A Study Flagged Bestseller Daggermouth as 60% AI-Written. The Author Says That's Wrong

A Stony Brook University research team scanned more than 14,000 Kindle ebooks with the AI-detection tool Pangram and found that the BookTok hit "Daggermouth" scored 60% AI-generated, the highest of any popular title in their dataset. The author, H.M. Wolfe, denies it, and her publisher, Simon & Schuster, is standing behind the book. The dispute is a useful test case for every author who writes with AI assistance, because it shows exactly how uncertain an accusation like this really is.

What the study found

The research was led by Stony Brook computer science professor Tuhin Chakrabarty, who ran the Pangram classifier across a large, randomly sampled set of Kindle titles. Beyond the raw AI-detection score, the team also checked each book for what they called "rare expressions": strings of at least five consecutive words that turn up verbatim across multiple suspected AI-written books but almost never in human-written fiction. According to reporting on the study picked up by Book Riot and Tech Times, Daggermouth contained a number of these shared phrases, which the researchers treated as a second signal on top of the Pangram score. The study has not been peer-reviewed.

Both sides are pushing back

Wolfe rejected the finding directly through her lawyer, calling the claim "wholly untrue" and arguing that "accusations like these have real consequences for authors, and they should not be made on the basis of technology that is known to be unreliable." Simon & Schuster, which reportedly paid a seven-figure sum in February for the book and its sequel, said through a spokesperson that Daggermouth went through "the same editorial and production process as our other published titles."

Both statements point at the same underlying problem: an AI-detection score is treated by readers as a verdict, but the tool behind it was never built to deliver certainty on any single book. Independent testing by the University of Chicago's Booth School found Pangram's aggregate accuracy to be very high on clearly AI-generated or clearly human text, with a lower false-positive rate than competing detectors. But that same testing found accuracy drops sharply, into the 60 to 80 percent range, on text that blends AI drafting with heavy human rewriting, which is exactly the gray zone a working novelist's process tends to fall into.

Why this matters for you as an author

Daggermouth is not an isolated incident. It follows a string of recent cases, including a $2 million book deal that collapsed over AI suspicion and Hachette pulling a novel over the same kind of doubt, where a detector score or a reader's hunch was enough to put a career on hold before anything was proven either way. Detection tools are getting better at spotting fully AI-generated text, but they are still weak at handling the far more common case: a human writer who outlines, drafts, or revises with AI help and then rewrites the result in their own voice.

That gap is why documenting your process matters more than ever, something we covered when authors started filming themselves writing to prove a book wasn't AI-generated. A detector score is not proof of anything on its own, in either direction, and an accusation can travel faster than any explanation you're able to give afterward.

PageWriter Studio keeps a record of that process for you automatically. Every outline, draft, and revision stays tied to your account, with a style profile that keeps the writing consistent with your own voice instead of a generic AI default, so you have something concrete to point to if a detector or a reader ever questions how a book came together. If that kind of built-in proof of process sounds useful for your next project, you can start a free trial.