AI · 19 November 2026
Pangram AI Detection Flags Peruvian Novel; Publisher Backs Author
Pangram classified 99 percent of a prize-winning Peruvian novel as AI-generated, but Mexican publisher Almadía has stood by the award and the author after its own review. The case shows why detector scores need a fair review process.
What happened
An AI detection tool, Pangram, has classified 99 percent of a prize-winning Peruvian novel as AI-generated, yet Mexican publisher Almadía is standing by the award and the author after its own review. The finding has drawn attention because the same tool also classified more than 80 percent of the author's personal texts as machine-written.
The case has also prompted a reaction in the French market. Gallimard, one of France's best-known publishers, plans to begin screening manuscripts with Pangram, according to reporting by The Decoder.
The dispute therefore sits at the point where detection software, editorial judgement and an author's reputation meet. Almadía's decision to uphold its review suggests that a high detector score, on its own, has not been treated as proof of how a book was written.
Why it matters
For organisations that publish, judge or commission written work, this episode shows how quickly AI detection has moved from a background technical question into a practical operational decision. Once a vendor score can influence a prize, a contract or a publishing schedule, the output of the tool becomes a service interaction with the author, not just an internal check.
The wider lesson applies beyond publishing. Any organisation that uses automated classifiers to assess people, whether for recruitment, education, compliance or customer communications, faces the same question: what happens after the score arrives, and who has the right to explain the context behind it?
By the numbers
- 99 percent of the prize-winning Peruvian novel was flagged as AI-generated by Pangram.
- Over 80 percent of the author's personal texts were classified as machine-written by the same tool.
The Renascence take
Most readers will focus on whether the detector was right. The more useful question is whether the process around it was designed for a probabilistic output. A classifier score is a signal for review, and treating it as a verdict turns a measurement tool into an accusation engine.
The principle at stake is one of procedural fairness, which service designers already know well: any system that can deliver an adverse judgement needs a documented route for the person affected to respond, with evidence. Publishers that screen manuscripts should define in advance what a high score triggers, who reviews it, and how an author can supply drafts, notes and version history. A customer-obsessed operator will treat the detector as the start of a conversation with the writer, not the end of one. Anything less will cost trust faster than any false positive.
Sources
This briefing was written by our Newsdesk, synthesising reporting from the outlets below. Follow the links for the original coverage.
FAQ
Questions we get on this topic
More in AI
Stay ahead of CX
Get the signal, not the noise.
The stories shaping customer experience — plus the Journal and Experience Loom — in your inbox.


