University of Washington Finds Fiction in a Third of Sampled Chatbot Chats

For a subset of chatbot users, the appeal is not a polished novel or a one-off answer. It is a story that can change direction on demand, continue indefinitely, and cater to tastes conventional publishing may never serve. A University of Washington analysis of more than 500,000 public chatbot prompts found that more than one-third of sampled interactions involved fiction generation.

The category was broad: prose, poetry, fan fiction, erotica, and role-play all counted. The researchers, Melanie Walsh and Neel Gupta, examined prompts from WildChat, a public dataset of ChatGPT conversations, to understand a behavior often obscured by debate over whether professional authors use AI.

The story is the product

WIRED’s interviews show what that interaction can look like. Petra Ferraz de Novaes, a software engineer in Brazil, uses SillyTavern and open-source models to build fantasy stories and private role-play scenarios. She sometimes produces self-contained stories the length of a small novel; in other sessions, she moves among worlds of her own invention and existing franchises.

A small group writes a great deal

That concentration complicates the headline figure without erasing the pattern. After controlling for heavy use, the team estimated that 7 percent of chatbot users engage in fiction generation. The estimate suggests custom storytelling is not only an extreme hobby, but it does not establish how frequently those users create fiction or what share of their overall chats it represents.

A large signal with a narrow window

That means the research is better read as evidence of a visible, highly engaged practice than as a census of chatbot behavior. Its importance lies in what the practice reveals: some people want fiction not simply to be written for them, but to be continually negotiated with them.

Personal control does not solve the craft problem

For now, that gap helps explain why chatbot fiction may function differently from traditional publishing. The model’s value can be less about literary finish than about responsiveness: it offers a reader control over the premise, pace, characters, and next turn. Whether that leads people toward more reading, away from unfamiliar writing, or simply into a separate entertainment category remains unresolved.