Caine Prize winner Kundizesa denies using AI to write his story

Why the story was questioned

Wara ran the story through Pangram, an AI-detection program that also uses AI. She said it identified only the epigraph, from an unpublished poem by South African poet Vangile Gantsho, as human-written.

Wara also pointed to patterns she associated with AI-generated writing: “not X, but Y” constructions, abundant metaphors and familiar aphorisms. She called the story unsuccessful, sentimental and full of clichés.

Journalist Vauhini Vara said she had run Kundizeza’s story through the AI-detection software Pangram, which itself uses AI, and it had deemed the story to be almost entirely AI generated. Photograph: Andrew Altschul

Source: theguardian.com

Kundizesa said in a post on X that he wrote “The Blue Jockey” himself and did not use generative AI. He said he would continue working with the Caine Prize team, share the award with the people who helped create the story and keep writing.

The allegation follows other recent disputes over literary awards:

What the prize can establish

The Caine Prize organizers said they are investigating the allegations against Kundizesa. They said they have a duty to look into such claims and intend to do so with human involvement, while supporting and caring for writers during the process.

Prize chair Ella Wakatama questioned the reliability of AI detectors in an appearance on BBC’s Front Row. She said she was not opposed to AI, but objected to treating a detector as the final arbiter. In her view, The Atlantic had called the winner a fraud without conducting a proper investigation.

Wakatama also said she did not want writers’ work, particularly that of Black writers, used as test material for technology that remains raw and unreliable.

Wara said she had read a 2021 story by Kundizesa and judged its tone and style markedly different from “The Blue Jockey.” Kundizesa replied that the claim assumes writers cannot work in different styles and genres, or substantially improve their craft over five or more years.

Wara cited her own research on Pangram, which she said found false positives to be rare, including when the tool assessed English writing by people who do not speak English as a first language. She also referred to a peer-reviewed study in which Pangram showed “high accuracy” when assessing student work written entirely with AI, partly with AI or without it. But she acknowledged on BBC that independent research into the accuracy of these tools remains relatively limited.

A detector is not a verdict

I think the central weakness in this dispute is not that one side has a detector and the other has a denial. It is that the evidence described publicly still depends on judgments that are difficult to verify: a tool’s score, a reader’s impression of style and a comparison with an earlier story.

The prize organizers’ investigation may establish more, but the announcement offers no standard for what would count as conclusive evidence. That gap matters beyond this award. When detection tools are still sparsely studied, treating their output as proof risks turning uncertainty into an accusation; refusing to examine it at all would leave writers and prizes without a credible way to answer one.