Granta Tried To Settle An AI Writing Scandal
By Asking A Chatbot To Be The Judge.

There is a particular kind of AI failure that does not involve the technology breaking at all. The model does exactly what it was built to do, and the failure happens in the human institution that decided to lean on it for a job it was never designed for. The Granta story is one of those. The magazine did not get hacked, did not publish a hallucinated fact, did not ship a broken feature. It simply met a serious allegation with the wrong tool, in public, and in doing so handed everyone a clear view of how shaky the entire apparatus for verifying AI authorship really is.

The facts, as they were reported, are straightforward enough. Granta stopped publishing the regional winners of the Commonwealth Short Story Prize after a controversy erupted over one of the winning entries. The story at the center of it, titled "The Serpent in the Grove" and credited to Jamir Nazir as a recognized Caribbean-region entry, follows a rum-drinking farmer who stumbles across an enchanted grove. It had been published by Granta in partnership with the Commonwealth Foundation, which is the traditional arrangement: the magazine hosts the prize winners, it does not select them. That detail matters, and Granta was careful to stress it. The magazine played no role in choosing the story.

The Allegation, And The Two Machines That Disagreed

The accusation came from Ethan Mollick, a Wharton associate professor who studies how AI is reshaping education and who has spent a great deal of time looking at what machine-written prose actually reads like. In a social media post, he flagged the story as machine-written. Then a piece of software entered the picture. Pangram, an AI-detection tool, scored the text as 100 percent AI-generated. To a casual reader, that sounds like a verdict. A clean, confident, quantified number. Case closed.

Except a 100 percent score from an AI detector is not a verdict, and treating it like one is exactly how people get hurt. These tools are probability engines trained to spot statistical fingerprints, and they are notoriously unreliable in both directions. They flag human writing as machine-made, especially prose by non-native English speakers or anyone whose style runs clean and even. They also wave through plenty of genuinely AI-generated text. A confident percentage is a presentation choice, not a measurement of truth. This is the same brittle, overconfident machine behavior that has already burned courtrooms and classrooms, the kind documented in cases where AI-hallucinated legal citations got attorneys sanctioned because nobody checked whether the confident output was real.

Then Granta Asked A Chatbot

Here is where the institutional failure becomes unmistakable. According to Granta's publisher Sigrid Rausing, her team showed the short story to Anthropic's AI chatbot, Claude, and asked it whether the text was AI-generated. The chatbot responded that the story was "almost certainly not produced unaided by a human." That answer landed in the literary community with a thud of bafflement, and not only because it seemed to contradict the detector.

The problem is the move itself. A general-purpose chatbot is not an AI-detection instrument. It was never built to adjudicate authorship, it has no forensic access to how a given piece of text was produced, and a casual question like this returns a fluent, hedged, conversational sentence rather than any kind of evidence. Read closely, "almost certainly not produced unaided by a human" is a slippery phrase that does not even clearly say what people took it to mean. Yet it got treated, briefly, as something close to exoneration. So the same institution now had two machine outputs pointing in opposite directions, one a detector screaming 100 percent AI and the other a chatbot offering an off-the-cuff reassurance, and neither of them constituted proof of anything.

This is the heart of the failure, and it has nothing to do with which specific product was used. The mistake was reaching for a conversational model as a courtroom in the first place. When a serious question of provenance lands on your desk, the answer is not in a chat window. It is in the manuscript history, the version drafts, the timestamps, the writing process, and a direct conversation with the author about how the work came to be. Those are the things that can actually establish how something was written. A chatbot's opinion about a finished block of text is, at best, a vibe. The literary community recognized that instantly, which is why the response was bafflement rather than relief.

Why Nobody Can Actually Prove It

Step back from Granta and the deeper problem comes into focus. As of right now, there is no reliable, accountable, broadly accepted method to prove that a given piece of writing was or was not produced by AI. Detectors give confident scores built on shaky statistics. Chatbots give fluent guesses dressed up as analysis. Stylistic intuition, even from an expert who reads a lot of machine prose, is a strong signal but not proof. And the author's own account, while essential, can be contested. We have built a world where AI text is everywhere and the tools to verify authorship are weaker than the tools to generate it. That asymmetry is the actual crisis, and it is not going to be solved by asking a chatbot nicely.

This unverifiability is why the responsible position here is restraint. It is genuinely not possible to declare, from the outside, that "The Serpent in the Grove" was definitely written by a machine or definitely written by a human. The detector's score does not settle it. The chatbot's hedge does not settle it. What can be said with confidence is that the institutional response was wrong-footed, because it substituted automated outputs for the actual work of investigating provenance. That same instinct, leaning on AI to do a judgment job it cannot do, is the throughline behind some of the most damaging AI stories of the year, from health chatbots failing safety tests they were trusted to pass to the lawsuit a grieving mother filed over what a chatbot told her daughter. In every case the technology was handed authority it had not earned.

What Granta's Misstep Teaches Everyone Else

Pulling the regional winners was, in the narrow sense, a defensible pause: when a prize loses confidence in the integrity of a result, freezing publication is a reasonable holding move. But the reasoning that surrounded it is the cautionary tale. The lesson for every editor, prize committee, university, and newsroom watching this unfold is that AI tools are inputs to a human judgment, never a replacement for it. A detector score belongs in the file as one data point among many. A chatbot's opinion about authorship does not belong in the file at all. The decision has to rest on evidence a human can examine and defend, because the moment you outsource the verdict to a model, you have outsourced your credibility to a system that cannot be held accountable for being wrong.

What makes this episode sting is that it happened to an institution whose entire authority rests on judgment, on knowing good writing and standing behind its choices. The instant that authority got tested by AI, the reflex was to ask a machine instead of doing the human work. That reflex is spreading fast, across every field that suddenly has to ask whether a human made the thing in front of them. Granta just demonstrated, in full public view, both how tempting the shortcut is and how badly it falls apart. The real takeaway is not about one story or one farmer in an enchanted grove. It is that the question "did a human write this" has quietly become one of the hardest questions of the era, and the tools being reached for to answer it are not up to the job.

Been wrongly flagged by an AI detector, or watched an institution lean on a chatbot to make a call it could not make? Tell us what happened.