On 21 September, hours after Thélyson Orélien received the Prix du Roman Fnac, an anonymous account posted AI detector results claiming that his bestselling debut novel had been written almost entirely by AI. One month later, on 21 October, his Canadian publisher ended its relationship with the author. The turning point was not another detector score. It was an external analysis of the novel's working files.

We wrote about the case on LinkedIn on 25 September, when detector scores were all the public had: Back to the Witch Hunt: Why AI Detectors Are Creating a Crisis of Proof. We warned against turning statistical scores into verdicts. The subsequent events have added important evidence, but the underlying lesson remains: a score can raise a question. To understand how a text was created, we need evidence of the process as well.

The case in eight facts

• A bestseller under suspicion. C'était ça ou mourir had sold more than 83,000 copies in France by 11 October, according to Edistat figures reported by ActuaLitté. The novel had also featured in selections for France's major autumn literary prizes.
• High scores, conflicting results. Pangram gave the full text a 92% AI score, according to France 24. Lucide reported classifying 34 of 35 sections as entirely AI-written. Both tools reportedly identified texts Orélien had published before ChatGPT existed as human-written. Yet AFP's tests on the same excerpts using other tools produced results ranging from "very likely" human to 100% AI.
• The literary prizes reacted. On 25 September, the Académie Goncourt removed the novel from its first selection, citing concerns about earlier texts and converging analysis results. The Médicis jury removed it on 15 October.
• The publisher initially held back. On 14 October, Boréal co-director general Renaud Roussel told ICI Première that the available material remained inconclusive and that a detector could not provide a definitive answer.
• Reconstructing the process proved difficult. Boréal tried to work with the author to assemble a dossier documenting how the novel had been created. The publisher could not independently establish the writing history from the material then available. Orélien says he began writing in 2017 and completed a first draft in 2019. On 17 October, he said he would provide no further material.
• The working files changed the decision. Boréal gathered the available working versions and commissioned an external analysis. According to the publisher, it found traces of AI-related tags in the metadata of certain documents. On 21 October, Boréal announced that it would not reprint the novel or publish further books by Orélien.
• The extent of AI use remains unknown. Boréal said the new evidence indicated that AI had been used during the writing process, but acknowledged that it could not establish how extensively. Orélien says he used AI for research and language-related tasks, while maintaining that he wrote the novel himself.
• The publishing industry is responding in different ways. Publishers are considering more explicit disclosure of AI use. A new literary prize, the Prix Pascal, asks entrants to describe the tools and prompts used. French media have also reported plans for AI screening of manuscripts at the Madrigall publishing group from January 2027.

What an AI detector score cannot show

A score describes how a tool classifies a finished text. It does not reconstruct the process that produced it. It cannot, on its own, tell us which passages were drafted, prompted, revised or translated with AI, in what order the work developed, or how much of the author's own research and thinking went into it. A percentage should not automatically be read as the proportion of words written by AI.

That distinction matters in the Orélien case. The later metadata analysis reportedly found traces associated with AI in certain documents. Even that finding does not establish precisely what the tools did or how much they contributed. As Clubic noted, a trace linked to an AI tool does not, on its own, reveal whether it was used to correct, rephrase, research or generate passages.

The lesson is not that detector results are worthless. They can be a reason to investigate further. The problem begins when an indication is treated as a conclusion without corroborating evidence.

Why this matters even more in education

Universities face the same challenge, but with an important difference: AI use is increasingly permitted under defined conditions. The aim is not simply to establish whether a student used AI. It is to determine whether the assessment still demonstrates the intended learning and whether the submitted work reflects the student's understanding and personal contribution.

A 2026 resource commissioned by Australia's higher education regulator TEQSA calls for institutions to build evidence of learning processes and move towards process-focused assessment. For an examiner, the relevant evidence might include the development of a text over several weeks, the sources consulted, the revisions made and the decisions taken along the way. A discussion in which the student explains those decisions can add further context.

A detector alone cannot answer those questions, even when its classification is accurate. Process documentation adds a different kind of evidence. It helps educators distinguish between the use of AI as a legitimate research or writing aid and the outsourcing of work that the student was expected to perform themselves.

There is a second lesson here. Evidence assembled only after an allegation depends on what the author still has available and is willing to share. Documentation that begins when the work begins gives honest students a contemporaneous record of their effort. They do not have to reconstruct months of research and writing under suspicion.

For publishers: establish transparency before a dispute arises

Boréal said it did not want to become an investigator of its authors. The publisher also described trust between authors and publishers as fundamental. That is a reasonable position. Yet trust is easier to maintain when expectations are clear and the writing process can be documented from the outset.

In this case, the dispute escalated after prize selections, bestseller lists, international translation agreements and a second book contract were already in place. Publishers now face a practical question: how can they discuss permitted AI use openly with authors without relying entirely on declarations or resorting to detective work after publication?

A declaration is useful, but it cannot independently verify itself. A detector result is a signal, not a complete account of the writing process. A record created as the work develops gives both parties something concrete to review and discuss.

What process documentation can and cannot do

Process documentation is not a lie detector, and it does not record every thought or action outside the tools being used. Its value depends on what is recorded, how reliably the record is maintained and whether the rules were clear before work began. It cannot prove every aspect of authorship, but it can provide evidence that a finished document alone cannot supply.

The practical change is significant. Instead of arguing over a percentage, an institution or publisher has a record of the work's development and a starting point for a fair, evidence-based conversation.

---

How Mentafy approaches writing process transparency

Mentafy Trail helps make an author's personal contribution visible through regular snapshots or drafts submitted by the author, together with a research log. Authors can review their own documentation and add context or comments. The approach does not rely on keylogging, camera monitoring or screen recording.

Mentafy also offers AI text classification. We treat its results as an indication that may warrant a closer look, never as a verdict on authorship. By combining signals from the finished text with evidence of the writing and research process, institutions can make better-informed assessments of AI use, source use and personal contribution.

Whether you assess student work or publish books, see how our tools work or talk to us directly.