Researchers at the University of Washington and the University of Colorado Boulder analyzed 573,453 real-world ChatGPT-style conversations and found that a small group of heavy users accounted for most of the fiction-writing activity. In the fiction subset, just 2% of users were responsible for more than 80% of the conversations.
The finding appears in the paper AI Fiction in the Wild, uploaded to arXiv on June 22, 2026, under identifier 2606.22748. The authors are Neel Gupta and Melanie Walsh of the University of Washington Information School, and Maria Antoniak of the Department of Computer Science at the University of Colorado Boulder. The paper was previously presented at Purdue University’s MFS Cultural AI workshop in September 2025 and is slated to appear in the literary journal MFS: Modern Fiction Studies.
A single scene repeated thousands of times
One example in the paper centers on an anonymous user who repeatedly asked the model to continue nearly the same scene over the course of several months. The prompt describes a high school clubroom, with characters including Natsuki, Sakura, and club president Monika, and revolves around a childbirth scene. The text cuts off mid-sentence, leaving the model to take over. The same or nearly identical setup was then pasted into the chat box again and again, eventually more than a thousand times.
That passage drew wider attention in early July after reports by Japanese outlet AUTOMATON and gaming publication Dexerto. Screenshots then spread widely on X.
Coverage often described the user as an “extreme outlier,” but the paper does not use that phrase. The researchers instead refer to a “particularly prolific outlier” and introduce their own label, “infinite story demander,” for users who repeatedly request the same or highly similar story beats over long periods.
34% of the dataset involved fiction generation
The dataset used in the study is WildChat, which contains 573,453 English-language conversations. After classification, 195,271 of them were labeled as fictional content generation, equal to 34% of the full set.
Within those roughly 195,000 fiction-related conversations, fan fiction made up 95,450 chats, or 49%, while erotic fiction appeared in 52,231 chats, or 27%. The paper says those categories overlap substantially.
The most common fictional universe was Doki Doki Literature Club!, with 22,381 conversations, or 11.5% of the fiction subset. It was followed by Freedom Planet with 5,204 chats, or 2.6%, then League of Legends with 4,514 and Naruto with 4,342.
The authors note that this ranking looks nothing like the popularity charts on Archive of Our Own, or AO3, the largest fan fiction platform. AO3 is dominated by large Western franchises such as Harry Potter, Star Wars, and Marvel. WildChat showed a very different pattern.
2% of users produced more than 80% of fiction conversations
The paper’s most striking line is simple: in the fiction subset, 2% of users were responsible for more than 80% of the conversations.
The unit matters. The paper measures conversation counts, not interaction volume, message counts, or token counts. That distinction is important because conversation totals are closer to the kind of usage figures AI companies often publish.
To see whether heavy users were repeatedly asking for the same material, the researchers calculated a prompt repetition rate, defined as the share of a user’s prompts that fell into the same semantic cluster.
- Regular users with at least two conversations had a repetition rate of 42%
- The top 2% of heavy users reached 69%
- The 10 most prolific users reached 85%
That points to a narrow group generating large volumes of fiction while returning to the same themes over and over.
The data came from a free GPT interface, not OpenAI’s internal logs
The source of the dataset also limits what the paper can claim. WildChat was collected by the Allen Institute for AI, or AI2, between April 9, 2023, and May 1, 2024.
The data did not come directly from OpenAI. Instead, AI2 hosted free GPT-3.5 Turbo and GPT-4 chat interfaces on Hugging Face Spaces. Users who clicked through two consent pop-ups could access the models without logging in and without rate limits, with the trade-off that their conversations could be made public.
The methods section states clearly that WildChat is not a representative sample of all ChatGPT users. The researchers suggest these users may have been more technically literate and more online than average. They may also have come from lower-income regions, from countries where ChatGPT was blocked, or from groups interested in probing model limits and generating content restricted on mainstream platforms.
Fiction fell from 34% to 7.1% after user-level deduplication
The researchers then changed the counting method. Using hashed IP addresses to estimate identities, they reduced 12,947 IPs to roughly 10,082 users and kept only one conversation per user.
With that adjustment, fiction’s share of the dataset dropped from 34% to 7.1%.
Same dataset. Same question. Different counting method. The answer changed by nearly a factor of five.
The paper adds in a footnote that once it moved to a one-user, one-chat view, Doki Doki Literature Club! fell sharply in the rankings. It was replaced by more mainstream works such as Game of Thrones and Pokémon.
That means the impression that ChatGPT users were especially interested in writing Doki Doki Literature Club! fan fiction was, to a large extent, the product of repeated activity by a small number of users.
How the paper interprets repeated prompting
The researchers do not frame the anonymous user simply as a joke or an anomaly. They place that behavior in the psychology of repeated consumption: rewatching the same film, revisiting the same city, or returning to the same museum. Some work in that area argues these experiences are not as repetitive as they first appear.
The paper also borrows the role-playing game term “re-rolling,” where players repeatedly restart until they get the stats they want. The authors observed that this user would often take a model response they liked, fold it back into the next prompt, and start again, suggesting that one version had finally landed close to the desired result.
The paper describes this pattern as a form of storytelling in which every ending generated by a large language model is slightly different, and the same story is remade into a new version each time.
In that framework, the user is not treated as a simple exception. The authors describe this as an extreme form of a new kind of reader, a “solipsistic reader-writer,” where one person both produces and consumes the story inside a closed conversational loop, with no other human on the other side.
Under that view, writing and reading can collapse into the same role when fiction is generated with AI.
The broader point: usage totals can misstate user behavior
The paper ends on a larger measurement problem. Claims such as “AI is changing human creativity,” “AI-generated content is surging,” or “a platform generated hundreds of millions of pieces of content this month” are not necessarily false. But they often leave out the same missing subject: who is generating that output, and how many times.
In this dataset of more than 573,000 conversations, 34% being fiction-related can sound like a major cultural shift. Once each user is weighted more evenly, that share shrinks to 7.1%.
The study does not argue that AI fiction is unreal. The researchers explicitly say they believe similar heavy users are likely also generating fiction compulsively on ChatGPT, Claude, and Gemini.
The point is narrower and more structural. Using usage volume to describe users can produce a distorted picture when a small group accounts for a huge share of activity. The paper argues that many of the headline metrics released across the AI industry rely on exactly that kind of measurement.
This report is based on the arXiv paper AI Fiction in the Wild and BlockTempo’s coverage.