Thélyson Orélien, and why AI detectors are almost never wrong

Kevin A. Bryan and Dror Shvadron are professors of strategic management at the University of Toronto’s Rotman School of Management, where they study the economics of AI and science.

The Haitian-Québécois author Thélyson Orélien won many of the prestigious prizes for French literature this year for his book C’était ça ou mourir, and was tipped as a leading candidate for the Prix Goncourt, the French equivalent of the Booker Prize. Only one problem: the book appeared to be largely written with AI.

We are both University of Toronto professors who do research on AI, including on its use in science. A new study by one of us finds that up to 30 per cent of all recent STEM PhD dissertation abstracts were written with AI assistance. Beyond academia, an article by Harvard professor Ricardo Hausmann, a Commonwealth Prize-winning short story published in Granta by Jamir Nazir, an Akutagawa Prize-winning novel by Rie Kudan, and a book review in the New York Times were all at least partly written with AI. AI use in creative tasks, whether admitted to or not, is increasingly common.

Canadian publisher Boréal has broken ties with Mr. Orélien and suggested it had evidence of AI use in his writing. Mr. Orélien maintains this is wrong. He called in an insult, insisting not one sentence in his book was written by a machine. Unfortunately for these writers, it is generally very straightforward to identify AI-generated text.

Author Thélyson Orélien denies using AI for novel, apologizes for past plagiarism

Models like Claude and ChatGPT tend to produce distinct patterns, in the same way that iambic pentameter, certain bawdy puns and slightly-off descriptions of continental Europe are fingerprints of Shakespeare. Patterns like “lists of three” – a word, an idea, and a thought – or negative parallelism – it’s not this, but that – appear much more frequently in AI-generated than in human-generated text. There is nothing wrong with these patterns, and many great writers use them, but their frequency in AI text is striking.

We can do better than trusting our gut, however. Algorithms called “classifiers” have been trained to identify text written by AI, with Pangram 4 and GPTZero among the most reliable. Consider trying to figure out whether a newly-discovered play was written by Shakespeare. Beyond pentameter and puns, there are thousands of minor stylistic habits that Shakespeare, like any writer, has. Indeed, Shakespeare himself may not have known that these patterns appeared in his work. If we had millions of Shakespeare plays, an algorithm would not have trouble spotting thousands of patterns that are more common in his work than that of other writers. That is essentially how AI detectors work.

Opinion: In the debate over AI use, Thélyson Orélien emerges as a tragic figure

This means that when a story is said to be “100 per cent AI” by the best AI detectors, it is because of a combination of thousands of patterns that becomes extraordinarily unlikely for a human. Recent studies have shown a false positive rate, where human writing is said to be AI, of one in 25,000 or lower; older studies claiming much more error or bias against non-native speakers used classifiers much less accurate than those available today.

On longer texts, such as a chapter of a book, the false positive rate is effectively zero. One of us offered a $1,000 bounty for any reader who could produce a single 1,000-word selection of French literature written prior to 2022 which would be flagged by Pangram, like the majority of Orélien’s novel, as 100 per cent AI. After 150,000 views, we did not have to pay out the money. A similar test by Radio-Canada checked a number of pre-2022 novels by writers of Haitian descent – all were scored 100 per cent human.

One may wonder whether works written by Creole authors or in particular styles are likely to be incorrectly flagged as AI, even if the classifiers are accurate more broadly. There is no evidence that this is the case with recent classifiers.

Readers and prize committees are right to feel misled about the origin of these words. But what is the role for AI in creativity? What if AI helps authors craft a clever metaphor or brainstorm new ideas? Or what if it helps an author with a creative idea for the plot of a story, but who lacks the gift for beautiful turns of phrase in a given language? Editors, authors, and readers will have to decide how to adjudicate this. If the past few months is any guide, it won’t be the last example of prize-winning prose in that set.