A 99% AI score can't referee a literary prize, and neither can any detector

What do you do when a number says a prize-winning novel is 99% machine-written, the author says it isn't, and the rules say a mere suspicion is enough to disqualify it? You find out, fast, that the rules were written for a world that doesn't exist.

That is the bind around Donde el agua trabaja, the novella by Peruvian author David Silva Rojas. According to EL PAÍS, Pangram attributes 99% of the roughly 26,000-word text to an AI model. The book won the first Almadía Ventosa-Arrufat novella prize, worth 20,000 euros, chosen from 2,204 manuscripts. The rules required a declaration that the work was not written, in whole or in part, with AI tools, and promised disqualification "ante cualquier indicio o sospecha". The publisher has suspended the book tour. It says the author denies using AI, and it promises a detailed review "sin anticipar conclusiones".

A suspicion clause is a trap for everyone

Look at that rule again. It sets the bar for disqualification at suspicion, and it offers no way to resolve a suspicion. In 2025 that looked tough. In 2026, with detectors a browser tab away, it means any reader with a scan can put a prize on trial. The contest has handed its authority to whoever runs the software first.

We don't think this means Pangram is junk. The evidence says otherwise, and it's worth being fair about it. A June 2026 peer-reviewed study from the Vrije Universiteit Brussel, which Pangram itself cites, found it the only one of four tools to produce satisfactory results. The Authors Guild tested five detectors on human writing published in 2022 or earlier, and Pangram returned 0 percent on all ten articles. The company claims a false positive rate around one in 24,000. If you had to bet, a 99% score on a long text is a heavy signal.

But a heavy signal is not a verdict. Those low error rates are the company's own figures or controlled benchmarks, not a guarantee about one specific book. Pangram's confidence labels are not calibrated probabilities of being right, and the company selling the detector is not the neutral party this situation needs. Tim Requarth walked back part of his earlier critique after talking to independent computer scientists, yet he still argues that company-reported error rates should be taken "with a grain of salt" in the real world. We agree with that narrower skepticism.

The author's style is a confound nobody can resolve

The most interesting detail in EL PAÍS's reporting is the one that cuts both ways. Silva Rojas has said he took the idea of the haiku from Japan, its ability to concentrate an experience in very few elements, and that it shaped the novel's rhythm. The jury praised "oraciones breves". Very short sentences are also something some AI models tend toward. The newspaper itself says there's no evidence that this style confuses Pangram, but that it can't be ruled out either.

That honesty is the whole point. A detector learns statistical patterns of style, and a writer who deliberately cultivates a spare style is working in the same territory. The Authors Guild makes the same argument: polished, controlled prose may resemble what the tools flag. Add the fact that the same newspaper found Pangram scoring the author's interview answers 97% AI while the journalist's questions came out fully human, and you have circumstantial evidence that is suggestive. It's still circumstantial.

Contests should ask for drafts, not just run scans

The fix isn't to abandon rules. It's to put them on something provable. In our view, a serious prize in 2026 should:

• Require disclosure of AI use, tiered, instead of a flat ban that invites lying.
• Ask finalists for process evidence, such as drafts, version history and notes, before announcing a winner.
• Treat a detector score as a reason to ask for that evidence, never as the evidence itself.
• Publish what happens after a flag, so nobody is judged by a mob.

EL PAÍS says it asked the publisher about drafts or version records and was told the matter is "compleja". In our view, that question should have been answered before the prize was announced, not after the headline.

Our verdict: if the author did use AI, the prize rules were broken and he deserves the consequences. If he didn't, he is being tried by a score no one can cross-examine. Either way the institution failed. Detectors are fine as smoke alarms, but they are no substitute for the fire inspection.