AI-detection claims against a prize-winning novel reopen debate on literary authenticity

What happened

David Silva Rojas, a Peruvian lawyer and writer born in 1981 in Ica, won the first International Short Novel Prize Almadía Ventosa-Arrufat in July 2026 for his manuscript "Donde el agua trabaja." The work was selected from more than 2,200 submissions by a jury made up of writers Cristina Rivera Garza, Selva Almada and Mónica Ojeda, who praised the novel's use of silences and its spare, drip-like sentences, according to El Español.

The prize carries a 20,000-euro award and, according to its published rules cited by El Español, explicitly states that "no partial or total use of artificial intelligence tools for writing works presented to the prize will be admitted under any circumstance."

The AI-detection findings

On 23 October 2026, Spanish newspaper El País published an investigation into 175 Spanish-language literary works, reported on by Infobae and El Español. Using the detection tool Pangram, version 4.1, the analysis concluded that 99% of the roughly 26,000-word novel was produced by an AI model. The same tool flagged 97% of Silva Rojas's written interview answers as AI-generated — while the journalist's own questions tested as entirely human-written — and found 80% likely AI origin in a separate 924-word essay he had published, per Infobae's report.

Pangram's maker claims an extremely low false-positive rate, cited by Perfil as 0.0041%, but the tool's own documentation acknowledges varying performance across genres, languages and writing styles. Perfil noted that such statistical detectors can misfire even on texts written long before generative AI existed, since they flag patterns rather than confirmed provenance.

Publisher and author response

Almadía suspended the novel's promotional tour, cancelling a Madrid presentation that had been scheduled for the Friday after the report, along with planned stops in Barcelona, Seville and Cádiz, according to El Universal.

Almadía said Silva Rojas had signed a declaration affirming he had not used AI tools, and that neither the jury nor the editorial team had suspected AI involvement during their review. The publisher added that detection software offers "valuable information" but is not sufficient on its own to reach conclusions that could damage an author's career, per El Universal.

Silva Rojas did not respond to the journalists behind the original investigation. Almadía said its internal review remains open, with no final conclusion announced as of the most recent reporting.

Not an isolated case

Perfil linked the Silva Rojas case to an earlier 2026 controversy involving Haitian-Canadian author Thélyson Orélien, whose novel "C'était ça ou mourir" became a bestseller of the French literary season (rentrée littéraire). Anonymous posts on X using Pangram analysis prompted France's Goncourt Academy to remove the book from its prize selection. In that case, publisher Boréal reportedly found AI-related metadata tags inside the manuscript's document files — physical evidence beyond a statistical score. Orélien acknowledged using AI tools for historical research but denied that AI had generated his sentences; his publisher said it remained uncertain about how far the AI's role extended, per Perfil.

Why publishers and prizes are under pressure

The two cases expose a gap many literary contests did not anticipate: rules banning AI use existed on paper, but no contest cited in this reporting described a standard verification process — such as draft-history checks or mandatory metadata review — applied before a winner is announced. Pangram-style detectors are statistical estimates, not legal proof, and Perfil's report stressed that treating a classifier's output as a "public verdict" without corroborating evidence would be risky for both authors and institutions.

For publishing houses, the dispute raises practical questions about contracts and submission rules: whether to require manuscript drafts, revision histories or metadata checks before announcing winners, and how to handle a flagged book that has already been printed, marketed or shortlisted. For readers, it raises the separate question of how literary value should be judged if a text's origin cannot be settled with certainty.