When AI accuses the writer

David Silva Rojas won the Premio Internacional de Novela Breve Almadía with Donde el agua trabaja. On 23 October, El País reported that Pangram had classified 99% of the analysed text as AI-generated. The launch was suspended. The author denied having used artificial intelligence. A number entered the picture and changed the course of a novel. [1]

On 6 November, Almadía upheld the award after reviewing writing materials and consulting specialists. It also urged El País to apologise. In the publisher's view, the newspaper had failed to distinguish clearly between suspicion and established fact. [2]

The outcome matters: after reviewing the process, the publisher upheld the award. The percentage accused; human review decided. A suspicion deserves investigation. A reputation deserves more than a screen.

The starting idea is powerful: AI was trained on human books and now confuses that writing with its own. The detector does not detect a machine; it detects what seems familiar or regular to it. And people write regular prose too.

The scope is this: familiarity explains well-known false positives in a family of detectors based on predictability. It does not show that “most books” are mistaken for AI, nor does it explain every detector. That is the boundary of the argument. Within it, the question is clear: why turn regularity in a text into an accusation against the person who wrote it?

People write predictable prose too

In 2023, GPTZero flagged a passage from the US Constitution as probably artificial. ZeroGPT did the same with Génesis. They are different tools. Edward Tian, the creator of GPTZero, attributed the constitutional case to its repetition in training. Margaret Mitchell explained perplexity: how much a text surprises the model. [3]

Perplexity measures predictability. If words arrive as the model expects, surprise is low. Some detectors used that regularity as a sign of automated generation. So-called burstiness added another observation: how much those patterns varied within the document. The conceptual problem soon becomes clear. A regular style is also part of the human repertoire. Predicting the next word and finding out who wrote the previous one are different tasks. [4]

In those systems, recognisable writing can seem too easy to have been written by a person. Familiarity becomes suspicion. The tools change too: GPTZero says it abandoned perplexity and burstiness in autumn 2023 and moved to deep learning. [4]

Pangram, the tool at the centre of the publishing controversy, says it does not use those metrics. It describes a neural classifier trained to distinguish human documents from similar artificial versions. It learns differences between examples; it does more than measure surprise. [5]

Reading its percentage correctly also matters. According to the provider's documentation, a result of 99% expresses the proportion of the text labelled as artificial. The system's confidence is a separate output. It does not mean 99% certainty that the author cheated. Confusing the two turns an internal measure into an accusation that the measure itself does not contain. [6]

Books are indeed part of this story. The documentation for The Pile includes Books3 and literature from Project Gutenberg among its materials. The June 2025 ruling in Bartz contra Anthropic documents books used to train Claude. Training on human writing is documented. The regularity sought as an artificial signal is also part of the texts from which models learn. [7][8]

The errors go beyond anecdotes

The problem also appears beyond books. Liang and his colleagues studied seven detectors in 2023. In a sample of 91 English essays by non-native speakers, the average false positive rate was 61.3%. Changing the vocabulary reduced the accusations. Writing style was influencing suspicions about its origin. [9]

OpenAI withdrew its own classifier in July 2023 because of its low accuracy. It is a significant precedent: a company capable of producing automated text acknowledged that its tool for identifying it was not reliable enough. [10]

The relevant error is not just getting a label wrong. It is using that label to decide who deserves to publish, pass a course or keep a job. The greater the consequence, the greater the need to examine the method and allow a response.

Spain: evidence depends on the procedure

In a Spanish publishing or contractual dispute, the first step is to identify the agreed obligation: writing the work oneself, disclosing assistance or complying with a specific prohibition. Then comes proof of the breach. Spain's Civil Procedure Act (Ley de Enjuiciamiento Civil, LEC) allocates burdens according to the facts alleged and requires account to be taken of the availability of evidence. It admits documents, witnesses and experts; expert opinions are assessed according to principles of sound reasoning. A percentage carries no special evidential weight merely because it appears in a report. [11]

The burdens change with the procedure. If a dismissal is challenged, the employer must prove the facts set out in the dismissal letter that justify it. Under the disciplinary rules for students at public universities, the law provides for investigation, submissions, the production of documents, requests for evidence to be taken and a hearing. Each procedure determines who must prove what and how the person accused can defend themselves. [12][13]

The professional response is to demand the text examined, the tool's version, the full result and an explanation of the method that can be checked. In a private contractual investigation, hearing the author allows the accusation to be tested and the writing process to be examined. The decision must reflect the agreed obligation and the evidence gathered.

Norway: assess the evidence and allow it to be challenged

Norwegian civil litigation follows the principles of free assessment of evidence and the right to challenge it. Høyesterett (Supreme Court of Norway) has explained both principles when applying tvisteloven (Norwegian Dispute Act): the court examines the evidence as a whole, and the parties must be able to comment on the facts used to reach a decision. A commercial score is not equivalent to the legal probability of a breach. [14][15]

In employment governed by arbeidsmiljøloven (Working Environment Act), dismissal requires objective justification and, as a rule, a prior discussion with the employee. The percentage does not replace that assessment. [16]

Universities offer a particularly recent indication. The committee that submitted NOU 2026:14 (Official Norwegian Report) on 6 November considers Pangram the most reliable tool currently available. Even so, it advises against using detectors to assess academic work or handle suspicions of cheating. Even with good detectors and few false positives, separating the student's intellectual contribution from the system's remains difficult. The committee also raises issues of transparency, legal certainty and privacy. Its recommendation addresses something deeper than an error rate. [17][18]

Correctly classifying a text does not settle what its author did either. Deciding whether cheating occurred requires reconstructing the author's contribution and applying the relevant rules. That work requires an investigation that can be explained and challenged. Technical accuracy and the justification for a sanction answer different questions.

Reconstruct the writing before judging it

Drafts, version histories, notes, interim submissions and testimony from those who followed the work should be kept. These materials make it possible to reconstruct the process that the detector tries to infer from the result. They must be examined with attention to their authenticity, integrity and context. The investigation concerns the writing: how it developed, what decisions the author made and what assistance they used. The statistical label must be tested against that history.

The answer to the problem is to investigate authorship before judging it. A publisher, university or employer must establish what happened, what was permitted and what consequence is appropriate. People write familiar and regular prose too. No percentage relieves the decision-maker of the responsibility to prove and explain. Almadía reviewed the process and upheld the award. That is the order: examine the writing, hear its author and decide on the evidence.

Sources

[1] El País.

[2] El Cultural.

[3] Ars Technica.

[4] GPTZero.

[5] Pangram.

[6] Pangram interpretation guide.

[7] EleutherAI.

[8] Court ruling.

[9] Study published in Patterns.

[10] Updated OpenAI announcement.

[11] LEC, Articles 217, 299 and 348.

[12] Spain's Labour Procedure Act (LRJS), Article 105.

[13] Spain's University Coexistence Act, Articles 7 and 19.

[14] HR-2020-2021-A.

[15] HR-2014-02467-A.

[16] Articles 15-1 and 15-7.

[17] Official report, Chapter 7.4.2.5.

[18] Legal annex.