A Third of ChatGPT Prompts Are Stories. The People Writing Them Are Fine With That.

A new study of 500,000 real ChatGPT conversations finds fiction is far more common than the publishing industry's AI panic suggests.

AI2Day NewsdeskEditor: Lee Brown5 min read
A glowing smartphone screen showing a grid of colourful Instagram-style photo thumbnails, surrounded by soft abstract digital light patterns suggesting data flo
Share

Key points

  • A study of more than 500,000 ChatGPT prompts found that over a third of interactions involved fiction, including prose, fan fiction or role-play.
  • Researchers estimate 7 percent of chatbot users actively generate fiction; a separate 2025 OpenAI study puts the figure at 1.4 percent, excluding role-play.
  • A single user prompted ChatGPT to generate variations of the same fan-fiction story thousands of times over several months.
  • Current AI models struggle with ornate prose and book-length plot coherence, though researchers expect those limits to shrink.
  • Participants valued AI storytelling for being judgment-free, instant, and effectively inexhaustible.

Petra Ferraz de Novaes, a 36-year-old software engineer from Brazil, had always wanted to tell stories. The problem was the sheer volume of decisions involved: choosing a style, building a character, steering a plot. So in 2024 she asked an AI model to do it with her.

She hasn't stopped since.

What are people actually doing?

Novaes uses a program called SillyTavern to run open-source AI models directly on her own computer. She sketches a premise, an elven warrior seizing the throne of the Amazon queen for instance, and the model fills in the rest. In a single session she might hop between several fantasy worlds, some original, some borrowed from Star Wars or Pokémon. Some sessions produce self-contained stories the length of a short novel.

"The limit is your imagination," she told Wired AI.

Researchers at the University of Washington studied 500,000-plus prompts from WildChat, a public dataset of real ChatGPT conversations. They found more than a third of interactions involved some form of fiction. Most came from a small group of heavy users: in one case, a single account prompted the chatbot to generate variations of the same fan-fiction narrative, set inside the anime video game Doki Doki Literature Club!, thousands of times across several months.

Controlling for those power users, the team estimates 7 percent of ChatGPT users engage in fiction generation. A 2025 study by OpenAI and the National Bureau of Economic Research put the share of fiction-related prompts at 1.4 percent, though that figure excludes role-play and may not count erotica.

WildChat has real limits. Prompts are anonymised, so no demographic data exists, and participants opted in to making their chats public, which means the group probably skews toward enthusiastic users. OpenAI told Wired that the dataset "is not a representative sample of all ChatGPT users."

Why does AI storytelling appeal to people?

AI models are available at any hour, never tire, never judge a request, and can produce a new scene in seconds. For readers who want to steer the action rather than follow an author's decisions, that matters.

Erotica readers in particular can write themselves into hyper-specific fantasies that published fiction rarely covers. Razrien, a 40-year-old man from Indiana who gave only his online username, uses AI to build elaborate role-play scenarios with characters he describes in detail. His preferred model is Google's Gemma 4, an open-source AI that runs without an internet connection. "I have a thing for large, angry women slapping me around. Some models are perfectly good for that," he told Wired AI.

"I'm not dating at the moment. So, hey, this works."

What are the real limits right now?

Today's AI models are noticeably better at genre fiction built on familiar story shapes than at the kind of prose you'd find in a literary novel. Complex characters, stylistic originality and book-length plot coherence all remain weak spots. Melanie Walsh, an assistant professor at the University of Washington Information School who co-authored the study, thinks these are real constraints, not merely temporary ones, though she expects them to shrink as models improve.

Publishing has spent months consumed by scandals over authors suspected of using AI without disclosure: cancelled book deals, withdrawn novels, accusations over prize-winning short stories. AI2Day's earlier reporting on the 24-person Brooklyn startup that publishers are using to detect AI writing shows how high the commercial stakes have become. Walsh argues those scandals say little about what readers actually want. The data suggests a committed group genuinely prefers custom AI-generated fiction to anything on a shelf.

For Novaes, AI storytelling didn't replace reading. It created it. "I was never much of a book reader," she said. "But now I do read other people's stories. These are the only novels I'm reading."

That's the finding worth sitting with: for at least some users, AI fiction is an on-ramp, not a shortcut.

Common questions

Is it legal to generate fiction with AI, including adult content?

Generating personal fiction with AI is broadly legal in most countries, but platforms have their own rules about adult content, and some open-source models are fine-tuned specifically to remove those restrictions. Check the terms of whichever tool you use.

Does AI-generated fiction affect professional authors?

This research focuses on personal, often private use rather than commercial publishing. The concern for working writers is mainly in the commercial market, where AI-generated books published for sale compress prices and crowd catalogues, not in hobbyist role-play.

How accurate are the usage estimates in the study?

The 7 percent figure comes from a self-selected dataset and almost certainly overstates general use. The 1.4 percent OpenAI and National Bureau of Economic Research figure is narrower because it excludes role-play. The true share sits somewhere between the two, and neither number should be treated as settled.

© 2026 AI2Day