Tuhin Chakrabarty is a computer scientist at the State University of New York at Stony Brook who studies how to detect AI-generated text. His research drew global attention earlier this year after a journalist cited it as supporting evidence to claim that significant portions of the bestselling 2026 novel Daggermouth were produced using artificial intelligence — an allegation that the book’s author, H. M. Wolfe, has denied. Nature spoke to Chakrabarty about his work, why it matters and what the rise of AI means for the future of creativity.
How did you become interested in detecting AI-generated writing?
I live in New York City, and I’m surrounded by creative people. I can understand why artists are enraged about AI. It can take years to write a book, but an AI chatbot can churn one out in 15 minutes. That got me interested in the subject of copyright and creativity. As AI use spreads, the future is going to get messier. What qualifies for copyright protection? How do you quantify the original human contribution when something is co-created with AI? How can you even tell the two apart?
What should count as AI-written text?
Disclosure is important. At scientific conferences, we are seeing increasing numbers of disclosures stating that AI was used and for what — to polish writing, say, or for copy-editing. Once you disclose, things become easier to judge.
Can Anthropic’s invisible watermarks curb ‘AI slop’? Researchers remain sceptical
Over time, you could imagine separate markets developing for text written purely by people, hybrid writing and mass-market AI-written prose. That might create better economic incentives for people who don’t use AI or are morally opposed to its use. Everyone’s process is different, and there is a huge spectrum in how AI can be used. My moral concern relates to people who use AI, but pass off AI-written prose as their own. That is where the problem lies.
How can AI-written prose be detected?
In a July preprint, my colleagues and I traced relatively unusual sequences of words that appear verbatim in different books (T. Chakrabarty et al. Preprint at arXiv https://doi.org/rsp2; 2026). We wanted to measure how much language books derive from books that came before them.
We compared the number of these rare expressions across three types of book for sale on Amazon. We looked at 200 titles flagged by the AI-detection software Pangram as containing substantial amounts of AI-written text, 200 self-published bestsellers with no AI-written text, and 200 literary fiction books that had won or been nominated for prizes. In each case, we identified rare expressions that appeared, word-for-word, in a sample of hundreds of thousands of titles on Google Books.
We found that literary fiction contained the least overlap with existing books, writing by self-published authors contained more and the AI-written books contained the most.
This suggests that AI models will find it easiest to automate more formulaic writing. A lot of self-published genre fiction uses tropes, clichés and formulae. These are relatively easy for AI to produce. Literary fiction requires a lot of originality and subtext. In work in progress, we divided books into narrative beats — the individual events or plot points that make up a story. There, too, we found that AI-generated books copy specific beats from earlier books written by people.
Why does AI-written text use words or expressions from existing books?
Large language models are trained on enormous amounts of human writing, including books, and sometimes specific sequences of language resurface in what they generate. We also find AI-written books containing expressions that occur in other AI-authored books. It isn’t necessarily that the model is deliberately copying a sentence from a particular book. Its output is a function of the material it was trained on, and sometimes that produces these unusually specific overlaps.
Couldn’t human writers repeat expressions from the published record, as well?
Absolutely. Two writers could reasonably produce an overlap of five or six words independently, particularly in genre fiction, because of the way it relies on formulae. But as an expression gets longer and more specific, it becomes harder to explain through coincidence. Many of the expressions we found in genre fiction were imagery, metaphors, clichés and hackneyed language. I can see why it is easier for AI to borrow those things, because if you’re writing romance or science fiction, you are also conforming to genre norms.
Your work has been used to make claims about particular books. How do you feel about that?
In our study, we chose not to name authors. As academics, we have little power. If someone brought a defamation suit against me or the university where I work, as a pre-tenured faculty member, that would be difficult to deal with.
But when a journalist is writing about your work, especially something as explosive as this, you have to share the underlying data. I shared them with The Atlantic. I was comfortable with that because I could stand by my results, and I trusted the reporter to do his due diligence.
The author of Daggermouth emphatically denies using AI. Why do you stand by your analysis?
The book contained specific rare expressions. One was: “They collapsed together, a tangle of sweat-slicked limbs and racing hearts.” That entire expression also appears in another book. It is too long and specific to occur in exactly the same way by coincidence. We found several examples like that.
Could AI help you to write your next paper?
Our checks with the AI-detection tool Pangram also flagged more than half of the book as AI-generated (see the results here: go.nature.com/4dghtvg). I don’t think the book is 100% AI-generated, but I stand by our conclusion that it substantially used AI.
[H. M. Wolfe did not respond to a query from Nature. But, in a statement made to The Atlantic through her lawyer, she said it was “wholly untrue” that she used generative AI to write Daggermouth. Both the US publisher Simon & Schuster and the UK publisher Tor Bramble stand by the author and the book.]
Could AI also be used creatively in literature?
Absolutely. Brilliant writers have used AI. Searches (2025) by Vauhini Vara is a fascinating example of non-fiction that engages with AI-generated text. Vara uses AI responses in the book in a way that interrogates what the technology is doing. Sheila Heti also experimented with AI in her 2023 New Yorker story ‘According to Alice’. Heti conversed with a customizable chatbot on the Chai AI platform, then removed her own questions and edited the responses into the published story.
I don’t think there is necessarily a stigma about using AI. There is a stigma about using AI and hiding it. If you use AI in a profound way that adds new meaning to literature and art, why not? If an accomplished writer uses it, discloses it and produces something that wows people, I’m sure publishers will publish that.