UW study finds 7 percent of ChatGPT users write AI fiction
University of Washington researchers Melanie Walsh, an assistant professor at the UW Information School, and Neel Gupta, a PhD candidate at UW, analyzed more than 500,000 ChatGPT prompts drawn from the public WildChat dataset. They found that more than a third of chatbot interactions involved some form of fiction generation, including prose and poetry, fan fiction, erotica or role-play.
The researchers found that the vast majority of these fiction prompts came from a small group of power users who generate and tweak stories obsessively. One user, for example, prompted ChatGPT to produce versions of the same fan-fiction story set in the world of the anime video game Doki Doki Literature Club! thousands of times over several months. But even after controlling for this heavy use, the researchers estimate that 7 percent of chatbot users engage in fiction generation.
The WildChat dataset has limits: prompts are anonymized, so no demographic data can be captured, and participants had to opt in to making their prompts public, so the sample may not reflect ChatGPT's user base as a whole. OpenAI told WIRED that WildChat is not a representative sample of all ChatGPT users. A separate 2025 study by OpenAI and the National Bureau of Economic Research put the figure at just 1.4 percent of ChatGPT prompts being fiction-related, well below the UW estimate; OpenAI's figure excludes role-play, and the company could not confirm whether it counts erotica.
Walsh and Gupta argue the finding is relevant against the backdrop of a string of publishing-industry scandals over suspected AI authorship: one publisher canceled a $2 million book deal, another withdrew a novel from shelves, and readers alleged a prize-winning short story was AI-generated. Those episodes implied that readers do not want AI-generated fiction, but the researchers say that reaction has obscured what readers actually do: a committed group is generating custom, personal fiction and does not mind that it comes from a model.
WIRED profiles two of these users. Petra Ferraz de Novaes, a 36-year-old software engineer from Brazil, started using an AI model to build out her story ideas in 2024. She now runs open-source models locally through a program called SillyTavern, sketching a concept, such as an elven warrior battling the queen of the Amazons for her throne, and letting the model fill in the gaps. She moves between her own settings and borrowed ones like Star Wars, Pokemon and Mass Effect, sometimes producing novel-length stories she shares online, and otherwise taking part in private role-play. She says she was never much of a book reader before, but now reads other people's AI-generated stories as her only novels.
Razrien, a 40-year-old man from Indiana who spoke under his online username, uses AI role-play to explore specific kinks, casting chatbots as characters such as a 'big angry rat wife' or an eight-foot werewolf. He favors Google's Gemma 4 among the models he has tried, saying it suits his tastes well, and describes the appeal as a choose-your-own-adventure story with infinite swipes that lets him pursue fantasies more precise than what he could find in existing pornography or erotica.
Walsh says models offer qualities human writers do not: they are judgment-free, produce stories instantly and are nearly inexhaustible. Gupta says the pattern undercuts an idealistic view of fiction as communication between writer and reader, suggesting that for many readers it is a simpler economic exchange focused on getting entertainment. The article notes current AI models still struggle with ornate prose, complex characters and stylistic flourishes, do better with genres built on familiar tropes, and cannot sustain coherent narratives at book length, though the researchers expect some of these limits to ease as models improve. Walsh floats a possible tiered future for publishing: AI-generated and self-generated fiction serving the median reader, alongside a smaller stream of human-authored fiction for readers seeking formal experimentation.
Key facts
- A UW study of over 500,000 ChatGPT prompts from the WildChat dataset found more than a third involve fiction, fan fiction, erotica or role-play.
- Controlling for a small set of obsessive power users, researchers estimate 7 percent of chatbot users engage in fiction generation.
- A separate 2025 OpenAI/NBER study found only 1.4 percent of ChatGPT prompts are fiction-related, a figure that excludes role-play.
- The study comes amid publishing scandals over suspected AI authorship, including a canceled $2 million book deal and a withdrawn novel.
- Profiled users include a 36-year-old Brazilian engineer who builds fantasy stories with local open-source models via SillyTavern, and a 40-year-old man who uses Google's Gemma 4 for personalized erotic role-play.
Why it matters
Recent publishing scandals over suspected AI-written books have carried an implicit assumption: that nobody wants machine-generated fiction and that using AI for it devalues the art. This study, based on real ChatGPT usage rather than industry reaction, suggests a sizable and largely invisible readership already consumes and produces AI fiction on its own terms, and does not mind that it comes from a model.
Who it affects
It affects the people writing and role-playing with chatbots for personal entertainment, from hobbyist writers building fantasy worlds to people pursuing specific erotic role-play, as well as publishers and authors navigating a market where AI-authorship suspicions can sink a book deal or pull a title from shelves.
How to use it
The profiled users describe two approaches: one runs open-source models locally through the program SillyTavern for full control and privacy, while the other tries different models to match specific kinks. Razrien singles out Google's Gemma 4 as his preferred model for a specific kind of character-driven role-play, though the article gives no pricing or subscription details for any of these tools.
How solid is it
The two figures at the center of the piece disagree sharply: the UW team's 7 percent estimate versus a separate 2025 OpenAI/NBER study's 1.4 percent, with OpenAI's number excluding role-play and its coverage of erotica unconfirmed. The WildChat dataset itself is anonymized, carries no demographic data, and only includes users who opted to make their prompts public, so it may not represent ChatGPT's broader user base; OpenAI itself told WIRED as much.
Risks and caveats
The article raises a concern that AI fiction generation could narrow some readers' exposure to writing that would otherwise stretch them. The article also notes today's models still struggle with ornate prose, complex characters and book-length coherence, doing best with genres that lean on familiar tropes, though the researchers expect some of these gaps to close as models improve.
“These models are judgment-free; they can produce stories instantaneously; they are nearly inexhaustible.”
— Melanie Walsh, University of Washington Information School