A study of more than 500,000 prompts fed into ChatGPT found that over a third of chatbot interactions involve some form of fiction generation — prose, poetry, fan fiction, erotica or role-play. The data comes from WildChat, a public dataset analyzed by researchers at the University of Washington Information School. The finding matters because it contradicts the assumption, widespread after several publishing scandals, that readers have no appetite for machine-written stories.
What the WildChat analysis shows
The researchers found that the vast majority of fiction prompts come from a small collection of power users who generate and tweak stories obsessively. In one case, a single user asked ChatGPT to produce versions of the same fan-fiction narrative set in the world of the anime video game Doki Doki Literature Club thousands of times over several months. Even after controlling for heavy use, the team estimates that 7 percent of chatbot users engage in fiction generation. The dataset has limits: prompts are anonymized, so no demographic data can be captured, and participants had to agree to make their prompts public. OpenAI told WIRED that WildChat is not a representative sample of all ChatGPT users, and a separate 2025 study by OpenAI and the National Bureau of Economic Research puts fiction-related prompts at 1.4 percent — a figure that does not include role-play.
Melanie Walsh, an assistant professor at the University of Washington who coauthored the study, says users know the stories are produced by an AI model and do not care — in fact, that appears to be a large part of the attraction. The mechanics differ from ordinary reading. A user sketches a concept, such as an elven warrior battling the queen of the Amazons and seizing her throne, and the model fills in the gaps, carrying the narrative in unexpected directions. Petra Ferraz de Novaes, a 36-year-old software engineer from Brazil, runs open-source models on her own computer through a program called SillyTavern and switches between several fantasy worlds in a single session, including Star Wars, Pokémon and Mass Effect. She says the unpredictability is what fascinates her: what people call hallucination is, in this case, often what the user wants.
The study lands in the middle of a run of publishing scandals over the suspected use of AI by professional authors. One publisher canceled a $2 million book deal, another withdrew a novel from the shelves, and readers alleged that a prize-winning short story was AI-generated. The implicit assumption in each case was that nobody wants to read machine-generated stories and that using AI devalues the art form. Walsh and her colleagues argue that the commotion has distracted from what readers actually do, and from what that behavior signals about demand for AI fiction. Their explanation for the demand is practical: the models are judgment-free, produce stories instantaneously and are nearly inexhaustible, which suits people who want to iterate quickly or direct the action themselves instead of accepting an author's decisions.
What this means for business
For companies that build or buy AI tools, the practical consequence is that fiction and role-play are not a marginal workload but a core consumer use case, and one that behaves differently from search or summarization. It is long-session, iterative and highly sensitive to model personality, which changes what vendors must optimize: memory within a session, tolerance for repeated regeneration of the same scene, and freedom from refusals. A small studio can run open-source models locally, as Novaes does, and keep costs and content policy in its own hands; a larger company integrating a hosted API inherits the provider's moderation rules and cannot tune them per user. The same mechanics apply to internal training simulations and scenario role-play, where staff rehearse negotiations or customer conversations.
What the news does not mean is that AI fiction has become a mass market on the scale of the headline figure. The two estimates differ by an order of magnitude — over a third of interactions in WildChat against 1.4 percent of prompts in the OpenAI and NBER study — and the gap is explained partly by methodology and partly by what counts as fiction. WildChat overrepresents people willing to publish their prompts; the OpenAI figure excludes role-play and the company could not confirm whether it counts erotica. A business evaluating the segment should ask the vendor which workloads the usage data covers, how sessions are counted, and whether role-play and adult content are included or filtered out. The researchers also note that current models struggle with ornate prose, complex characters and stylistic flourishes, fare better with well-trodden story arcs and tropes, and fail to hold a coherent narrative at book length.
The marker to watch is whether the usage gap narrows as measurement improves: if independent datasets and provider telemetry converge on a similar share of fiction and role-play traffic, the segment is a stable consumer category rather than an artifact of one public dataset. A second signal is product behavior — whether model releases start advertising session memory, character consistency and lower refusal rates as headline features, which would confirm that vendors treat this audience as a paying market. For business, that would mean the same capabilities arrive in enterprise tools for training, simulation and customer-facing agents.
