Can AI Detectors Prove a Novel Was Written by AI?

In 2026, AI detectors stopped being a classroom tool and became a way to put novelists on trial in public. The pattern repeats: someone pastes passages from a buzzy book or prize-winning story into a detector, posts a screenshot showing a high AI score, and the argument begins before the author has said a word.

The most dramatic case so far is French. On September 25, 2026, Reuters reported that the Académie Goncourt had removed a bestselling debut novel by a Haitian-Canadian author from the running for France’s most prestigious literary prize. The accusations came from a little-known X account that said running passages through the detector Pangram showed the book “was written almost entirely by artificial intelligence.” Reports that the author had plagiarised in earlier, unrelated works followed. The Académie cited both the AI and plagiarism accusations.

On October 13, 2026, the author told AFP that “not a single sentence in this book was created by a machine,” said he had started the book in 2017 and finished a first draft in 2019, and admitted to plagiarism “mistakes” in earlier writing. AFP noted he had not made early drafts available, and that his Canadian publisher said it could not categorically confirm or deny the AI allegations. As of mid-October 2026, that dispute is unresolved.

The most useful detail in AFP’s reporting was its own test. AFP ran an excerpt through Pangram and got a 100% probability of AI. It ran excerpts through other detectors and got everything from “very likely” human-written to 100% AI-generated. Same text, opposite verdicts.

That spread is the core problem. The Authors Guild tested five popular detectors in May 2026 on ten of its own articles, all published in 2022 or earlier, before generative AI was widely available, so any accurate tool should have scored them as human. Pangram and Originality.ai did well: Pangram scored all ten at 0% AI. But ZeroGPT ranged from 5% to 76%, and Sidekicker.ai flagged every article as mostly AI, two of them at 100%.

So the honest summary is not that every detector is useless. Some are good at recognising clearly human text, and Pangram’s maker says its false-positive rate is 0.01%, according to the Christian Science Monitor. The trouble is that a reader with a screenshot rarely tells you which tool they used, on what passage, or how long it was, and the tools change constantly.

Detectors are themselves AI models trained to spot statistical patterns in machine text, such as predictable word choice and rhythm. The Authors Guild points out the paradox: polished, edited human prose shares many of those patterns, because language models learned from polished, edited human prose. The more controlled a writer’s style, the more it may resemble what the tool is built to flag.

A low error rate also stops sounding low at scale. A University of Michigan doctoral student who studies detectors told the Christian Science Monitor that they do work in some contexts, better on fiction and on longer text, but that even a 1% false-positive rate is high for publishing, where one accusation can end a career. And a 2023 Stanford study, the Monitor noted, found detectors biased against non-native English speakers (Pangram was not part of that study).

Finally, a detector answers a narrower question than the one people ask. It estimates whether text looks machine-generated. It cannot tell you whether an author drafted the book and then ran it through an AI editor, used AI for research, or wrote every word by hand in a style that happens to look smooth.

If the instinct is to fall back on a reader’s judgement, the research is not encouraging. In a 2026 study in Judgment and Decision Making, Sydney Sears and Deena Skolnick Weisberg had 1,682 participants read short stories of about 1,000 words. Readers rated ChatGPT’s stories higher than the human ones on quality, 1.54 versus 0.97 on a scale from -3 to +3, according to The Decoder. In two further experiments, people asked to tell AI and human stories apart did no better than chance.

The twist is the label. Regardless of who actually wrote a story, participants rated it higher when they were told a human wrote it. That makes “it reads like AI” a weak basis for an accusation, and it shows how much a label can colour a verdict. (We look at what AI romance does and doesn’t do well in can AI write good romance?.)

The clearest counterexample to the Goncourt affair is the 2026 Commonwealth Short Story Prize. After readers flagged winning stories in May, one of which Pangram scored as 100% AI, the Commonwealth Foundation investigated for about a month. According to Book Riot and the Christian Science Monitor, it reviewed working drafts, time-stamped documents and notes, did not use AI-detection software, and found no evidence of AI use. The writer under the most scrutiny went on to win the overall prize. Granta, which had published the winning stories for more than a decade, still ended its partnership with the prize, citing editorial integrity.

The Authors Guild’s advice to publishers points the same way: never cancel a contract or pull a book on accusations or detector results alone, disclose methodology, and give authors a full chance to respond. For writers, the practical takeaway is simple. Your process is your evidence. Keep dated drafts, revision history, notes and outlines. A file history that shows a story growing over months says more than any score.

We do not run AI detection on stories and do not certify anything as human-written, and we would be wary of any platform that claims a detector can settle authorship on its own. Our AI editor works alongside a writer on their own manuscript, giving notes on a chapter, adding transitions or expanding a passage, and it is free within a daily allowance. Our creator agreement asks that published work is original and that every character is an adult, and publishing submits a story for review before it appears in the fiction library.

If you use the editor, keep your own drafts as you go. That record protects you far better than a detector score ever could. For how disclosure works on the largest store, see our explainer on Amazon KDP’s AI disclosure rules.