Most self-published authors can solely dream of writing one thing as widespread as Daggermouth. Set in a surveillance state dominated by a masked elite, the novel traces the unlikely romance between the president’s son and an murderer employed to kill him. Since H. M. Wolfe uploaded Daggermouth to Amazon as a Kindle e-book, it has been a viral hit. “Did I end Daggermouth, or did Daggermouth end me?” a TikTok ebook influencer gushed. In February, Simon & Schuster paid seven figures to safe the publishing rights to the ebook and its sequel, which is due out subsequent 12 months.
Daggermouth has spent months on USA Right this moment’s best-seller checklist and is ranked No. 1 in Amazon’s “science-fiction romance” style. It additionally tops a much less prestigious checklist, one compiled by teachers learning AI writing. Amongst 1000’s of self-published Amazon ebooks that they recognized as containing “substantial AI textual content,” none was extra widespread than Daggermouth.
Till not too long ago, AI fiction has been straightforward to dismiss as slop. On Amazon, new-book releases practically tripled from 2022 to 2025. (That’s in all probability not as a result of people out of the blue devoted themselves en masse to the written phrase.) Many are blatant copycats of current titles, seemingly meant to trick readers into shopping for them. Their gross sales figures are typically meager; the evaluations, indignant. A suspicious variety of them function protagonists with the identical names. Final 12 months, readers of 1 fantasy novel seen a line that gave the impression to be responding to a immediate asking a chatbot to put in writing within the fashion of one other writer. But when a ebook as beloved as Daggermouth had been written with a chatbot, it might increase an ungainly query: Might AI fiction be getting … good?
In a brand new examine, researchers randomly chosen greater than 14,000 Kindle ebooks and ran them by Pangram, an AI-detection instrument. One in 5 got here again as greater than 25 p.c AI-written; Daggermouth returned 60 p.c, in keeping with an underlying knowledge set that the authors shared with me. These numbers, they acknowledge, inform us solely a lot. The researchers grouped collectively textual content that Pangram had flagged as both “AI-generated” or “reasonably AI-assisted” (the latter of which might imply that AI was used to edit or polish human writing, or vice versa). And though Pangram has a low fee of false positives, it’s not excellent.
So the examine’s authors, led by Tuhin Chakrabarty, a professor of laptop science at Stony Brook College, devised one other verify. They scanned every ebook’s textual content for “uncommon expressions”—strings of not less than 5 phrases in a particular order—that additionally seem in different suspected AI books however virtually by no means in human-written ones. Daggermouth had lots of them. For example, the next sentence seems close to the top of a multipage intercourse scene in Chapter 27: “They collapsed collectively, a tangle of sweat-slicked limbs and racing hearts.” That very same sentence, phrase for phrase, additionally reveals up in one other self-published e-book by a distinct writer. Or think about this line from Chapter 38 of Daggermouth: “Her pulse thundered in her ears, drowning out the murmurs of the gang.” It’s additionally in an e-book by one other writer who has printed greater than 400 totally different titles on Amazon previously three years.
The examine has but to be peer-reviewed, and to be clear, there isn’t a definitive proof that Wolfe used AI. “The suggestion that I used generative AI to put in writing Daggermouth is wholly unfaithful,” Wolfe mentioned in a press release despatched by her lawyer. “I wrote this ebook myself. I’ve been outspoken about my opposition to generative AI and what it’s doing to writers, artists, and the artistic neighborhood. I don’t consider it belongs within the writing course of.” Referencing Pangram’s limitations, Wolfe added that “accusations like these have actual penalties for authors, they usually shouldn’t be made on the idea of know-how that’s identified to be unreliable.” Susannah Lawrence, a spokesperson for Daggermouth’s writer, Simon & Schuster, advised me in an electronic mail that the corporate stands behind the ebook, noting that it went by “the identical editorial and manufacturing course of as our different printed titles.” Neither Wolfe nor Lawrence addressed my questions concerning the “uncommon expressions” that Daggermouth shares with different latest ebooks.
Chakrabarty, who researches AI writing, advised me that Daggermouth has “too many telltale indicators” of AI for him to purchase that rationalization. He cited Chapter 18, “Thank You,” for instance: His analysis group broke it into 12 chunks for Pangram to guage, and the instrument labeled all components as AI-generated. (In my very own spot checks, the instrument labeled the primary few paragraphs as human-written and the remainder of the chapter as one hundred pc AI.) I ran this by Mohit Iyyer, a professor of laptop science on the College of Maryland who wasn’t concerned with the examine. An knowledgeable on AI-language detection, Iyyer advised me that even accounting for Pangram’s fallibility, it’s “virtually statistically unattainable” for a completely human-written ebook to register as 60 p.c AI. He added that the examine’s broader methodology, which connects AI writing to e-book gross sales, is sound.
Whether or not or not you consider that Wolfe deserves the good thing about the doubt, the identical questions on how Daggermouth was written apply to a rising variety of different hits on Amazon. Chakrabarty’s examine discovered that books written with “substantial” AI help made up 20 p.c of Amazon’s e-book catalog, 12 p.c of gross sales, and 10 p.c of finest sellers of their style.
But Daggermouth stands out for extra than simply its recognition. The second-best-selling e-book within the knowledge set is Inked Athena, by Nicole Fox, an writer who has 418 titles listed on Amazon. Daggermouth, in distinction, seems to be certainly one of simply two novels to Wolfe’s title in 2025, and its sequel shouldn’t be due out till 2027—a sign that Wolfe’s course of doubtless includes vital quantities of real human craft. Iyyer mentioned that utilizing AI in a means that leads to a novel that’s really gratifying to learn takes “an enormous quantity of effort by a human writer.”
One other distinction: Jennifer Prokop, who evaluations books for Kirkus and co-hosts Fated Mates, a preferred podcast about romance novels, advised me that folks in her world hold speaking about Daggermouth. “Nicole Fox, I’ve by no means heard a single soul speak about her,” she mentioned. (Fox didn’t reply to a request for remark.) Prokop sees mechanical prose cropping up principally in self-published ebooks, for which the dominant platform is Amazon’s Kindle Limitless subscription service. Simply because an e-book is charting on Kindle doesn’t imply individuals are devouring it. Grifters can now use AI to crank out titles by the dozen after which use all types of methods to entice readers, similar to shopping for tons of of bogus five-star evaluations and paying click on farms to juice their metrics. “If you will get 10,000 individuals to learn 10 pages of your ebook, and you’ve got 300 books you’re placing out,” she mentioned, you may make actual cash. “However you’re not really writing romances individuals wish to learn.” (Amazon didn’t reply to a request for remark.)
If Daggermouth is AI, then possibly it’s simply an outlier—a ebook that caught the proper wave of social-media hype. (A collection of controversies surrounding the ebook, together with accusations of the writer’s AI use and a backlash to the quilt artwork for a particular hardcover model, appeared to provoke its followers and produce it much more consideration.) The notion that AI writing requires an ideal confluence of occasions to high the charts is a comforting thought. However it may not be true. “Everybody thinks they’ll inform AI writing,” Christine Larson, a journalism professor on the College of Colorado at Boulder who research the impact of know-how on artistic work, advised me. “However I don’t assume we are able to inform a lot.” Privately, a number of authors and editors have advised her that they’re utilizing the know-how to speed up or enhance their output. “I completely purchase into the concept a human working with AI can write one thing that’s compelling,” she mentioned. “I’ve completely little question that we’ll see extra of this sooner or later.”
If she’s proper, then the query shouldn’t be about whether or not we’ll see AI finest sellers. It’s about whether or not anybody concerned with them will admit it. Coral Hart, a South African writer who makes use of AI to assist her publish tons of of books a 12 months underneath totally different pen names, advised me that she works arduous to maintain her pseudonyms secret after she obtained loss of life threats and 1000’s of one-star evaluations following a New York Occasions story about her. Hart claimed that she has had profitable authors be a part of the courses that she teaches on AI writing, although she declined to call them. “Till they cease the bullying, the witch hunts, and the ousting, I don’t assume authors are going to be clear, and I don’t blame them,” Hart mentioned. “I’m definitely not going to.”
The taboo in opposition to using AI in literature is comprehensible: Amongst different issues, it’s incorrect to take credit score for a ebook that you just didn’t write your self. Shunning authors suspected of using AI may work if there have been any solution to inform for positive—which there isn’t. That shaming ritual places the ethical burden on particular person authors as an alternative of on the mammoth firms accountable for the system by which they’re making an attempt to eke out a livelihood. It additionally tends to break down any distinction between pure AI slop and human writing that’s assisted by AI. The upshot is a state of affairs that feels untenable: AI writing will be profitable, however provided that it’s stored secret. Ultimately, one thing’s acquired to provide.
