The Novel Nobody Could Prove Was Written by a Machine

On 21 September, the day Thélyson Orélien's novel C'était ça ou mourir won the Prix du roman Fnac in France, an anonymous X account called Balance ton Claude — the name is a joke about Anthropic's Claude — posted that it had run the book through an AI detector and found it "almost entirely" written by artificial intelligence. Within days the Académie Goncourt had removed the novel from the longlist for France's most prestigious literary prize, saying its aim was "to affirm that we neither endorse nor condone in any way the creation of texts generated with the help of AI." No jury had read drafts. No auditor had been appointed. The biggest authorship dispute in European publishing this year was opened by a stranger with a screenshot.

The stranger later had a name. Le Figaro reported that the anonymous account belonged to one of its own columnists, Samuel Fitoussi, who described himself as "techno-optimistic" and said he had begun the campaign because he feared "AI will cause the disappearance of reading and writing." What tipped the academy from initial calls for caution to removal were plagiarism allegations surfaced by Libération; Orélien later admitted to past "unacceptable mistakes" in his writing, including a competition text from 2012 that had borrowed from a preface by Louis Pauwels. And the publisher, Canada's Les Éditions du Boréal, initially stood by the author: "Without credible, concrete evidence," it said, "it is not our place to decide whether this book has a right to exist." That sentence, and what replaced it, is the story.

What the detector actually said

Pangram deserves better than the way this affair used it. It is not the junk detector that flattered and ruined reputations in 2023. Its own benchmarks, published on the company's blog on 16 October, report a target false-positive rate of 1 in 10,000 and an observed rate of 0.0041% across 2,046,273 human-written documents, with a headline false-negative rate of 0.3396% on nearly 520,000 AI-generated documents. On creative writing specifically — the slice this case sits in — it misses 0.269% of AI text. Those are the vendor's own figures, and it has published third-party evals alongside them.

Pangram is a tool that makes statistical predictions. We deliberately tune Pangram to be more inclined to produce false negatives rather than false positives; we have made this choice because while an occasional false negative can be annoying, a false positive can cause real harm.

— Pangram, "Is Pangram accurate?", 16 October 2026

A vendor tuning away from false accusations is behaving more responsibly than most of its competitors did three years ago. But look at what the score was asked to decide: whether a novel was "almost entirely" machine-written. Pangram answers one question — does this text resemble the statistical signature of machine output — and the prize committee heard it answer a different one: was this book written by a machine. The gap between those questions is the entire affair.

Three questions, one score

There are three different claims people collapse into one: this text resembles AI output; this text was produced by AI; and this book is the work of its credited author's mind. A classifier sees an artefact. It cannot see a process — the draft typed by hand then heavily reworked, the passage generated then rewritten three times, the research assistance the author conceded from the start. "Here and there I did use tools for research, particularly for foreign idioms," Orélien told AFP, while denying that "not a single sentence in this book" was created by a machine. Human prose, machine prose, and the mixed, edited, revised thing that most AI-assisted writing actually is live in one statistical space. No amount of tuning extracts authorship from a paragraph.

The field test of a detector is whether its scores track reality, and the affair collapsed exactly there. One detector family said 100% machine; others, as French media reported, said very probably human, on the same text. If that is what AI detection comes to when a book is scrutinised by half the francophone internet — the most examined novel of the year — then a detector score is not a verdict. It is one thermometer, read by a man with a motive.

The file talks when the prose cannot

On 21 October, Boréal announced it was ending its professional relationship with Orélien. It had gathered every working version of the novel it held and hired an outside party to analyse the drafts. The finding, in the publisher's own words:

The results obtained this week revealed traces of AI tags in the metadata of certain documents.

— Les Éditions du Boréal, publisher's statement, reported by AFP

That is the most interesting sentence in the whole affair, and not for the reason it was written. Documents record how they were made: application fields, author names, revision histories, and the properties that AI-assist features write into a file when they touch it. Orélien's own defence is forensic reasoning too — he began the book in late 2017, he says, and finished a first draft of more than 300 pages in 2019, years before ChatGPT existed; a timeline as an alibi. The case ended up weighing document traces because the text alone could never carry the verdict.

Then watch what Boréal conceded in the same statement: the publisher did not know the extent of AI's use in the writing process, "and we will likely never have a definitive answer." The author stopped cooperating with the inquiry. The metadata showed a tool had touched certain documents. The extent — a sentence, a chapter, the shape of the book — is now unknowable to the company that published it. That is as complete an admission of method failure as you will read this year, and it made the front page anyway.

Provenance, not forensics

Our position, labelled as the opinion it is: publishing will not solve its AI problem with better detectors. It will solve it — slowly, unevenly, and late — with provenance and disclosure. The regulatory piece already exists.

The direction of travel is clear: machine-made content will increasingly arrive carrying its own label, and the gap between a detector's guess and a document's record will keep widening. What a publisher that actually wants answers can demand is process, not vibes: dated drafts, version histories, and a written AI-use statement naming tools and purposes — the sort of disclosure this site makes on every post we publish, because the only verifiable fact about tool use is the one the writer states, and a statement is only trustworthy when it is the norm rather than the confession.

Authors should hear the warning in the same story: whatever you did with a writing tool, your file may know it. Metadata is testimony now. And scrubbing it — which every leak teaches the market within a year — will be the next scandal. That is one more reason the answer has to be a disclosure norm and not a forensics arms race: a publisher at war with its own authors' file properties is fighting the wrong enemy.

The strongest case against our side

There is a serious counter-argument, and it goes like this: publishing has never verified authorship. Ghostwriters, editors, agents, writing teachers reworking promising manuscripts — the industry built great books on a polite fiction that the name on the cover did all the typing, and nobody demanded anyone's revision history, because the effort of a book was human-sized and that assumption quietly carried the norm of authorship for centuries. AI is what broke the assumption. But the remedy the detector-boosters propose — secret classifiers rendering verdicts on prose — would be intolerable in any other domain, and it cannot even agree with itself on the most heavily examined novel of the year.

If a norm of disclosure had existed — written AI-use statements at contract, applied evenly, enforced for concealment rather than use — the Goncourt would not have needed an anonymous account and a screenshot to decide what to do about an AI-suspected book. Boréal nearly had it right in September, with its demand for credible, concrete evidence. It should not have needed an analysis firm and a metadata sweep in October.

What the metadata cannot say

But not a single sentence in this book was created by a machine, nor the meaning or the structure of a story that comes from my gut.

— Thélyson Orélien, statement to AFP

Nobody can check that sentence — not the Académie Goncourt, not the outside analysts, not the anonymous account with its detector — and Boréal has conceded in writing that nobody ever will. That is the real verdict of the Orélien affair: a book lost its prize and an author lost his publisher without anyone in the room owning a method. Provenance will keep growing: EU labels, tool markings, disclosure clauses in contracts. But until novels arrive with receipts, the honest answer to "did AI write this?" is the author's own. The only system that makes that answer worth having is one that asks early, asks everyone, and punishes concealment harder than use. Publishing is currently doing the opposite: letting a stranger with an unaudited detector ask late, and decide.

We use AI tools in our research and drafting on this site. Every judgement and every recommendation is a person's.