AI can write prize-winning fiction. Now what?
The controversy surrounding Jamir Nazir's The Serpent in the Grove, winner of a Commonwealth Short Story Prize and published by Granta, shows that reading is no longer enough to establish authorship
Last week the Commonwealth Foundation announced its 2026 short story prize winners. As it has done since 2012, Granta, a well-respected UK publishing house, posted the five regional winners online. Soon after, the Caribbean winner, Jamir Nazir, was accused of using generative AI to write his prize-winning story, ‘The Serpent in the Grove’. While the competition’s judges praised the story for its ‘precise yet richly evocative’ prose, an internal review by Granta that followed the accusations of AI use concluded that the story was ‘almost certainly not produced unaided by a human’.
So, let’s have a read.
The story’s opening lines describe the grove humming at noon and then define that hum by what it is not, ruling out the labour of bees and the work of cutlass on vine before arriving at a ‘belly sound’ beneath the earth. Negation staged as the setup for revelation is among the most reliable rhetorical defaults of current language models when asked to perform a literary register; the system has learnt that contrast yields a feeling of insight, and so it manufactures contrast at every opportunity. Further on, the story offers the aphorism ‘A story is a well. It eats sound until somebody throws a rope’, a sentence with the cadence of profundity but little real meaning, a pattern that repeats throughout the piece. There are other things: the deliberately strange verb-noun pairings such as ‘eats sound’, the gnomic aphoristic compression, the reflex toward parable in place of incident, and of course, the title noun is paired and re-paired in mirrored allegory (the serpent as ‘the thing in a man that slid along stone for dark’, then ‘the thing in a woman that wrapped a vine around herself and climbed’, which is the rigid binary symmetry one expects of a model optimised for literary register).
But maybe Nazir did write The Serpent, which wouldn’t be the first piece of formulaic fiction to reach award-winning status (I can honestly see ChatGPT receiving a Booker Prize some day).
Famous AI commentator Ethan Mollick posted on Bluesky that the story functioned as a ‘Turing test of sorts’. He had run it through Pangram, an AI detection tool, which returned a 100 per cent AI-generated reading, and posted the screenshot. The framing here is telling. By calling it a Turing test, Ethan Mollick implicitly conceded the central problem that readers (and publishers) face in the age of LLMs: whether a piece of prose was written by a human can no longer be reliably determined through reading alone. If it could, well then no test would be needed.
With reading insufficient, AI detectors have become the arbiter, but of course, AI detectors are useless. Tools like Pangram do not work, and Mollick knows they do not work, but a 100% reading was sufficient grounds for him to issue the accusation to his massive following. That his suspicion happens to be one that many experts and readers (myself included) believe is correct does not absolve the method by which he proceeded.
The consequences of this reach far beyond Nazir. If AI ‘detectors’ and Bluesky accusations are allowed to function as prima facie grounds for a public denunciation of authorship, the writers harmed will be the ones whose stylistic surface overlaps with generative prose for reasons that have nothing to do with AI use, such as non-native English speakers, whose work, as much research has shown, is flagged by detectors at disproportionately high rates. Academic writing (whatever that is these days), which can often be (intentionally) dry and formulaic, full of ‘fancy’, what might be described as ‘non-human’, words, is much more likely to be flagged as AI-written than creative work. Or maybe someone just writes in a way that is very LLM-y, which I’d predict will increasingly be the case as people continue to consume content generated by these systems.
The problem we now face is this: no, we do not want so-called authors winning literary prizes for AI slop, but equally, we don’t want a situation wherein the threshold for being publicly named as a fraud is a screenshot from an AI detector and a Bluesky post.
Reading has, since the spread of literacy, operated on a tacit warrant, that the prose in front of you is evidence of a mind having deliberated over it under conditions for which the writer can be held to public account (I wrote about this recently for Psyche). We read for fluency, and underneath the fluency we read for the trace of presence: someone chose this word, paused at the comma, struck out the previous sentence, settled on a phrasing they could answer for. Generative AI produces text of an entirely different ontological character. What Roland Barthes pronounced dead in 1967 was the author as metaphysical anchor for reading, a death announced for the sake of the reader’s interpretive freedom; but the death in front of us now is operational, because the act of text generation does not constitute authorship in any sense the literary tradition has recognised. The fluency is intact but the trace is not. When the Commonwealth Foundation awards a prize on textual surface alone, with no procedure for verifying that the surface corresponds to a deliberating author, the prize has begun rewarding fluency stripped of provenance.
Razmi Farook, the Foundation’s director-general, has explained that submissions are not run through detectors because doing so would raise consent issues over the commercial use of unpublished original work, a defensible position (that and the fact that they don’t work). Granta has added a notice to all five winners and signalled it will await the Foundation’s verdict. But none of this addresses the structural problem, which is that the judging procedure was designed for a world in which submitted prose could be assumed to have an author, and unfortunately, that world is gone. The Foundation processed 7,806 entries this year, a scale at which provenance cannot be verified by anything short of a redesigned submission protocol (one which I cannot imagine). It’s not realistic to think that prizes can start demanding a provenance trace through drafts and revisions or other kinds of dated process records.
The question of whether AI can write prize-winning fiction has been settled, but to me, that’s really of little consequence (I’ve seen terrible writing celebrated long before generative AI). The real significance of The Serpent’s success is that the procedure by which literary institutions decide a prose work is praiseworthy has collapsed at the same moment the procedure by which literary publics decide a prose work is fraudulent has acquired a force it has not earned, and both procedures used to rest on the same assumption about authorship. Whatever Nazir did or did not do with a language model, the lasting consequence of the affair has been the demonstration that anyone—especially a public figure like Mollick—can mobilise a public verdict on authorship by pasting prose into a black box, and that none of his peers will object to the method.
Redesigning literary judgement for a world without assumed authorship is an extremely difficult task. We need to take our time and get it right.