The Caine Prize and the limits of AI detection

Earlier this month, Zimbabwean writer Simbarashe Kundizeza won the Caine Prize for African Writing for his short story "The Blue Fighter". Within days, The Atlantic reported that Pangram, an AI-text detector, had marked almost the whole story as AI-generated, apart from a quoted excerpt of another writer's poem.

Fíor is a natural language classifier that predicts whether a piece of text was more likely written by a person or by AI. That makes this an uncomfortable story to read, and one worth reading carefully.

The prize has not taken the award back. Its organisers say the story is under review and that they have a duty to examine claims like this. Kundizeza denies using AI. He told AP the story draws on his own family's life during Zimbabwe's years of economic collapse, and that the technology should not be used in a way that creates bias against people writing about their own experiences and cultures.

The journalist Karen Attiah put a sharper version of that point on Bluesky. She noted that the Zimbabwean head of the prize board had said the story resonated deeply with her. Then she asked whether it was colonial to insist on trusting "AI programs and non-Africans more than Africans reading their own stories".

What a score is

A detector does not see a writer at work. It reads the finished text and estimates how closely it matches patterns learned from human and machine writing. The output is a probability. It is a claim about the text, not about the person.

That is why Fíor reports that a piece "reads as AI" and never that it "was written by AI". The difference is not a technicality. It is the whole case.

The bias question

Kundizeza's charge of bias is not new, and there is evidence behind the general point. In 2023, Stanford researchers ran 91 essays by non-native English speakers through seven popular detectors. More than half were wrongly labelled as AI-generated. Essays by American eighth-graders were classified almost perfectly.

The researchers' explanation was simple. Writing with plainer vocabulary and more regular sentences sits closer to what language models produce. Anyone whose English does not match the writing a detector learned from is more likely to be flagged.

That study tested older tools and says nothing about this story. It does show that a high score can say as much about what a tool was trained on as about who wrote the text.

The case for checking

There is a real argument on the other side, and it deserves more than a line. Prizes are open to anyone, and some entries will be machine-written. A prize that ignored every warning sign would be failing every writer who did the work themselves. A detector flag is a reason to look more closely, and a human review is exactly that. On that reading, the system is working.

The trouble is timing. The review is still going on. The score has already been published, and Kundizeza's name now sits next to the word AI in headlines around the world. A probability, printed in a major magazine, works like a verdict long before anyone has reached one.

Fíor's articles score news coverage from outlets, never the work of a named individual. This story has not been run through Fíor, and it will not be.

A detector can say that a piece of writing looks like something. Only readers who know the writer, the place and the work can say what it is. The first answer arrived within days. The second is still being read.

Read more about how Fíor works.