Pangram, a three-year-old AI-detection startup, has become the referee of publishing's AI scandals
What's going on? Pangram, a three-year-old AI-detection startup, has become the de facto referee of publishing's AI scandals. The latest flashpoint is the Caine Prize for African Writing: on 6 Oct Zimbabwean author Simbarashe Kundizeza won the £10,000 prize for his short story "The Blue Fighter", and within a week The Atlantic reported that "the AI detector Pangram flags almost the entire story as AI-generated." Kundizeza denies it: "I did not use generative AI to write it. It is entirely my own work." The prize says it will review his draft history using a "human centered approach", and its chair Ellah Wakatama told the BBC she wants "proof that does not come from a machine."
A year of literary AI fights: CNN counts a run of cases in which Pangram scores were cited. Hachette cancelled the US release of Mia Ballard's "Shy Girl" in March. The Commonwealth Short Story Prize winner was accused in May. Researchers put the bestseller "Daggermouth" at 60% AI-written in July, the same month a debut thriller sold for around $2m was withdrawn by its agents. In September, the Goncourt jury dropped Thélyson Orélien from its longlist after an accusation built on a Pangram test. Substack has built Pangram into its platform since the summer, so readers can check whether a newsletter post looks "human-made".
Why it matters: A US startup with a small team and about $13m raised, most recently a $9m round led by Menlo Ventures in July, now effectively sets the evidence standard for authorship disputes in an industry it has no roots in. CEO Max SperoDealroom has a profile for this one. Try Dealroom →, a former Google and Nuro engineer, told CNN: "I'm an avid reader. I've never been a writer." Pangram pitches itself as a defence for magazines and prizes flooded with machine-written submissions. Critics, including authors' groups and academics quoted by CNN, Wired and the BBC, say the accusations have fallen mostly on non-white writers, and that a probability score can create doubt it can never fully settle. Institutions are drawing different lines. The Commonwealth Foundation cleared its winners without using detectors but has since asked Pangram for input on next year's process. Granta cut ties with the prize after running the story through Claude.
What's confirmed and what isn't: Pangram's own figures put its false-positive rate at 1 in 10,000 (0.0041% in its latest technical report), and some university studies support low error rates. A 2025 University of Maryland study found a 2% false-positive rate, and an independent French test by Next found 1 wrong flag in 460 human articles. No allegation cited here has been proven, and the accused authors named above have denied using AI or have not admitted to it.
Sources: CNN / BBC / Wired / The Bookseller / Declic Media
