Pangram flags 99 percent of Peruvian novel as AI-generated
Pangram classified 99% of David Silva Rojas’s award-winning Peruvian novel as AI-generated. Almadía reviewed the author’s documents and backed the author, while Gallimard plans manuscript screening.

Why it matters
- Pangram flagged 99% of *Donde el agua trabaja* as AI-generated.
- Pangram attributed more than 80% of the author’s interviews, essay and Instagram denial to AI.
- Almadía reviewed Silva Rojas’s documents and stands by the award.
- Gallimard reportedly plans to screen manuscripts with Pangram starting next year.
- The Authors Guild says Pangram reliably identified human-written texts in its tests but should not be the sole basis for decisions.
By Rebecca Stone
Pangram’s detector flagged 99 percent of the award-winning Peruvian novel Donde el agua trabaja as AI-generated, according to an account in The Point Magazine.
The result challenged the standing of the novel’s author, David Silva Rojas. Mexican publisher Almadía has answered with a review of documents supplied by the author and a decision to stand by the award.
The stakes extend beyond one controversy. The Authors Guild warns against using AI detectors as the “sole basis” for decisions, while Silva Rojas’s essay in The Point Magazine reports that Gallimard plans to screen manuscripts with Pangram starting next year.
What did Pangram flag?
The 99 percent figure describes the scope of Pangram’s classification within the novel. It is not a finding that Silva Rojas used AI, and it is not a determination of misconduct. The Point Magazine account presents Pangram’s output as an accusation that requires review.
El País reported that Pangram had also attributed more than 80 percent of the text across three categories:
- Interviews with Silva Rojas.
- An essay by the author.
- Silva Rojas’s Instagram denial of the allegations.
Read together, those results extend the dispute beyond the novel to statements in which the author addressed the accusation directly.
The available account does not disclose how Pangram handled the interviews, essay and denial, or whether the tool tested the materials separately or together. It also does not identify the passages that produced the 99 percent score or the passages that made up the remaining 1 percent.
That missing detail matters because a high classification share can identify material for scrutiny without explaining how the detector reached its result. The report gives readers the percentage but not the measurement behind it.
How does Silva Rojas frame the dispute?
In her essay, Silva Rojas says Latin American critics ignored Pangram’s results and dismissed the allegations wholesale as “racist.” She compares that reaction with two other cases in which AI use became part of a literary dispute.
One is Thélyson Orélien, whose use of AI the essay describes as confirmed. The other involves the Caine Prize and Jamir Nazir, whom Silva Rojas notes kept his award despite the accusations against him.
Those references serve a specific argument. Silva Rojas uses them to argue that critics can dismiss detector findings without first examining their substance, and that an accusation can become a verdict before an author receives a fair review.
She also argues that literary juries often favor clichéd stories about minorities, which language models can readily imitate. The report presents this as Silva Rojas’s explanation for why the detector may have scored her own work as AI.
The account does not supply data that test her claim about juries or model imitation. It does show why the case extends beyond detector accuracy: the allegations concern whose writing counts as credible and how literary institutions respond to suspicion.
Why does Almadía stand by Silva Rojas?
Silva Rojas’s essay in The Point Magazine says Almadía reviewed documents from the author and then stood by the award. The decision gives the publisher’s review a concrete role after Pangram’s classification.
The account does not identify the documents, describe their contents or explain what findings connected them to the award. It also leaves the award and the date of Silva Rojas’s recognition unnamed.
That gap prevents an outside reader from independently evaluating the publisher’s decision. It does not erase the decision itself: Almadía had the author’s documents, reviewed them and continued to support the award.
Standing by the award also does not necessarily amount to a separate finding that every sentence in the novel passed a human test. The report frames Almadía’s response as a rejection of the practical consequence of Pangram’s challenge.
What does Gallimard’s screening plan change?
Silva Rojas’s essay also reports that Gallimard plans to screen manuscripts with Pangram starting next year. The report gives no calendar year, selection criteria or rules for handling manuscripts that receive high scores.
For Gallimard’s editors, the tool would become a first-pass filter, much as a spell checker flags possible errors before a person reads the full text. That is a more consequential role than analyzing one disputed novel after publication.
The difference is temporal and institutional. A retrospective score can support an accusation or prompt a document review. Pre-publication screening can shape which submissions reach an editor, even when the announcement gives the score no final authority.
Almadía’s response offers the counterpoint. Its staff reviewed documents and backed the author. Gallimard’s plan would put Pangram into the intake process, giving the same type of score a direct role in evaluating new work.
Gallimard’s reported plan does not say whether a Pangram result can disqualify a manuscript. Without that policy, the announcement establishes interest in screening, not a rule for rejection or a verdict on any author.
What does the Authors Guild warning establish?
In the United States, the Authors Guild represents writers’ interests and warns against using AI detectors as the sole basis for decisions. The warning matters most when institutions are deciding whether a submission or an existing work should advance.
The group’s statement also acknowledges evidence in Pangram’s favor: the Authors Guild says Pangram reliably identified human-written texts as human-written in its tests. The report provides no numerical false-positive rate for the novel or those trials.
Those positions can coexist. Results in tests of human-written texts do not by themselves settle whether a detector can assign a 99 percent AI score to a literary work, or whether a publisher’s document review resolves the result.
The warning implies that detector output should inform, rather than replace, human review. The reports about Almadía and Gallimard show where that distinction will be tested.
Where does the case stand?
The factual record now consists of a sequence of actions and several unanswered questions. Pangram has issued two large classification claims. Almadía has reviewed documents and backed the award. Gallimard has a reported plan for broader screening. The Authors Guild has urged restraint.
The next test is procedural: what happens to a manuscript when Pangram flags it, whether the publisher can explain the decision, and whether the document review behind Almadía’s choice becomes a standard rather than an exception. The available reporting does not answer those questions.
For now, a 99 percent detection claim, a publisher’s documentary review and a reported screening plan remain separate events. Gallimard’s screening starting next year will show whether they become one policy.
Original: thepointmag.com
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
754 articles
Related articles
- Goncourt Academy Pulls Novel From Longlist Over AI Claim
- Caine Prize Winner Accused of AI Writing Denies the Claim
- Up to 20 Percent of New Wikipedia Articles Are AI-Generated, Cleanup Editor Says
- Kevin Roose Is AGI-Pilled: The AGI Chronicles, Machine Gods, and Why Belief Drives the AI Race
- Guardian columnist: shame over AI use misses the real target