An AI detection company says three of the five regional winners of this year's Commonwealth Short Story Prize were probably written, at least in part, by a machine. Pangram Labs published the claim this week after running every winning entry since the prize launched in 2012 through its detector.
The numbers, and who's counting
The pattern Pangram lays out in its analysis is tidy enough to be suspicious: nothing flagged before 2025, one story flagged that year, then a cluster in 2026. The 2025 overall winner, "Descend," came back at an 88% AI fraction, which the company reads as mostly machine-written with a human hand somewhere in the mix. This year, three of five winners tripped the alarm.
The story that kicked all of this off never needed a lab. Jamir Nazir's "The Serpent in the Grove," the Caribbean regional winner, set off readers days before Pangram weighed in. Wharton professor Ethan Mollick ran it through the detector and got 100%. Plenty of other people did the same and got the same number. Judges chaired by novelist Louise Doughty had picked it from 7,806 entries and praised its "precise yet richly evocative" language.
Worth a pause here. Pangram claims 99% accuracy and a small false-positive rate, and outside testing has treated it more kindly than most detectors. But "the company that sells AI detection found a lot of AI" is the sort of headline that should send you to the methodology, and a clean year-by-year curve is exactly what a vendor would want its data to look like. The percentages on the two newly flagged 2026 stories aren't something I can independently confirm.
What the prize people say
The Commonwealth Foundation is not budging. It says it doesn't use AI detectors in judging, calling the tools not infallible, and that all five shortlisted writers personally stated no AI was used. Director-general Razmi Farook said the Foundation confirmed this "upon further consultation," whatever that consultation involved. It has also opened a transparent review of its selection process, which is a curious move for an organization as sure of itself as those statements suggest.
The detail that lingers comes from Granta, which published the winning entries. Asked whether the Nazir story was machine-made, the magazine reportedly consulted a chatbot, which answered that the text was "almost certainly not produced unaided by a human." A literary journal asking ChatGPT whether something was written by ChatGPT is the closest this saga gets to a punchline.
Does any of it hold up?
The honest reading is that nobody can prove authorship either way, and that's the problem. A detector gives a probability, not a confession. A signed declaration gives a signature, not a guarantee. Readers flagged the prose as off well before anyone ran software, and "feels off" isn't something you can adjudicate a prize with.
The overall 2026 winner is due in June. The Foundation's review runs alongside it, with the disputed stories still online.
