Literary Doping

I’ve been reading about two literary controversies this autumn, one involving a Haitian-Canadian novelist and the other a Norwegian crime writer, and they keep reminding me of professional cycling in the days when doping was apparently ubiquitous.

The first concerns C’était ça ou mourir, the debut novel by Thélyson Orélien. It was one of the successes of the French literary autumn until an anonymous account on X claimed that an AI detector had identified it as largely machine-written. With allegations of plagiarism in Orélien’s earlier writings also coming to light, the Académie Goncourt removed the novel from its selection.

Orélien admitted to past plagiarism but denied using AI to write the novel. His Canadian publisher, Boréal, initially stood by him, but ended its collaboration with him after an external analysis found traces of AI tags in the metadata of some of his working files. The publisher says it still doesn’t know how much AI was involved in the production of the book.

Meanwhile, the Norwegian newspaper Aftenposten has reported that six books by Jørn Lier Horst, one of Norway’s most successful crime writers, were flagged by an AI detector. Horst readily acknowledges using ChatGPT for research, for recording his ideas and for discussing possible improvements to his sentences, but insists that he has written all his books himself. A detector score isn’t proof to the contrary.

I’m less interested in establishing exactly how these two authors wrote their books than in what happens if large numbers of novelists start using AI extensively, and whether we really want to deal with this in the way cycling eventually dealt with doping.

Imagine being a professional cyclist in the late 1990s. You’re talented, you train extremely hard and you’re understandably reluctant to take a drug that might damage your health, but you’ve gradually come to realise that many of the riders who consistently beat you are taking EPO. If you don’t do the same, your career may be rather short, and if you admit to doing so, it may be shorter still.

I wonder how many novelists are starting to feel something similar. If you’ve spent two years writing a novel and you know that an AI could identify weaknesses in the plot, improve the dialogue and help you with passages you’ve never been entirely happy with, it must be rather tempting to use it. It’s even more tempting if you suspect that some of the books competing for the same prizes have been through precisely this process.

Horst says he agreed to be interviewed by Aftenposten because he wanted to help remove the taboo surrounding AI use. His openness became part of an exposé in which the newspaper’s chief literary critic accused him of betraying his readers’ trust. I can’t think of a better way of persuading other Norwegian writers to keep quiet about their own use of AI.

The Norwegian Arts Council, Kulturrådet, has meanwhile announced new rules for its literary funding and purchasing schemes, which will come into force in 2027. Books wholly or partly produced by AI will generally no longer be eligible. This matters because the state purchases large numbers of new books for public libraries, and these purchases help sustain a flourishing literary culture in Norwegian.

Under the new rules, I can ask an AI how my novel ought to end and use its ideas, provided I write the actual ending myself. If I retain the paragraph it’s written for me, however, my book will count as partly AI-produced. Horst wonders where this leaves him when he asks ChatGPT to improve the word order of a sentence. I suspect many writers who use AI will find it difficult to work out whether they’ve crossed the line, let alone document that they haven’t.

Cycling had an interesting problem of its own before a direct test for EPO became available. In 1997, the cycling authorities introduced a rule whereby male riders whose haematocrit (the proportion of their blood made up of red blood cells) exceeded 50% were excluded from competition for two weeks, officially on health grounds. For cyclists taking EPO, the limit became something to aim for rather than a reason to stop doping.

I suspect AI detectors will have a similar effect on literature. Authors who’ve never used AI will run their manuscripts through them to make sure nobody can subsequently accuse them of cheating, and those who do use AI will do precisely the same thing. Both groups may end up avoiding certain words and constructions because they’re liable to trigger a detector, irrespective of whether they improve the novel. Horst says he’s heard of people deliberately introducing spelling mistakes into their emails for this very reason.

It would be a peculiar achievement for literary policy if it ended up encouraging writers to make their prose worse.

Cycling eventually introduced the biological passport, which allows the authorities to monitor changes in an individual rider’s blood over time. Perhaps publishers will introduce a literary equivalent, requiring a complete version history alongside the finished manuscript. Boréal has already done something similar by commissioning an analysis of Orélien’s working files. I can quite easily imagine a future in which somebody who writes a novel longhand and throws away earlier drafts has difficulty establishing that it’s their own work.

All this might make sense if there were as compelling a reason to eradicate AI from literature as there was to eradicate doping from cycling. If we’d legalised EPO, we’d effectively have made taking it a condition of employment for professional cyclists, and there’s something fundamentally wrong with a profession in which people have to take drugs that may damage their health in order to keep their jobs.

I’m not convinced there’s an equivalent argument in literature. There are economic consequences, of course, and I can understand why established novelists are worried about a technology that might allow their competitors to write more books. However, it’s also worth considering who might become a novelist if the technology is available to everybody.

Imagine a nurse who’s spent twenty years working in a small Norwegian town. She knows the place and its people intimately, she has a good ear for dialogue and she’s devised a fascinating plot, but she’s never been particularly good at writing prose and can’t afford to employ somebody to help her with the manuscript. AI might make the difference between her novel getting written and it remaining an idea in her head. I’m not sure why cultural policy should favour somebody who’s an excellent prose stylist over somebody who has a wonderful story to tell but needs help telling it.

There’s a complication here, however. For some readers, part of the pleasure of a book is knowing that another human being spent months or years constructing its sentences, much as there’s pleasure in owning something made by hand. I have no difficulty with that, and there may be good reasons for subsidising people who write books entirely without AI assistance, just as we subsidise various traditional crafts that would otherwise disappear.

Personally, however, what I want is to read the best possible book. If somebody could prove that the Mona Lisa had actually been painted by a robot that Leonardo da Vinci had built, I’d find the discovery fascinating (and I’d probably be even more impressed by Leonardo), but her smile would remain as enigmatic as ever. What I feel when I look at the painting isn’t the same as what I think of the achievement of the person who produced it.

The distinction seems even clearer in mathematics. If an AI produces a valid proof of a theorem that mathematicians have been struggling with for a century, the theorem has been proved, whether or not a human being could have come up with the proof. We might still want a mathematician to explain it to us, of course, but to deny that it’s a proof at all would be rather like arguing in the nineteenth century that only cloth woven by hand was real cloth.

I wonder whether part of the difficulty in literature is that authors, publishers and literary critics inhabit a relatively small world in which they know one another and the identity of the author matters enormously. A new novel by an established writer is interesting not just as a novel but as a development in that writer’s career, and the critic reviewing it may have been following that career for thirty years. I’ve read lots of books by people about whom I know absolutely nothing, and my enjoyment hasn’t been diminished by my ignorance.

Perhaps the literary establishment is mistaking what matters to its members for what matters to its readers. If AI makes it possible for people outside that world to produce excellent novels, it may pose a much greater threat to the literary establishment than to literature itself.

There’s another interesting possibility here. An AI that knew my literary tastes and had access to the books being published might be considerably better at identifying novels I’d enjoy than a critic writing for a general readership. Perhaps our Norwegian nurse will eventually find her readers with the help of AI, too, without ever attracting the attention of an established literary critic.

The implications for smaller languages are particularly interesting. There are vast numbers of books that will never be translated into Danish or Norwegian at current translation costs, and if AI could reduce the cost of producing a good translation, it might make a great many of them available. Yet the new Norwegian rules also exclude translations produced by machine translation and subsequently reworked by a professional translator, irrespective of the amount of work the translator has put into them.

I can understand why translators are worried about their incomes, although cheaper translations could also mean that more books get translated. It would be rather ironic if a policy intended to protect Norwegian literature made reading in Norwegian less rewarding than reading in English, because so many more books were available in the latter.

There’s an important distinction here between what a literary prize is intended to reward and what the Norwegian purchasing schemes are intended to achieve. A prize for the achievement of writing a novel without AI would be perfectly reasonable. However, the purchasing schemes are intended, amongst other things, to make excellent literature available to Norwegian readers. I’d have thought that the quality of the books they get to read ought to matter rather more than whether the novelist or the translator could have produced them without AI.

I can also understand why Boréal felt unable to continue working with Orélien after finding evidence that contradicted what he’d told them. There’s a difference between defending the use of AI and defending dishonesty about its use. I’d be quite happy for books to carry a note in the colophon explaining how AI was involved in their production, although we’re unlikely to get particularly informative notes as long as being open about the use of AI puts an author’s career at risk.

In May 2007, a number of former riders from Team Telekom admitted to having used EPO, amongst them Erik Zabel and Rolf Aldag, and on 25 May Bjarne Riis admitted that he’d been doping when he won the Tour de France in 1996. Each admission made it a little easier for the next cyclist to come clean.

I suspect we’ll eventually have a similar moment in literature, when a few sufficiently prominent novelists admit to having used AI extensively for years and a great many other writers suddenly feel able to admit that they’ve been doing precisely the same thing. We may then discover that some of the books we’ve enjoyed most over the preceding decade were amongst those written with the greatest amount of AI assistance.

The cycling authorities could go back and revise the results of various races. We won’t really be able to revise the pleasure we got from reading those books.

Leave a Reply