RESOLVING THE LEDGER
This is one outlet's own report from The Atlantic — the article as it was filed. Other outlets are covering the same event; open the full story to compare every source side by side.
See the full story · 1 sources
Most self-published authors can only dream of writing something as popular as Daggermouth . Set in a surveillance state ruled by a masked elite, the novel traces the unlikely romance between the president’s son and an assassin hired to kill him. Since H. M. Wolfe uploaded Daggermouth to Amazon as a Kindle ebook, it has been a viral hit. “Did I finish Daggermouth , or did Daggermouth finish me?” a TikTok book influencer gushed . In February, Simon & Schuster paid seven figures to secure the publishing rights to the book and its sequel, which is due out next year.
Daggermouth has spent months on USA Today ’s best-seller list and is ranked No. 1 in Amazon’s “science-fiction romance” genre. It also tops a less prestigious list, one compiled by academics studying AI writing. Among thousands of self-published Amazon ebooks that they identified as containing “substantial AI text,” none was more popular than Daggermouth .
Until recently, AI fiction has been easy to dismiss as slop. On Amazon, new-book releases nearly tripled from 2022 to 2025. (That’s probably not because humans suddenly dedicated themselves en masse to the written word.) Many are blatant copycats of existing titles, seemingly intended to trick readers into buying them. Their sales figures are generally meager; the reviews, indignant. A suspicious number of them feature protagonists with the same names . Last year, readers of one fantasy novel noticed a line that seemed to be responding to a prompt asking a chatbot to write in the style of another author. But if a book as beloved as Daggermouth were written with a chatbot, it would raise an awkward question: Could AI fiction be getting … good?
In a new study , researchers randomly selected more than 14,000 Kindle ebooks and ran them through Pangram, an AI-detection tool. One in five came back as more than 25 percent AI-written; Daggermouth returned 60 percent, according to an underlying data set that the authors shared with me. Those numbers, they acknowledge, tell us only so much. The researchers grouped together text that Pangram had flagged as either “AI-generated” or “moderately AI-assisted” (the latter of which can mean that AI was used to edit or polish human writing, or vice versa). And although Pangram has a low rate of false positives, it is not perfect .
[ Read: America has a Pangram problem ]
So the study’s authors, led by Tuhin Chakrabarty, a professor of computer science at Stony Brook University, devised another check. They scanned each book’s text for “rare expressions”—strings of at least five words in a specific order—that also appear in other suspected AI books but almost never in human-written ones. Daggermouth had many of them. For instance, the following sentence appears near the end of a multipage sex scene in Chapter 27: “They collapsed together, a tangle of sweat-slicked limbs and racing hearts.” That same sentence, word for word, also shows up in another self-published ebook by a different author. Or consider this line from Chapter 38 of Daggermouth : “Her pulse thundered in her ears, drowning out the murmurs of the crowd.” It’s also in an ebook by another author who has published more than 400 different titles on Amazon in the past three years.
The study has yet to be peer-reviewed, and to be clear, there is no definitive proof that Wolfe used AI. “The suggestion that I used generative AI to write Daggermouth is wholly untrue,” Wolfe said in a statement sent by her lawyer. “I wrote this book myself. 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.” Referencing Pangram’s limitations, Wolfe added 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.” Susannah Lawrence, a spokesperson for Daggermouth ’s publisher, Simon & Schuster, told me in an email that the company stands behind the book, noting that it went through “the same editorial and production process as our other published titles.” Neither Wolfe nor Lawrence addressed my questions about the “rare expressions” that Daggermouth shares with other recent ebooks.
Chakrabarty, who researches AI writing, told me that Daggermouth has “too many telltale signs” of AI for him to buy that explanation. He cited Chapter 18, “Thank You,” as an example: His research team broke it into 12 chunks for Pangram to evaluate, and the tool classified all parts as AI-generated. (In my own spot checks, the tool classified the first few paragraphs as human-written and the rest of the chapter as 100 percent AI.) I ran this by Mohit Iyyer, a professor of computer science at the University of Maryland who wasn’t involved with the study. An expert on AI-language detection, Iyyer told me that even accounting for Pangram’s fallibility, it’s “almost statistically impossible” for an entirely human-written book to register as 60 percent AI. He added that the study’s broader methodology, which connects AI writing to ebook sales, is sound.
Whether or not you believe that Wolfe deserves the benefit of the doubt, the same questions about how Daggermouth was written apply to a growing number of other hits on Amazon. Chakrabarty’s study found that books written with “substantial” AI assistance made up 20 percent of Amazon’s ebook catalog, 12 percent of sales, and 10 percent of best sellers in their genre.
Yet Daggermouth stands out for more than just its popularity. The second-best-selling ebook in the data set is Inked Athena , by Nicole Fox, an author who has 418 titles listed on Amazon. Daggermouth , in contrast, appears to be one of just two novels to Wolfe’s name in 2025, and its sequel is not due out until 2027—an indication that Wolfe’s process likely involves significant amounts of genuine human craft. Iyyer said that using AI in a way that results in a novel that’s actually enjoyable ...
AIPROPX is an independent multi-source news index — we track, compare, and connect coverage from across the web into one place you won't find anywhere else.