Power users drive nearly all the fiction in a 500,000-prompt ChatGPT dataset

In a single session, Novaes moves between worlds she invented and worlds she borrowed, among them Star Wars, Pokemon and Mass Effect ^[12]. Some sessions end up as self-contained stories the length of a small novel, which she sometimes posts online ^[13]. "It can be never-ending. You find a quest, you complete it, you go for the next one," she told WIRED ^[14].

The two published counts of this behaviour use different denominators. The 7 percent from Melanie Walsh's team is a share of users, reached after controlling for heavy use ^[4]. The 1.4 percent in the 2025 study by OpenAI and the National Bureau of Economic Research is a share of prompts, it leaves role-play out, and OpenAI could not confirm to WIRED whether erotica is counted ^[7]. Seven divided by 1.4 is five, so the per-user estimate runs five times the per-prompt one ^[17].

Inside WildChat, the vast majority of fiction prompts came from a small collection of power users ^[2]. For anyone reading their own logs, that concentration matters more than the topline.

Every number in this record comes from prompts sent to ChatGPT ^[1][7], and none of them count the people running models on their own machines. That route appears here as one named person: Novaes runs models on her own machine through a program called SillyTavern ^[10].

Walsh's account of the appeal is about what the models lack. "These models are judgment-free; they can produce stories instantaneously; they are nearly inexhaustible," Walsh said ^[16]. Novaes described the same trait as a feature: "People say that AI hallucinates a lot. But in this case, it's probably what you want," she told WIRED ^[11].

The publishing fights have been about disclosure. One publisher canceled a $2 million book deal, another withdrew a novel from shelves, and readers alleged that a prize-winning short story was AI-generated ^[15]. The people in this study ordered the text themselves. Walsh, an assistant professor at the University of Washington Information School, said these readers know the stories are produced by an AI model and do not care ^[5][6].

Start with grain: the same dataset yields more than a third of interactions and 7 percent of users, and the distance between those two numbers is the heaviest accounts ^[1][4]. Then scope: OpenAI's 1.4 percent excludes role-play, which makes a count that includes role-play and a count that does not incomparable ^[7]. Demographics are not available either way: WildChat prompts are anonymized, and participants had to agree to make their prompts public ^[9].