Proof is needed. Caine Prize: software accuses the winner of using AI. But an algorithm cannot pass judgement

Artificial intelligence promised to help us recognise what is false. Instead, it is creating a more subtle problem: what happens when we entrust another artificial intelligence with the power to determine who has cheated? The case of Simbarashe Kundizeza is a perfect example. On 6 October, the Zimbabwean writer won the Caine Prize for African Writing with his short story "The Blue Fighter". Five days later, Vauhini Vara reported in *The Atlantic* that she had run the text through Pangram, an AI detection programme: the story appeared to be almost entirely artificially generated. Kundizeza categorically denied this, claiming to have a verifiable account of the work’s creation and to have drawn inspiration from family experiences. The Caine Prize has launched an investigation, without revoking the prize. The issue is not whether to believe the writer or the detection tool, but to understand what kind of evidence we are prepared to accept.

A detector can flag anomalies and statistical similarities. But a flag is not a verdict. Even highly accurate systems produce false positives, and human writing is not a chemical substance with an unchanging formula: a writer changes tone, imitates, gets worse, gets better. Turning all this into a percentage of guilt amounts to confusing an investigative tool with a judge. Ellah Wakatama, who chairs the prize, put the matter in the right light: we need evidence that does not come solely from a machine. Drafts, revision histories, notes, emails, and editors’ testimonies. They are not infallible, but they allow us to reconstruct a process rather than simply accepting an algorithmic verdict. It is also a lesson for journalism: a detector can indicate that something is worth verifying; it cannot be the verification itself. In the first case, the algorithm raises a question; in the second, it settles it. Then there is a second risk. Do you write too well? AI. Too badly? AI. Do you use metaphors? AI. Do you change your style? AI. A tool designed to safeguard authenticity may end up imposing a rigid idea of how a human being ought to write. The problem certainly exists: awards, universities and newspapers must establish rules governing the use of generative models.

But precisely because this is a serious matter, we cannot take the easy way out. The rule of law in the age of AI begins here: not in the rejection of automated tools, but in the refusal to regard them as oracles. Between suspicion and public accusation, there must remain something very ancient and very human: evidence. The Kundizeza affair does not yet tell us how "The Blue Fighter" was written. It does, however, tell us how we should not judge it: by pressing a button and turning a machine’s response into a condemnation.