Book News
A Bestseller Was Accused of Being AI-Written. Here's What's Actually Known.
This year's biggest BookTok crossover hit just got tangled up in a very 2026 kind of controversy. Daggermouth, the viral fantasy novel by H.M. Wolfe published by Simon & Schuster, is the subject of a new academic study claiming a large share of its text was AI-generated. Wolfe says that's wrong. So does her publisher.
Stories like this tend to get compressed as they travel: a study result becomes “proof,” a denial gets buried below the headline. We'd rather slow it down. Here is what the study actually found, what the detection tool behind it can and can't tell you, and what both Wolfe and Simon & Schuster have said in response — in full, not trimmed to a soundbite.
What the study actually claims
Researchers at Stony Brook University, in a paper by professor Tuhin Chakrabarty, ran 14,000 randomly selected Kindle ebooks through an AI-detection tool called Pangram. Their headline finding: roughly one in five of those 14,000 books “contained more than 25% AI-generated text,” according to the tool.
That 14,000-book, one-in-five figure is the actual substance of the research. It is a claim about a broad slice of the Kindle marketplace, not about any single author. Among the individual titles the tool flagged, though, was Daggermouth. Pangram scored it 60% AI-written.
That is the number that has been circulating. It is worth being precise about what it is and isn't. It is not a court finding, an admission, or a peer-reviewed conclusion. It is a probability score, produced by one detection tool, applied to one manuscript, generated as part of a study that has not yet been peer-reviewed.
It is also worth noticing what got lost in the jump from “one in five of 14,000 books” to “this specific bestseller.” The study's actual finding is about the Kindle marketplace broadly: a large, essentially untested claim that a meaningful share of self-published and traditionally published ebooks alike may contain substantial AI-generated text. That is the bigger, less viral story. Daggermouth became the headline because it's a book people already knew, attached to a name people could search. A number is more compelling when it has a face on it, even when the face wasn't the point of the research.
What a detector score can and can't prove
The paper has not been peer-reviewed. That matters more than it might sound like it does. Peer review is the process by which other researchers in the field check a study's methods before it gets treated as established science. Chakrabarty's paper hasn't gone through that yet, which means its sample, its methodology, and its conclusions haven't been independently checked by anyone outside the team that produced them.
AI detectors have a documented false-positive problem. Tools like Pangram work by scoring how closely a piece of text matches patterns the tool has learned to associate with AI-generated writing. That is a fundamentally different thing from proving how a specific person actually wrote a specific sentence. Simon & Schuster made exactly this point in its public response to the study, stating plainly that AI-detection tools “have been shown to produce false positives.” We aren't going to claim more insight into how Pangram works internally than what's actually been reported. No source we can verify explains its underlying mechanism, so neither will we.
The researchers also used a second, more specific method: checking suspect books for what they call “rare expressions” — phrases of five or more words, in a particular order, that show up rarely in human-written text but more often in text suspected of being AI-generated. It's a narrower signal than a single percentage score. Book Riot, reporting on the study, noted that Wolfe has not specifically addressed this rare-expressions question in her public statements.
Wolfe's denial
Wolfe has denied the accusation directly, and more than once, in her own words:
“The suggestion that I used generative AI to write Daggermouth is wholly untrue.”
— H.M. Wolfe
“I wrote this book myself.”
— H.M. Wolfe
“I have been outspoken about my opposition to generative AI and what it's doing to writers, artists, and the creative community. I don't believe it belongs in the writing process.”
— H.M. Wolfe
That isn't a hedge or a non-answer. It's a direct, on-the-record denial from an author who, by her own account, has been publicly critical of generative AI in creative work.
Simon & Schuster's response
Wolfe's publisher backed her just as directly, in two separate statements:
“[Daggermouth went through] the same editorial and production process as our other published titles.”
— Simon & Schuster
“We do not believe conclusions about an author's work should be drawn from AI-detection tools that have been shown to produce false positives. H.M. Wolfe wrote 'Daggermouth,' and we stand behind her and her work.”
— Simon & Schuster
That's about as unambiguous as a publisher statement gets. Simon & Schuster isn't hedging on whether it believes its author; it's directly challenging the evidentiary weight of the tool that produced the number in the first place.
What's actually still unresolved
Strip away the noise and here's what genuinely remains open: whether the “rare expressions” pattern researchers found in Daggermouth's text has an explanation Wolfe hasn't yet addressed publicly, and whether Pangram's 60% score reflects something real about how the book was drafted or is itself a false positive of the kind Simon & Schuster describes. Neither the study's authors nor anyone else has published an independent replication of the finding on this specific book, and the underlying paper still hasn't cleared peer review.
It's also worth sitting with why this particular gap is so hard to close from the outside. A detection tool can only ever describe a pattern in finished text. It has no access to a manuscript's drafts, an author's notes, or the editorial back-and-forth that produced the final version on shelves. Readers, journalists, and researchers are all working from the same limited vantage point: the published book, run through a piece of software none of us can fully audit ourselves. That's not a reason to dismiss the study. It's a reason to hold its conclusion the way the researchers themselves seem to, and the way a non-peer-reviewed paper deserves: as a finding that invites further scrutiny, not one that has already survived it.
What we can say honestly is this: a non-peer-reviewed study, using a detection tool with a documented false-positive problem, produced a number. The author it names has denied it in detail, more than once, in her own words. Her publisher has backed her by name and challenged the reliability of the tool itself. Nobody outside H.M. Wolfe knows how Daggermouth was actually written, and a detector score by itself doesn't settle that question. It raised it.
If you want the fuller picture on how AI-detection tools work in general, where their real limits are, and what actually separates AI-written text from AI-assisted or unfairly flagged human writing, we wrote a companion piece: How to Tell If a Book Was Written by AI.
Sources
- Book Riot — “Study Claims Viral Bestseller Daggermouth Is AI-Generated”
- Tech Times — “Daggermouth Hardcover Drops Today; BookTok Hit Flagged AI by Researchers”
- Breitbart Tech — “Suspected AI Usage by Authors Is Causing Chaos for Book Publishers”
- AI Weekly — “Bestseller Daggermouth Flagged as 60% AI in Kindle Study”