The literary storm around Thélyson Orélien’s C’était ça ou mourir has exposed a problem that reaches far beyond one novelist or one prize.
Orélien, a Haitian-born Canadian writer, won the 2026 Prix du Roman Fnac for a novel about migration, violence and survival.
Soon afterwards, the X account Balance ton Claude published screenshots from the AI detector Pangram and claimed that large portions of the book were machine-generated.
Orélien denied the allegation, saying he began writing the book in 2017 and completed a first draft in 2019, before generative AI tools became widely available.
The controversy became more serious on September 25, when the Académie Goncourt removed the novel from its first selection.
The academy cited separate plagiarism revelations involving earlier writings, as well as converging analyses suggesting the novel was, in all likelihood, largely produced with artificial intelligence.
That decision deserves to be taken seriously. But it also demonstrates why publishers, universities and employers urgently need a better evidentiary framework for AI-era authorship.
Plagiarism can be tested by comparing text with an earlier source. AI authorship is different: a detector generally cannot reconstruct the creative process behind the final document.
Pangram’s own benchmarks are impressive. Its latest technical material reports a false-positive rate of about 0.0041 per cent across a test set of two million human documents — roughly one false positive in 24,000 — while another breakdown reports about 0.003 per cent for books.
Those figures make the tool useful. They do not make its output forensic proof.
They are vendor-reported measurements produced under defined benchmark conditions, and Pangram itself acknowledges that performance varies according to the domain and characteristics of the text being tested.
Independent research gives further reason for caution.
A 2026 large-scale audit found that state-of-the-art AI text detectors can still produce false positives in real-world settings.
Other studies have found substantial variation among detectors and shown that human editing or paraphrasing can alter detection outcomes.
Earlier research also raised concerns about the disproportionate classification of writing by non-native English speakers.
The evidence does not mean AI detectors have no value. It means their results require context and corroboration.
The vendors themselves acknowledge these limits. Turnitin tells educators that its AI-writing indicator may misidentify both human- and AI-generated text and explicitly advises that it should not be the sole basis for adverse action.
That principle matters because a percentage can create an illusion of certainty. A result reading “100 per cent AI” sounds like a verdict, even though it represents a statistical classification of linguistic patterns.
This distinction matters especially in multilingual societies.
Writers from Africa, the Caribbean and other linguistically diverse regions may write in English or French while carrying rhythms, syntax, repetition and oral traditions shaped by other languages.
None of this proves that Orélien’s novel was or was not AI-assisted.
It does mean that institutions should be careful about turning statistical resemblance into an accusation of misconduct.
A model can identify patterns associated with machine writing; it cannot see a writer drafting, revising or arguing with an editor.
There is already a better direction. When the 2026 Commonwealth Short Story Prize faced AI allegations involving Jamir Nazir and other regional winners, the Commonwealth Foundation examined working drafts, time-stamped documents, notes and evidence of the writers’ creative processes.
After its investigation, the winning stories were upheld. That approach is slower than uploading text into a detector, but it asks the stronger question: what evidence exists about how the work was created?
Publishing should build on that principle. Authors can preserve drafts, version histories, notes, editorial correspondence and document metadata.
Publishers can require disclosure of material AI assistance while distinguishing proofreading, translation, structural advice and text generation.
Scientific publishing bodies are already emphasising disclosure and human accountability for AI-assisted work. In digital media, standards such as C2PA Content Credentials are developing tamper-evident provenance records showing how content was created and modified.
The wider lesson is not that AI detectors are useless. They can serve as screening instruments, just as fraud scores, spam filters and anomaly detectors are useful elsewhere in computing.
The mistake begins when a probabilistic signal becomes a finding of guilt without corroborating evidence and due process.
That standard should apply whether the person being judged is a novelist, a Kenyan university student, a journalist, a researcher or an employee.
Generative AI has created a genuine integrity problem for publishing, education and professional life.
The answer cannot be another black box that tells us whom to distrust. High-stakes institutions need transparent rules on acceptable AI use, evidence of the creative process and systems of verifiable provenance.
In the AI era, the strongest question is no longer whether a text looks human. It is whether we can establish, fairly and credibly, how it came to exist.
Writer is an ICT, dybersecurity and digital governance professional
