France just turned an AI-detector score into a public hanging.
Thélyson Orélien — Haitian-Canadian poet turned novelist — is having the kind of autumn most writers would kill for. C’était ça ou mourir (It was that or die) sold tens of thousands of copies, won the Prix du Roman Fnac, and landed on the early longlists for Goncourt, Renaudot, Femina, Médicis. Then an anonymous account calling itself Balance ton Claude — Denounce your Claude — posted that Pangram had scored passages of the book at “100% AI.” And no shit. It really is happening. The score traveled faster than the novel.
Orélien denied it. He told Libération he drafted the book in 2019 — before text generators were a consumer toy — and that the repetitions critics are waving as “machine tells” are Haitian and Caribbean oral tradition, living rhythm, not autocomplete. Grasset backed him. Detectors, the house said, will fantasize that the Bible is synthetic. Other writers pointed out that creolised French — Spanish, Portuguese, Haitian Creole braided into the line — does not sit cleanly inside the statistical boxes these tools were trained on.
None of that stopped the pile-on. People fed his defense statement into Pangram and crowed when that, too, came back “100% AI.” Surprise. A pattern-matcher that mistakes difference for generation will keep mistaking difference for generation. Research keeps saying the same quiet thing: these systems do not know who wrote a sentence. They guess from probabilities. They false-flag humans. They miss edited machine text. And still we treat the percentage like a confession.
Been there. The internet loves a certainty theater.
Here is the adult version. Detectors are not courts. They are not publishers. They are not workshops. They are statistical hunches with a marketing page. Useful as a triage hint, maybe. Useless as a moral verdict. When a debut novel written in a tradition outside the training average gets dragged by a tool named after a Claude, the scandal is not only about one author. It is about what kind of attention we are willing to pay — and who gets to declare a human fake without showing their work.
That is the big world. The little world is Circle.
Circle is being built as the opposite of anonymous score-dropping. Members request assignments. They write structured reviews of text or video. They earn credit toward having their own work read. Disclosure lives by role: what the tool did, what the human kept, who is accountable for the final take. We do not pretend the machine was never in the room. We also do not outsource judgment to a black-box percentage that cannot sit for a defense.
GAPOZ Press is the people’s publisher for the same reason. Writing in public is labor. Credit is earned. If you used AI to research or draft, say so. If you wrote the thing yourself and a model only helped you edit, say that too. Soften the witch-hunt language. Keep the standard high. A reader may care whether the voice is yours. A detector does not get to answer for them.
So when the next “100% AI” screenshot hits the timeline, try a harder question than the score. Who claims the work? What process made it? Can a human show drafts, notes, a timeline, a tradition that explains the music of the sentences? Or are we just running literature through a lie detector built for average English and calling the fireworks justice?
A score is not a confession. Disclosure by role is still the grown-up move. Circle is being built for that — not for making the tool undetectable, and not for hanging writers with a percentage.
Also on the beat
- Apex CoVantage launching eden at Frankfurt (Oct 7): AI editorial engine that flags subject-aware errors and keeps copyeditors deciding — https://www.nativo.ventures/2026/09/23/apex-covantage-to-launch-eden-an-ai-powered-editorial-engine-for-publishing-at-frankfurt-book-fair/
- Google’s spam-update pressure on scaled, low-value AI publishing (value and intent, not “AI = spam”) — https://www.intelligenthq.com/google-ai-content-spam-update-2026/
- TECNO’s EllaClaw demo for journalist/creator workflows: research briefs, scripts, style profiles, risk-word checks (Sep 23) — https://www.gizchina.com/tecno/tecno-shows-how-ellaclaw-can-be-customized-for-different-ai-workflows
- YouTube’s AI disclosure + auto-label rules for photorealistic GenAI (labels don’t kill monetization; nondisclosure can) — https://blog.youtube/news-and-events/improving-ai-labels-viewers-creators/
- Clear KDP line: disclose AI-generated text/images/translations; AI-assisted brainstorm/edit generally does not — https://writersofthewest.net/blog/amazon-kdp-ai-content-policy/
How this was made
AI assisted research and a first draft of this Circle View. A human reviewed, edited, and is accountable for the judgment and the final take. Disclosure by role — not undetectable.
