Publishers backed a prize-listed novelist while the AI evidence stayed private

01

WHAT ACTUALLY CHANGED

On September 21, the previously little-known X account Balance ton Claude accused Thélyson Orélien of using AI to write most of C’était ça ou mourir.

The anonymous account said it submitted selected passages to Pangram. It presented the detector result as evidence that the novel was written almost entirely by AI.

AFP and Reuters said they could not verify the identity behind the account. The account told Reuters that anonymity protected its independence and freedom to speak.

The accusation arrived on the day Orélien received the Prix du Roman Fnac for the book, turning a literary award into an immediate public authorship dispute.

C’était ça ou mourir is Orélien’s first novel. It follows a Haitian teacher fleeing gang violence through 12 countries toward Canada.

The novel appears on major French-language prize lists, including the Goncourt, Médicis and Femina selections.

Orélien denied using AI to write the novel. He said he has long written and created without needing a machine and is assembling supporting material with his publishers in France and Quebec.

Éditions du Boréal said it took the allegations seriously while maintaining confidence in Orélien’s work. Grasset condemned the claims and reaffirmed its confidence in him.

Reuters reported on September 23 that the publishers continued to stand behind the author. Their support is relevant institutional testimony, but it is not a substitute for independently reviewed authorship evidence.

AFP tested the accuser’s selected excerpts using InVID-WeVerify and seven other detectors. InVID-WeVerify assessed them as very likely human-written, while the other systems produced sharply divergent results.

Among AFP’s seven additional tests, one system returned a 100 percent AI result, two reached the opposite conclusion, two returned mixed probabilities and two disagreed between the tested passages.

The AFP comparison did not establish who wrote the novel. It showed that changing the detector can change the conclusion even when the sampled text stays the same.

Pangram says its detector is 99.7 percent accurate. Reuters presented that as the company’s claim, not as an independent validation of this novel, edited literary French or Orélien’s complete manuscript.

Philippe Claudel, president of the Académie Goncourt, said no decision had been taken about removing the book. He said the jury would consider the issue if indisputable information were available before its October 6 meeting.

02

WHY THIS MATTERS

A detector result is a probability produced by a model. It is not direct observation of the writing process and does not identify a human author, collaborator, editor or tool history.

Selected excerpts can create selection bias. An accuser can choose the passages that produce the strongest score while leaving out pages that score differently.

A novel is not a clean laboratory sample. Drafts are rewritten, editors suggest changes, sentences are compressed, idioms travel between languages and a writer may deliberately repeat structures for rhythm.

French shaped by Haitian and Caribbean traditions may not resemble the writing used to train or validate a detector. No public evidence shows that Pangram’s headline accuracy transfers cleanly to this literary context.

A 2023 Stanford-led study found serious false-positive problems for several detectors when evaluating non-native English writing. That study did not test Pangram, French or Orélien’s prose, so it cannot decide this case. It does show why performance claims must be tested on the population and language actually under judgment.

The accuser’s anonymity does not make the claim false. It does make verification, accountability, conflicts and the selection of evidence harder to examine.

Publisher support does not make the denial true. Publishers have commercial and reputational interests in a successful book. Their statements should be weighed beside evidence, not mistaken for it.

The author says he has faced racist attacks since arriving in France. That experience matters because public accusations do not land on every writer with the same force. It does not by itself prove that this specific accusation was racially motivated or technically wrong.

The novel tells a Haitian migration story and arrived from a debut author whose work uses a literary tradition that some French readers may find unfamiliar. A fair process must avoid treating stylistic unfamiliarity as machine evidence.

Prize organizations are improvising after publication. If AI-use rules are vague or unwritten, a jury can end up inventing standards in the middle of a scandal and applying them selectively.

Authorship is not a binary switch once tools enter the room. Spellcheck, translation aids, research search, transcription, developmental editing and generative rewriting perform different roles and should not be collapsed into one accusation.

Demanding every private notebook and draft can punish the accused again. A neutral reviewer can inspect provenance confidentially, verify dates and revision patterns and publish an attestation without releasing the manuscript’s entire workshop.

Writers also deserve an appeal. A machine score that can alter a career, contract or prize position should come with the model version, sample, threshold, validation data, limitations and a route to challenge the result.

The publishing industry needs a standard before the next accusation, because anonymous detector screenshots are cheap, viral and capable of moving faster than any careful investigation.

03

WHERE IT COULD HELP

• Write prize and publisher rules that distinguish brainstorming, research, translation, copyediting, accessibility tools, sentence rewriting and generation of substantial creative text.
• State those rules before submissions open, including what must be disclosed and which uses can lead to disqualification.
• Never impose a penalty from one detector score. Treat the result as a lead that may justify a careful review, not as proof.
• Require the complete tested sample, detector name, model version, settings, threshold and date so another reviewer can reproduce the result.
• Test the same text with multiple systems and include known human work from the author, comparable literary French and edited control passages.
• Ask whether the detector was validated on the language, genre, length, publication process and linguistic community involved in the decision.
• Use a confidential provenance review of dated drafts, version histories, editor comments, email attachments, notes, backups and early submissions.
• Let a neutral expert produce a narrow attestation that describes what was verified without publishing every private draft or source.
• Separate evidence of AI assistance from evidence of prohibited AI assistance. A disclosed research or correction tool may comply with rules even when a detector notices stylistic change.
• Give the author written notice, access to the evidence, time to respond, an independent reviewer and a documented appeal path.
• Preserve evidence before controversy changes it. Publishers can hash submitted drafts and editorial files so later reviewers can confirm dates and integrity without exposing the files publicly.
• Publish a short decision record explaining the rule, evidence considered, uncertainty and reason for any prize action.
• Audit detector error rates across French, Quebecois French, Haitian and Caribbean writing before adopting one as an institutional screen.
• Track who is accused and who is cleared. A system that disproportionately burdens writers from particular linguistic or racial backgrounds needs correction even when no individual motive can be proven.
• Keep the literary question separate from the forensic one. Readers can dislike a sentence or prize decision without pretending aesthetic judgment proves machine authorship.