One way of tracking the progress of AI is through milestones. Some are familiar: when, in 1997, Garry Kasparov lost at chess to IBM’s Deep Blue. More recently, in 2016, Lee Sedol lost at Go to DeepMind’s AlphaGo. Some milestones are yet to happen, such as the first AI to autonomously solve a Millennium Prize Problem. But arguably, this week a grim milestone may have been passed, this time in the realm of literature: the first significant literary prize won by an allegedly AI-written short story.
The story, “The Serpent in the Grove”, was published on Granta’s website as a regional winner of the Commonwealth Short Story Prize; it beat out 7,806 rivals for the honor, and was selected by a panel of judges chaired by the novelist Louise Doughty. I came across it by chance, on X, and started reading.
By now, I have read thousands of words of AI-written prose due to heavy use of various AI systems. The rhythms of AI writing were instantly familiar, but I was shocked to see them in a high-quality literary magazine such as Granta. Within a few sentences, I was certain I was reading GPT. So I posted about it on X, where I often post about AI milestones to my audience of mostly technical people. To my surprise, this set off a global literary scandal in the process, and one which – in my view – points to a new phase in human history. Artificial intelligence can generate text which appears to be human generated. The field of literature will need to adapt accordingly.
How could I tell? There are clear stylistic signatures. There’s the now-classic “Not X, not Y, but Z” or “It’s not X, it’s Y”. This occurs in the second sentence (“Not the bees’ neat industry or the clean rasp of cutlass on vine, but a belly sound”), and then throughout. Then there’s GPT’s enthusiasm for conjuring metaphors and similes when you ask it to generate creative fiction – but unfortunately metaphors and similes that do not really make sense: “The girl smiled like sunrise over a sink”; “she had the kind of walking that made benches become men”; “silence in a village is smoke; it sneaks from something burning”. All of these have the appearance of sense but become mysterious and odd the more you think about them. And then there’s the humming. In March 2025, OpenAI CEO Sam Altman tweeted a short story to tout GPT’s fiction-writing capabilities. In the very first paragraph you can see, “Already, you can hear the constraints humming like a server farm at midnight – anonymous, regimented, powered by someone else’s need.” It’s safe to say that “The Serpent in the Grove” hums along too. These signs – and the conclusions of AI detection software Pangram, which declares the text “100% AI written” – seem to overcome any reasonable doubt.
If this is so, it is worth acknowledging that it broke the rules and must be addressed head on, for the simple reason that this is a contest, for writers, in which the entrants are asked to affirm that their stories were not written with AI assistance. Apart from the more philosophical issues about AI assistance in writing, this should be taken as a potential case of cheating. Given this, I found the Commonwealth Prize’s response to be wanting: “When they submit stories to the Prize, writers accept our entry rules and guidelines. These include confirming that their submission is their own original work. All shortlisted writers have personally stated that no AI was used and, upon further consultation, the Foundation has confirmed this.”
This is not good enough, and it is not clear how the Foundation could have “confirmed” that no AI was used. At a minimum, prizes should use software such as Pangram as a simple check on their candidates, and have human experts investigate for a second opinion if a story is flagged as AI-written. This is an easy check to do and would go a long way to solving the issue; indeed, Pangram has analysed the past winners of this prize and flagged two other winning stories from 2025 as being potentially AI-generated as well, pointing to this becoming a recurring issue as models get better and better. Simply burying our heads in the sand about this issue does not meet the moment.
As we try and forecast how AI is going to affect various fields, it helps to look at those which have had to reckon with AI for years, and chess is one of the earliest. Deep Blue’s victory over Kasparov occurred in 1997. Chess had its own high-profile cheating scandal recently, when Magnus Carlsen tacitly accused Hans Niemann of cheating over the board after losing a game to him in 2022. As many people pointed out, it is impossible to prove that a chess player has cheated simply by inspecting their moves, since they could always have just had a good day; part of Carlsen’s reasoning was simply that Niemann did not seem stressed or concentrating at the board in what was a complex and hard-fought endgame, reasoning which is fundamentally subjective and cannot constitute any kind of proof. Despite all this, chess organisers do not throw up their hands and ask players to be honest; they have installed and tightened rigorous anti-cheating measures at live tournaments, so that chess tournaments can continue without being ruined by the spectre of possible cheating.
There are other lessons we can take from chess, too. A “doomer” perspective would have concluded that, with AI exceeding human ability in chess, chess was finished as a game; and yet, chess experienced an extreme boom in popularity after 2020, and participation in chess as a human activity is higher than ever. This example alone proves that, just because AI exists in a given field, one need not be overly pessimistic about that field’s prospects. Chess players were initially demoralised by Deep Blue, but the field as a whole resisted demoralisation. Every field will have its version of this, and as more milestones are passed, AI panic will come for us all.
A simple distinction we can make is between activities that are intrinsically human, and activities which are more instrumental in nature. In the former category we might put things such as art, religion, love, gardening, taking walks in sunshine, eating ice cream; the latter might include activities such as programming, manufacturing, making Powerpoint presentations, writing legal documents. The latter is done as a grim necessity; we do the former because we enjoy it. My belief is that literature is an activity of the intrinsically human type. AI will gradually absorb the economically productive types of activities: indeed, over 90 per cent of the code written for my software company is now written by AIs, which was not true even a year ago. I expect that much white collar work will follow the same path. But these human, relational activities are a more complicated story, like chess, and they may even grow more important as AI absorbs a growing share of economically productive work over the next few years. We will need literature, religion, art, and love more than ever. In that sense, I do not see a real “threat” to literature from AI: in fact, it could become more popular than ever.
There is also the question of merit. As many online have pointed out, the story in this case is not particularly good. I like to joke with my technical friends that we now have rigorous proof that the humanities are harder than STEM: writing a good short story or a poem is much harder than writing code or proving mathematical theorems, since we have AIs that can perform the latter very well but we do not yet have a clear example of a wonderful AI-written story or poem. So literature has not yet hit its “Deep Blue” moment. There is no AI Ulysses or Divine Comedy. Part of the reason for this is technical; but part of it is more fundamental.
The technical reason is simple. AI progresses fast in “friendly” domains, where there are large quantities of training data, and a clear, verifiable feedback loop. Domains like chess, Go, mathematics and programming meet both of these conditions: a maths proof is either true or false. A chess game is either won or lost. Thus, AI can train itself in a loop by generating results, checking them, and strengthening the circuits that led to success. But literature is much harder as a domain, because there are not so many literary masterpieces, and it is difficult for a programme to judge whether something is a masterpiece. Thus, it will remain much harder to produce really good, novel art using AI than it will be to surpass human ability in domains such as programming and mathematics.
The more fundamental reason is this. Literature is a highly complex expression of life, and AI does not have the deep, human experiences from which good literature ultimately springs. Each new masterpiece is in some sense a complete surprise to us; one could not have predicted Ulysses or The Golden Bowl or The Waste Land or My Last Duchess. There is a certain unaccountable, original quality to every masterpiece which makes it very challenging for AI to produce one, since the way AI is trained to generate texts works against producing such odd outliers. (I explore this in more detail in another essay on what makes art great.) Thus, although we can expect AI to become competent and perhaps even good at generating formulaic thrillers, romances, blockbuster superhero movies, and so on, a true masterpiece may be beyond its powers for quite some time; and it is precisely these masterpieces that are what literary fiction is looking for. It remains important to support human voices, therefore, in producing their writing, and part of this responsibility falls on literary prizes and magazines to ensure AI writing does not intrude.
One can make this argument in an economic fashion. All text in human history before the deep learning revolution was produced by humans. Now, however, the rate at which AI is producing text is increasing very fast. Eventually, AIs will produce more text than humans and the ratio will flip. This in fact makes human-written text more valuable, especially since it is that uncorrelated human text which goes into training AIs in the first place. As AI gets more pervasive, therefore, the value of human-produced text will only rise.
It is vital for civilisation that we continue to produce authentic culture. One form of risk for AI systems has been termed “mode collapse”; this is the idea that all AI outputs sound roughly similar. As AI becomes a greater fraction of our cultural output, this influences human output, too, to sound more like the AIs, resulting in a collapse in diversity and rich, original cultural production rooted in the deep contexts of ordinary life. If everyone ends up sounding like ChatGPT because we read so much AI-written prose, the total diversity of human written output goes down, and everyone ends up sounding like a peppy LinkedIn post. It is important, therefore, to encourage diverse human literature to continue to be produced, and indeed one hopes that human writing becomes more eccentric, wild, and experimental as time goes on. Part of the way we encourage this is through literary prizes. Maintaining their integrity is vital.
Storytelling is an old activity and, as anyone who has told a story to a young child knows, one deeply fundamental to who we are and how we learn. Many activities will become dominated by AI over time, and this may even be to the good; humans were not put on this earth to maintain Excel models. But in the arts and in literature, we must keep going. Resist demoralisation!
[Further reading: The silent coup]
Content from our partners
Related
Join the debate
Subscribe here to comment