How do we spot AI slop? Lessons from the Granta controversy

One would think that an editor at a publishing or media house would be able to tell the difference between human-written and AI-generated text with high precision, the way a seasoned jeweller appraises a stone. One look through the loupe, a knowing tilt of the head, and the fake is exposed.

The past few weeks have not been very kind to that assumption. On X, a user shared the very popular blog titled ‘The quiet grief of adult friendship’, noting that it was one of the most beautiful articles she had read in a while, the kind that hits hard.

A few hours later, Max Spero, CEO of Pangram Labs (one of the leading AI-detection companies), shared a screenshot suggesting this profound article was 100% AI-generated, which the author later denied. And this was not a singular incident.

Within the same week, an Indian author drew AI accusations, and then came the talk of the town: a short story by Jamir Nazir that appeared in Granta and won the Commonwealth Short Story Prize, with lines such as “The girl smiled like sunrise over a sink”, leading many readers to cry AI slop. According to Pangram, the text was authored by AI, along with two other stories.

At the crux of every one of these unmaskings is a simple fact: it was first caught by a trained eye and then corroborated by an algorithm. The mild irony is just that the trained eye rarely belongs to a veteran gatekeeper. Instead, it belongs to someone chronically online who has read too much ChatGPT output.

Coincidentally, a colleague Jenna Russell, now a PhD student at the University of Maryland, wrote a paper on how people who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text.

The problem, then, is epistemic. The editors and gatekeepers at prestigious literary and journalistic venues neither like nor use AI. This, of course, means they may often fail to spot AI-generated prose because they haven’t read enough of it to recognise its stylistic tics or idiosyncrasies.

The obvious safeguard is software built exactly for this job. Granta’s own publisher, Sigrid Rausing, half grasped this, noting the irony that AI is the best tool we have for catching AI. Her team then demonstrated the “half” by running the story through Claude (Anthropic’s general-purpose LLM), which further muddied the waters by concluding that it was “almost certainly not produced unaided by a human”.

The wrong AI detector, however, is only part of the problem. Many folks have talked themselves into a deeper blind spot. The prevailing wisdom, absorbed largely from a long-running and contentious fight elsewhere, is that the tools simply don’t work. Free, low-quality detectors really are easy to fool, as the internet keeps proving: feed one a passage of Frankenstein and it brands Mary Shelley a machine.

Another common objection is how AI detectors can be biased towards non-native speakers. Some have put it less gently. The Sri Lankan speculative fiction writer Yudhanjaya Wijeratne attacked the Commonwealth Foundation’s director-general Razmi Farook for claiming AI detection tools are trained on text from a “dominant culture,” calling it a shallow orientalist excuse and an insult to writing.

As a computer scientist, I want to offer my two cents. Blanket statements such as “AI detectors don’t work” are usually not backed by empirical evidence but instead rely on isolated anecdotes. Independent research from UChicago Booth found that some commercial detectors achieve near-zero false-positive rates that hold up even against ‘humanizer’ tools.

More recently, with support from Pangram Labs, the writer Vauhini Vara analysed past Commonwealth Short Story Prize winners and found no evidence of AI use before 2025. This fits the timeline, since the generative-AI boom only began in 2023 and adoption takes time.

It isn’t impossible to evade AI detectors. Research has shown that by significant editing or fine tuning, it is possible to pass off AI-generated text as human. But that effort proves the opposite point. If you have to heavily edit or fine-tune AI output to slip it past a detector, you’ve done the real work that good writing takes anyway.

A detector doesn’t need to be unbeatable but rather just make faking it more trouble than writing the thing. Using reliable tools to screen for LLM-generated text is critical, as AI slop floods submissions at a scale we as humans aren’t equipped to handle.

If we are to preserve human creativity and originality of expression, fears about potentially making false accusations must give way to enlightened conversations about how and why we’re using generative AI, about when it’s OK and when it’s not.

Editors, prize juries and publishers should be drawing the line between the legitimate use of AI to assist research and revision and the dubious use of it to generate prose that is then passed off as someone’s own work.

In her essay ‘Why I Write’, Joan Didion said she wrote entirely to find out what she was thinking. Outsource that and you’ve saved time, but skipped the only part that mattered. A detector cannot save culture on its own. But it can buy us the time to remember what we are protecting, and why.

END OF ARTICLE