Viral bestseller Daggermouth may be AI-written—and readers loved it

H. M. Wolfe's novel Daggermouth tells the story of an unlikely romance between the president's son and an assassin hired to kill him, set in a surveillance state ruled by a masked elite. Since Wolfe uploaded it to Amazon as a Kindle ebook, it has become a viral sensation. A TikTok book influencer gushed, "Did I finish Daggermouth, or did Daggermouth finish me?" In February, Simon & Schuster secured the publishing rights for seven figures and commissioned a sequel due next year. Daggermouth has spent months on USA Today's bestseller list and ranks No. 1 in Amazon's "science-fiction romance" genre.

But researchers at Stony Brook University, led by computer science professor Tuhin Chakrabarty, discovered something unsettling while studying AI writing on Amazon. In a new study, they randomly selected more than 14,000 Kindle ebooks and scanned them with Pangram, an AI-detection tool. One in five came back as more than 25 percent AI-written. Daggermouth returned 60 percent, according to the underlying data set the authors shared. The researchers acknowledged the tool's limitations—Pangram flags text as either "AI-generated" or "moderately AI-assisted" (the latter can mean AI was used to edit or polish human writing, or vice versa)—but devised a second check. They scanned each book's text for "rare expressions," strings of at least five words in a specific order that appear in other suspected AI books but almost never in human-written ones.

Daggermouth had many of them. Near the end of a multipage sex scene in Chapter 27 appears: "They collapsed together, a tangle of sweat-slicked limbs and racing hearts." That exact sentence also appears in another self-published ebook by a different author. In Chapter 38: "Her pulse thundered in her ears, drowning out the murmurs of the crowd." It's also in an ebook by another author who published more than 400 different titles on Amazon in the past three years. The study has not yet been peer-reviewed, but Mohit Iyyer, a computer science professor at the University of Maryland who specializes in AI-language detection and was not involved with the study, told the researchers that even accounting for Pangram's fallibility, it is "almost statistically impossible" for an entirely human-written book to register as 60 percent AI.

Wolfe denied the allegation. "The suggestion that I used generative AI to write Daggermouth is wholly untrue," she 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." She cited Pangram's limitations and noted that accusations "have real consequences for authors, and they should not be made on the basis of technology that is known to be unreliable." Simon & Schuster's spokesperson, Susannah Lawrence, wrote that the company stands behind the book and notes it went through "the same editorial and production process as our other published titles." Neither Wolfe nor Lawrence addressed the researchers' findings about rare expressions shared across ebooks.

Chakrabarty countered that Daggermouth has "too many telltale signs" of AI for him to accept the explanation. He cited Chapter 18, "Thank You," where his research team broke the text into 12 chunks and Pangram classified all parts as AI-generated. Daggermouth stands out in another way: it appears to be one of just two novels by Wolfe in 2025, with its sequel not due until 2027—suggesting a process that involves "significant amounts of genuine human craft," according to Iyyer, who noted that using AI in a way that results in an enjoyable novel takes "a huge amount of effort by a human author." By contrast, the second-best-selling ebook in the data set is Inked Athena, by Nicole Fox, an author with 418 titles listed on Amazon in three years.

The broader research found that books with "substantial" AI assistance make up 20 percent of Amazon's ebook catalog, 12 percent of sales, and 10 percent of bestsellers in their genre. Jennifer Prokop, who reviews books for Kirkus and co-hosts Fated Mates, a popular podcast about romance novels, told researchers that people in her world keep talking about Daggermouth but said, "Nicole Fox, I've never heard a single soul talk about her." Prokop observed that mechanical prose crops up mostly in self-published ebooks on Amazon's Kindle Unlimited subscription service, where grifters use AI to produce titles by the dozen and exploit tricks like fake five-star reviews and click farms to boost metrics. "If you can get 10,000 people to read 10 pages of your book, and you have 300 books you're putting out," she said, "you can make real money. But you're not actually writing romances people want to read."

Christine Larson, a journalism professor at the University of Colorado at Boulder who studies technology's effect on creative work, told researchers: "Everyone thinks they can tell AI writing. But I don't think we can tell so much." She noted that privately, several authors and editors have told her they're using AI to accelerate or improve their output, and said, "I totally buy into the idea that a human working with AI can write something that is compelling. I have absolutely no doubt that we will see more of this in the future." Coral Hart, a South African author who uses AI to publish hundreds of books a year under different pen names, said she works hard to keep her pseudonyms secret after receiving death threats and thousands of one-star reviews following a New York Times story about her. Hart claimed successful authors have joined her AI-writing classes but declined to name them, saying: "Until they stop the bullying, the witch hunts, and the ousting, I don't think authors are going to be transparent, and I don't blame them. I'm certainly not going to."

The Daggermouth case exposes a fundamental problem in how the publishing world is responding to AI-written fiction. For years, AI-generated text has been treated as obvious slop—cheaply produced, poorly reviewed, mechanically plagiarized. Amazon's new-book releases nearly tripled from 2022 to 2025, most of them assumed to be low-effort content designed to trick readers or exploit algorithmic ranking. Yet Daggermouth broke that pattern: it became a genuine bestseller, beloved by readers and book influencers, serious enough that Simon & Schuster paid seven figures to acquire it. The question is no longer whether AI can produce readable fiction, but whether readers—or even publishers—can tell the difference when it does.

The research methodology, led by Tuhin Chakrabarty at Stony Brook, tries to move beyond the limitations of any single detection tool. Pangram alone is imperfect, but cross-referencing AI-detection scores with the discovery of identical rare phrases across multiple ebooks creates a pattern that independent experts call "almost statistically impossible" in a purely human-written work. Yet Wolfe's denial and Simon & Schuster's backing illustrate a deeper challenge: without definitive proof, and with reputational stakes high enough to invite legal pushback, publishers and the writing community face a stalemate. The authors' own study acknowledges Pangram's fallibility, and the research remains peer-review-pending.

What makes the situation untenable, according to experts quoted in the article, is that AI writing can now be lucrative only if kept secret. Authors and editors privately use AI to accelerate or refine their work, but public admission invites backlash—death threats, one-star review campaigns, and professional shunning. This creates perverse incentives: someone using AI transparently risks severe social punishment, while someone using it covertly can profit without penalty. The taboo against undisclosed AI authorship is morally justified (taking credit for work you didn't write is wrong), but the enforcement mechanism punishes individual authors rather than addressing the system that created the economic pressure to use AI in the first place.

No comments yet. Be the first to share your thoughts!