A dispute over a novel shows why honesty about AI use matters—and why allegations still need evidence
A publisher has withdrawn its trust in a novelist over AI. It is walking away from their relationship while acknowledging that it cannot establish how extensively AI was used.
That publisher is Les Éditions du Boréal, the Canadian publishing house behind C’était ça ou mourir, a novel by the Haitian Canadian writer Thélyson Orélien. It had selected his manuscript and brought the book to readers. On October 21, it announced it would neither reprint the novel nor publish his future books.
I cannot treat AI’s presence in someone’s writing process as a moral failing.
What interests me is whether the people involved were honest about its role and whether the evidence supports the accusation. An allegation that could wreck a career deserves more than the atmosphere of certainty.
In its October 21 statement, Boréal says its efforts with Orélien to document the writing proved inconclusive. After he said he would provide no further material, it commissioned an external examination. It reports traces of AI tags in some drafts’ metadata, contradicting his repeated assurances that he had never used AI to create the novel.
That is a serious position, and the external examination deserves to be taken seriously. A publishing relationship depends on candor; an editor cannot be expected to keep defending assurances the editor believes have been contradicted. But a decision to end that relationship leaves a separate question open: did a machine write any of the novel’s prose?
Orélien disputes that premise. In a statement sent to AFP earlier in October, he denied that any sentence in the novel had been created by a machine and described using tools for research, including foreign idioms.
After the breakup, he told Libération that he had used AI in the creative process for historical research, according to AFP’s reporting.
Research assistance and the generation of finished prose are different activities. An author can use the same chatbot for either. Saying AI entered the process does not, on its own, tell a reader which activity occurred.
Orélien’s explanation deserves to be examined as an explanation, without being treated as either an automatic acquittal or a confession to something he denies.
There is uncomfortable context here, too. Orélien has acknowledged past plagiarism. That deserves scrutiny in its own right. It cannot establish who composed the sentences in this novel.
The publisher’s statement does not identify the examiner or reproduce the tags and methodology. It gives readers a conclusion without the technical material needed to assess its reach.
I would want to know which operation produced those traces and whether the examination linked them to passages in the published book. Without that explanation, I cannot tell whether the reported evidence contradicts the research account or establishes something more extensive. Neither can I responsibly dismiss it as meaningless.
NIST’s 2024 review of synthetic-content transparency explains a useful limit: metadata alone is neither reliably attributable nor tamper-evident, and even a digital signature cannot establish that the information recorded is true.
That general warning does not invalidate Boréal’s findings. It does put an obligation on anyone interpreting such evidence to explain what it actually establishes. The word “metadata” should not become a more sophisticated way of asking everyone to stop asking questions. The technical vocabulary ought to make the evidence clearer. Otherwise, we have merely upgraded the stationery on which uncertainty is delivered.
My concern is whether the process used matched what was agreed. A publisher may set stricter conditions for a novel than I would set for a newsletter. Both parties need to understand those conditions before a manuscript becomes a public test of someone’s honesty. A contract should make clear what assistance is permitted and what must be disclosed. If a house prohibits generated prose, the author should know what that prohibition covers, including whether it extends to rewriting a passage the author originally composed.
The Authors Guild’s October 13 guidance offers a practical starting point. The U.S. writers’ organization recommends explicit written AI policies and opposes using detector results alone to accuse an author or cancel a contract. It also opposes requiring writers to use technology that logs their writing to prove authorship. Those are recommendations, rather than a ruling on this case.
I do not want a future in which a novelist must produce a surveillance record to earn permission to be believed. Keeping drafts can be sensible. Making continuous documentation a condition of literary legitimacy would ask writers to become archivists of their own innocence. The time spent proving that a sentence belongs to you is time unavailable for writing the next one.
There will still be disputes. Clear agreements cannot make people honest or turn incomplete evidence into a complete history. They can, however, give an author something specific to answer and a publisher something specific to enforce. That seems a better use of a contract than discovering what everyone meant by “AI use” after the relationship has collapsed.
I keep coming back to the readers in that first admission. How a book was made can change what its words mean to someone who loved it. Readers deserve an honest account, including a clear statement of what remains unknown.
Orélien, meanwhile, has to live with the accusation beyond the next news cycle. Protecting literature should require enough care to distinguish a publisher’s broken trust from a settled account of authorship. I want clarity about what was promised, and care in judging whether that promise was broken.