An author accused of using AI: what the Goncourt case shows about detectors and drafts

Last checked 12 November 2026. This is an explainer based on public reporting, not legal advice, and TimeBinder takes no position on whether the author used AI. TimeBinder does not detect AI, and a dated record is not proof of authorship, ownership or originality.

What has been reported

The facts below come from the BBC, the Guardian, France 24 (with AFP), the Académie Goncourt’s own statement and French press reports. They are still developing, and we have kept to what those sources say.

• 21–22 September 2026 (reports differ on the day). A little-known group calling itself “Balance ton Claude” said on X that Thélyson Orélien’s debut novel C’était ça ou mourir was “almost totally” written by AI. The group said it had tested passages with the detector Pangram. The BBC reported that Orélien denied the claim at once and that his French publisher, Grasset, defended him. The book had just won the Prix du Roman Fnac and was on the selection lists for the Goncourt and the Renaudot.
• 25 September 2026. The Académie Goncourt said it was removing the title from its first list. Its statement cited plagiarism in the author’s earlier writing that had just come to light, and the converging results of several analyses by researchers and journalists, which it said showed the novel was most likely largely produced by AI. It said the aim was to protect the integrity of the prize.
• 13 October 2026. In a statement to AFP, Orélien said “not a single sentence” had been created by a machine. He said he had used tools for research, particularly for foreign idioms, and that he began the book in late 2017 and finished a first draft of more than 300 pages in 2019. AFP reported that he had not made early drafts available, and that his Canadian publisher, Boréal, said it could not confirm or deny the allegations. He also admitted “unacceptable mistakes” in connection with separate plagiarism accusations about his earlier work, and apologised.
• 21 October 2026. According to French press reports, Boréal ended its collaboration with the author, saying its own expert analyses indicated AI had been used in the writing process, while also saying it still did not know to what extent. Orélien now says any AI use was for research only. BFMTV reported on 1 November that his French publisher, Grasset, had declined to comment on Boréal’s decision and that the book was still on sale. As of that report, no court had ruled on the questions the case raises.

Plagiarism allegations about his earlier texts, which he has said involved “unacceptable mistakes” without naming specific cases, are a separate matter from the AI question. This article is only about the AI allegation and what writers can learn from the way it unfolded.

What the detector did and did not show

The allegation began with a detector result. AFP later ran excerpts through other tools and reported results ranging from “very likely” human-written to “100 percent” AI-generated, while on the excerpt AFP tested, Pangram (which AFP described as considered one of the most reliable) indicated a 100 percent probability of AI use. The BBC noted critics’ points that detectors look for patterns, not authorship, that those patterns also occur in human writing, and that some tools may be less accurate on dialect or non-standard language.

Detector makers say something similar about their own products. GPTZero’s official FAQ, for example, says results should not be used to punish students, that edge cases exist in both directions, and that its classifier is intended to open a conversation rather than settle one. It recommends asking for artefacts of the writing process, such as drafts, revision histories and notes.

So a score can be the start of a question. It is not an answer, and it is not a record of how a text was made. For the longer explanation, see why working drafts and timestamps matter more than AI detectors right now and whether file metadata can show AI use.

What the case suggests for every author

Without judging this author, the public record shows a gap that many writers share. When a claim arrives, the strongest reply is evidence that already existed before the claim.

• Contemporaneous beats retrospective. A first draft from 2019 is a strong claim only if there is something dated from 2019 that others can check. A folder re-saved last week, an email you now forward, or a file whose “date modified” you can edit does not do that. See what happens to file dates when you copy, zip or download.
• Several dated points beat one. A record of an outline, a first chapter, a full draft and the submission file shows a manuscript growing. One file dated the day before submission shows much less.
• Say what you used, in your own words, early. Authors increasingly use AI tools for research, translation, idioms or editing. An honest, dated note of what you used and for what is easier to defend than a later clarification. Publishers and prizes now ask; see when a prize winner discloses AI use.
• Know what a dated record cannot do. It cannot show that no AI tool was ever used, who typed the words, or that the text is original. It shows that an exact file existed no later than a recorded time and still matches.

A simple routine you can start today

1. Keep a dated milestone export. At each stage (outline, first chapter, full first draft, each revision, the file you submit) export a PDF or .docx and give it a clear name. Do not overwrite earlier exports.
2. Date the exports. Make a fingerprint of each file and have it timestamped. With TimeBinder, the fingerprint is made in your browser and the file stays on your device. The free SHA-256 checker shows what a fingerprint is, and how to timestamp a file walks through the steps.
3. Keep a plain AI-use note. One dated line per use: tool, purpose, which part of the work. If you use none, you do not need to write anything, but if you use some, write it at the time.
4. Keep a project backup. The full working project, not just the export, with its own history. See how to prove a Word document or Scrivener project existed before a date.
5. If you are accused, reply calmly and offer the records. Our calm reply template and the author-accused guide cover the wording.

For the wider picture of how these cases tend to unfold, read the AI accusation casebook and the guide to proving a manuscript existed before an AI accusation.

Tools that can help

Nothing here is required, and none of it proves who wrote the text, and none of it detects AI. Use what you already have if it does the job.

• Microsoft Word, Google Docs or LibreOffice, with exports kept at each milestone. Google Docs and Word both keep version history; that history lives inside the service, so export and date a copy as well. Not affiliate links.
• Scrivener (affiliate link) — a long-form writing app for macOS, Windows and iOS with a binder, named snapshots and project backups. Official guidance says snapshots live inside the project, that Compile uses the current Draft folder rather than a snapshot, and that File > Back Up > Back Up To… is the way to make a copy you can keep or share. It is a writing tool, not a dated record: back up and date an export at each milestone.
• A plain text file or notebook for your AI-use note, if you use any AI tools. Not an affiliate product.

Sources checked

• BBC News, “Thélyson Orélien: French uproar over author accused of using AI,” 23 September 2026.
• The Guardian, “Novelist accused of using AI to write book removed from French prize list,” 25 September 2026.
• Académie Goncourt, Communiqué, 25 September 2026.
• France 24 with AFP, “Novelist Thélyson Orélien rejects AI claims after bestseller pulled from French prize list,” 13 October 2026.
• BFMTV, “Affaire Thélyson Orélien…”, 1 November 2026 (Boréal’s 21 October statement; no court ruling).
• Publishing Perspectives, “In France, the Publishing Industry Endures an Autumn of Discontent,” November 2026.
• GPTZero, official FAQ (limitations; process artefacts), checked 12 November 2026. Scrivener limits as published by Literature & Latte.

Draft trails for authors are grouped on the writers and authors timestamping hub. For a different authors’ dispute — the US Anthropic settlement, whose claim deadline has passed — see what the payments mean, and what a dated manuscript still does and does not show. A dated manuscript cannot prove a book was in a training set.

Start a dated draft trail before anyone asks for one

Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.

Related: Office for Students AI assessment guidance: what students should keep.

Related: Caine Prize clears its winner of AI allegations: what writers should keep.