts caution against treating any single detector score as proof.
How can UncovAI help verify whether text was AI-generated?
UncovAI's Text Detector scans written content and returns a confidence score on whether it was likely written by a person or generated with AI, in seconds. Like any detector, it should be used as one input alongside context, not as an automatic final verdict, especially for high-stakes accusations.
Sources
- La Presse: L'affaire Thélyson Orélien en six questions
- RTS: Accusé d'utiliser l'IA, l'écrivain Thélyson Orélien se défend
- Franceinfo: L'écrivain Thélyson Orélien accusé, Grasset évoque "une campagne de haine"
- Moustique: Pourquoi les détecteurs de textes générés par IA ne sont pas fiables
- Le Parisien: Comment fonctionnent ces détecteurs d'IA dans un texte (quotes UncovAI founder Florian Barbaro)
As of publication, whether AI was used in the writing of C'était ça ou mourir has not been publicly established; this article reports the ongoing, contested dispute and does not assert a conclusion either way.
Check Text the Responsible Way
A single detector score should never be the whole story. UncovAI's Text Detector gives you a fast, clear signal on AI-generated writing, meant to be one part of your verification process, not the final word.