More than a third of the conversations in a public archive of 500,000 chatbot prompts involve making something up: prose, poetry, fanfiction, erotica, roleplay scenarios. That is the result of a study by researchers at the University of Washington, who worked through the open WildChat dataset. Set aside the small group of power users who generate and rework stories obsessively, and the researchers still put the share of people writing fiction with a chatbot at 7%. A 2025 study by OpenAI and the National Bureau of Economic Research arrived at 1.4%.
The finding lands on an industry that has spent its recent arguments on the supply side. One publisher cancelled a book contract worth $2 million; another pulled a novel from sale; readers claimed that a story which had won a literary prize was written by AI. Each case rested on the same assumption, that readers do not want machine-made text and that using AI cheapens fiction. Melanie Walsh, assistant professor at the University of Washington Information School and a co-author of the study, argues that the noise pulled attention away from what readers were themselves doing. Her point is that the people in the data know perfectly well a model wrote the story, and that for many of them this is part of the appeal.
Novaes is one of them. She had always liked telling stories, but making every decision herself — style, syntax, how a character develops — felt like more than she could carry. In 2024 she asked an AI model to render one of her ideas, and the model's odd and sometimes unpredictable readings of her prompts turned out to be the thing that held her. What gets called hallucination elsewhere is, in this format, the desired output.
She now spends long stretches assembling stories with a model. She sketches the outline — an elf warrior fights the queen of the Amazons and takes her throne — and the model fills the gaps. In a single session she moves between several fantasy worlds, some of her own invention, others lifted from Star Wars, Pokémon and Mass Effect. Sometimes the result is a finished piece the length of a short novel, which she posts online. Sometimes it is private roleplay that never has to end: one quest gives way to the next, and the only limit is what the user can imagine.
The 7% figure is the conservative one because the raw distribution is lopsided. Most of the fiction prompts come from a narrow band of heavy users. In one case a single person spent months asking ChatGPT, thousands of times over, for versions of the same fanfiction set in the world of the anime video game Doki Doki Literature Club!
WildChat has limits the authors acknowledge. The prompts are stripped of identity, so there is no demographic breakdown. Participants had to agree to have their prompts published, which may skew who ends up in the sample. OpenAI told WIRED that WildChat is not a representative sample of ChatGPT users.
OpenAI's own number is the more revealing artefact. The 1.4% it published with NBER did not count roleplay at all, and the company did not say whether erotica was included. Those two exclusions are not edge cases in this data — they are a large part of what people appear to be doing.
This is where the study is most useful, and it is not really a finding about literature. It is a finding about measurement. The gap between 1.4% and 7% is not a dispute about facts; it is a dispute about the definition of fiction, and the party with the complete logs is also the party with an obvious interest in a narrow definition. A company negotiating with publishers, regulators and advertisers has very little to gain from a headline saying that a meaningful slice of its traffic is erotic roleplay. An outside research team assembling its estimate from a consent-based public dataset has the opposite bias and a much worse sample. Neither number should be read as the truth; the interesting part is that the honest range is wide enough to change what ChatGPT is understood to be for.
The behaviour at the top of that range is vivid. Rasrien, a 40-year-old in Indiana who asked to be identified by his online name, describes it as a choose-your-own-adventure book with an unlimited number of moves. He builds characters who dominate him in roleplay, among them a big bad rat wife and an eight-foot werewolf woman with a bad temper. He has worked out which models suit which fetishes; his current all-purpose favourite is Google's Gemma 4. The appeal, as he describes it, is specificity: rather than settling for the nearest available thing in online pornography or published erotica, he gets the exact fantasy. He is not dating anyone and finds the arrangement suits him. He is attached to the rat wife, even though in the story she occasionally tears him apart and then heals him, which is how she expresses love.
Walsh's explanation for the pull is structural rather than psychological. Models have properties human writers do not: they do not judge, they produce instantly, and they almost never run out. Some users, the researchers suspect, are drawn by the ability to rewrite the plot quickly and repeatedly; others by the control, which relieves them of an author's decisions. Neil Gupta, a doctoral student at the University of Washington and a co-author, reads it as evidence that for a lot of readers, literature can be a simpler transaction in which a person gets the entertainment they came for.
What models still cannot do sets the current boundary. They struggle with ornate prose, complicated characters and distinctive style, they are best at genres built on familiar plot machinery and tropes, and they cannot hold a coherent narrative across the length of a book. The researchers expect some of that to erode as models improve, and sketch a tiered outcome: a large volume of AI-written books and personal stories serving most readers, and a small share of human-written fiction serving a niche appetite for formal experiment and ambiguous narration. Walsh notes that English literature academics, teachers and part of the publishing world have been a niche audience for a long time already.
There is a cost the study names directly. A reader who specifies their own story is less likely to run into the text that pushes them outside their existing tastes. And there is a symmetrical gain that cuts against the panic in publishing: Novaes barely read books before this. Now she seeks out other people's AI-generated stories, and those are the only fiction she reads. The fear was that machines would take readers away from writers. The data describes something narrower and stranger, which is an audience that was never reading in the first place.