Did AI Write It? The Percentage Doesn't Tell the Whole Story

A recent investigation has brought AI detectors back into the literary debate. For writers, publishers and readers, the challenge is to understand what a tool flags and what we need to establish before claiming that a machine wrote a book.

Imagine a publisher receiving two files: a novel and a report with some of its paragraphs highlighted. The first tells a story; the second offers a classification. As soon as a percentage appears, the conversation can take a turn: we stop asking how the novel works and start arguing about who wrote it.

The question deserves careful investigation. An accusation can damage the bond between an author and their readers. The possibility that a text is being sold as someone's own work while its origins remain hidden deserves scrutiny too. Both concerns call for evidence read with care.

From a sample to a headline is quite a leap

On 13 October, EL PAÍS published an analysis using Pangram to examine 119 popular books from 2025 and 2026. Of the 44 self-published titles in that selection, 13 had more than 50% of their text classified as AI-generated. Those figures describe a sample, not Spanish self-publishing as a whole. The EL PAÍS investigation.

That last sentence matters as much as the numbers. If we turn a selection of popular titles into a statement about all published books, we change the question the data can answer. And if we move straight from a text classification to a claim about someone's conduct, we're adding a conclusion that needs further evidence.

What the result actually means

Pangram 4's documentation distinguishes between passages classified as human-written, AI-assisted or AI-generated. It calculates the proportions for a document by counting the characters assigned to each category. The company also makes clear that its confidence score isn't a calibrated probability. These are different measures; we need to know which one we're looking at. Pangram 4 model card.

Consider a hypothetical example: a report classifies 80% of the material analysed as AI-generated. That result describes how the system classifies the text it was given. It doesn't mean there's an 80% chance that the person whose name appears on the book deceived the public. Nor can it, on its own, reconstruct the conversations, decisions and revisions that produced the file.

That doesn't mean we have to dismiss detectors. It means we have to describe their results precisely and weigh them alongside other evidence. Pangram itself acknowledges a non-zero error rate and warns of the harm false accusations can cause. That caution comes from the developer's own documentation; it isn't an objection dreamed up to sidestep the debate.

Generating and editing aren't the same thing either

The Kindle Direct Publishing guidelines put this distinction into practice. Amazon requires publishers to disclose AI-generated content, including images and translations. By contrast, it considers content AI-assisted when a person creates it and uses AI to correct or polish it, and doesn't require that assistance to be disclosed. It also continues to classify content as AI-generated even after subsequent editing. Official KDP guidelines.

These are one platform's rules, not a universal definition of literary authorship. But they help bring precision to a discussion that too often goes no further than the broad question of whether AI was used. Correcting a typo, exploring an idea and commissioning a scene are different kinds of intervention and need to be described as such.

For our publishing community, the useful questions are what AI was used to do, where in the process it was used and whether those involved in publishing the work knew about it. A specific account gives us firmer ground for discussing responsibility than a label that lumps everything together.

Research needs boundaries too

A study by Chakrabarty and other researchers examines 14,419 genre fiction books self-published on Amazon, using text detection and sales data. The latest version listed on arXiv is dated 3 August 2026 and is presented as a working paper under review. It shouldn't be announced as a discovery made this week or as a settled scientific finding. Study record and full text.

The size of a sample matters; so do its categories and the methods used to establish them. When reporting on studies like this, we need to distinguish the findings their authors report from the conclusions a publisher, reader or court might wish to draw from them.

An editorial conversation, not a verdict

At Tu Letra Online, we propose treating a detector's flag as the starting point for a documented review: establish exactly what was analysed, keep the result and hear how the manuscript was written. Drafts or a revision history can provide context where they exist. Their absence alone doesn't prove that AI was involved either.

Transparency calls for agreements that authors and readers can understand. And trust depends on our being able to investigate without letting a suspicion define someone for good. A percentage can be a useful clue. A conversation about a book also needs people who can make sense of it.

• What uses of AI would you expect an author or publisher to explain to readers?
• What other evidence would you ask for before making a decision about a manuscript flagged by a detector?
• When choosing a book, how much does the writing process matter to you, alongside the literary result?

Share your experience, make your case and disagree respectfully. We want to hear both from people who use these tools and from those who prefer to work without them.