50,000 ChatGPT conversations, with 34% of the dialogue volume used to write fan erotic novels

A research team from Washington University and the University of Colorado Boulder analyzed 573,000 real ChatGPT conversations, trying to figure out how humans write novels with AI. What they found was an anonymous user who, over the course of several months, asked the model more than a thousand times to keep writing the same novel scene. Even more crucially, in the accounts that used GPT to write novels, just 2% of people produced over 80% of the dialogue volume.

(Background: Is the human brain better than machines? Tens of thousands of people on websites pretending to be ChatGPT)

(Additional context: Deezer warns that 44% of newly uploaded music is AI-generated, and human creators’ hard-earned money is being collectively looted)

Table of Contents

Toggle

• This strange phenomenon appears in the paper
• One-third of the conversations are writing novels
• 2% of people write 80% of the conversations
• These 573,000 conversations were scraped from free GPT
• 34% drops to 7.1%
• Writing the “having a baby” scene over a thousand times
• Every sentence that says “AI is reshaping creation” is missing a subject

In a school club classroom, a female high school student named Natsuki clutches her stomach and says, “Sakura, right at this time—you’ve chosen to be born here. Can’t you wait two more months?”

Club president Monika asks her what’s wrong. She grits her teeth and answers, “My water broke. Sakura’s coming out—and I have no idea what to do.”

Then this script cuts off in the middle of a sentence. “My body… it—” The sentence stops right there, with no continuation.

This is not the original plot of the game. It’s a piece of text someone typed manually and pasted into a ChatGPT chat box. The punctuation break is deliberate—that’s where the model is supposed to take over.

Then this person pastes almost identical stuff again. And again. Again and again, for months—over a thousand times.

This strange phenomenon appears in the paper

The dialogue above is fully included in the paper “AI Fiction in the Wild” (ID 2606.22748) uploaded to arXiv on June 22, 2026. The authors are Neel Gupta and Melanie Walsh from the University of Washington’s Information School, as well as Maria Antoniak from the Computer Science department at the University of Colorado Boulder. The paper was presented in September 2025 at Purdue University’s MFS Cultural AI symposium, and is expected to be published in the literary journal “MFS: Modern Fiction Studies.”

In early July, Japanese outlet AUTOMATON and esports media Dexerto reported the scene in the paper almost at the same time. Screenshots then went viral on X immediately, becoming one of the biggest tech memes of July.

First, let’s correct one thing: reports in both Chinese and English generally call this user an “extreme outlier,” but the full text of the paper never uses that term. The researchers’ actual wording is “particularly prolific outliers,” plus a category they invented: “infinite story demander,” referring to people who repeatedly request the same, or highly similar, story scenes over a long period of time.

“Doki Doki Literature Club!” is a visual novel game released in 2017. On the surface it’s a wholesome campus romance theme; playing it reveals it’s a psychological horror work that plays with meta-narrative. Natsuki, Monika, Yuri, and Sayori are the four members of the Literature Club. The game isn’t especially well known in Taiwan, but in the English internet it’s a stronghold of fan creation.

One-third of the conversations are writing novels

The dataset used by the research team is called WildChat, consisting of 573,453 English dialogue exchanges in total. After classification, 195,271 entries involve fictional content generation, accounting for 34%.

In these 195,000-plus “novel-writing” conversations, fan-created works make up 95,450 (49%), while erotic novel content accounts for 52,231 (27%). The two overlap heavily.

And for the most popular work world/setting, number one is “Doki Doki Literature Club!,” with 22,381 conversations, or 11.5% of the novel subset.

Second is the action game “Freedom Planet,” with 5,204 conversations, or 2.6%. Next are League of Legends with 4,514, and Naruto with 4,342.

The paper specifically points out that this ranking is completely different from the hottest works on AO3, the world’s largest fan platform. On AO3, the top items are mainstream Western IPs like Harry Potter, Star Wars, and Marvel. WildChat’s list looks totally different.

2% of people write 80% of the conversations

There’s a very dramatic sentence in the paper.

“Within the novel subset, 2% of users are responsible for over 80% of the conversations.”

To be precise: the circulated version of the material says “2% of accounts contribute over 80% of interaction volume,” but the paper’s unit is the number of dialogues, not interaction volume, message count, or token count. This distinction matters a lot, because dialogue count is the kind of metric the AI companies most commonly publish.

So what are these 2% doing? The researchers calculated a “prompt repetition rate,” measuring what proportion of the same user’s prompts fall into the same semantic cluster.

For typical users (with at least two conversations), the repetition rate is 42%.

For the top 2% of heavy users, 69%.

For the top ten most prolific users, 85%.

These 573,000 conversations were scraped from free GPT

This is the easiest part to misread—and it determines how far the paper’s conclusions can be inferred.

WildChat was collected by Allen Institute for Artificial Intelligence (AI2), from April 9, 2023 to May 1, 2024. The collection method was not to obtain data from OpenAI. Instead, free GPT-3.5 Turbo and GPT-4 chat interfaces were set up on Hugging Face Spaces. Users only had to click the two consent pop-ups to get access to free usage without logging in and without rate limits, with the trade-off that chat records might be made public.

In its methodology section, the paper is very clear: WildChat is not a representative sample of all ChatGPT users. The researchers speculate these people “understand technology better than average users and live more on the internet,” and they may also come from low-income regions, from countries where ChatGPT has been banned, or simply be people trying to test model boundaries and generate content that the platforms explicitly prohibit.

34% drops to 7.1%

Next, the team did something simple but with a harsh result.

They inferred user identity using hashed IP addresses, and 12,947 IP addresses converged into roughly 10,082 users. Then, for each user, they kept only one conversation.

The proportion of novel-related content dropped from 34% to 7.1%.

With the same batch of data and the same question, swapping the algorithm changes the answer by nearly five times.

Not just the total. In a footnote, the paper adds another blow: after one conversation per person, the ranking of “Doki Doki Literature Club!” “plunged sharply.” In its place, works like “Game of Thrones” and “Pokemon,” truly mainstream titles, climbed onto the list.

So the conclusion that “ChatGPT users most love writing fanfiction for ‘Doki Doki Literature Club!’” is itself an illusion created by repeated behavior from just a small number of people.

Writing the “having a baby” scene over a thousand times

It’s worth noting that the paper’s attitude toward this anonymous user isn’t mockery.

The researchers place this behavior in a psychology context of “repetitive consumption”—watching the same movie again, visiting the same city again, revisiting the same art museum again. Some studies even suggest that this kind of experience “isn’t as repetitive as people think.”

They also borrowed a role-playing game term: “re-rolling.” Players repeatedly restart until the numbers come out satisfactory. The researchers observed that after a user got a version they liked, they typically took the model’s freshly generated reply and fed it back into their own prompt, then restarted again—meaning they had finally found that version worth sticking with.

The paper gives a sharp description of his narration: “In stories generated by large language models, every ending is slightly different; the same story gets remade as something new.”

Within the framework of this paper, he isn’t an abnormal case. He’s an extreme form of a new kind of reader. The researchers call it a “solipsistic reader-writer,” someone who simultaneously produces and consumes stories within a closed loop of dialogue, with the other end having no other humans.

Writing and reading always require two people. But with AI creation, those two roles are very likely to be the same person.

Every sentence that says “AI is reshaping creation” is missing a subject

Back to the original question.

We read sentences like this every day. AI is changing human creation. AI-generated content is growing explosively. Some platform generates hundreds of millions of pieces of content per month. Some model’s weekly active users break into the hundreds of millions.

Not one of these sentences is false. They just all omit the same thing: who is using it, and how many times.

A dataset of 573,000 conversations, with 34% related to writing novels, sounds like a shift of an era. Spread it out, though: in every ten thousand people, a couple hundred are digging several thousand times in the same hole. When you equalize each individual’s weight, this shift of the era shrinks to 7.1%.

Next time you see an AI product announce “570k pieces of content generated this month,” it’s worth asking: how many people generated those 570k pieces?

This isn’t meant to deny that AI creation is happening. The researchers themselves also say they believe similar heavy users are “very likely compulsively generating novels” on ChatGPT, Claude, and Gemini too. That part is true.

The real problem is structural misreading: describing “users” using “usage.” And every number the AI industry publishes right now uses the former.

This article is sourced from arXiv: AI Fiction in the Wild, compiled and reported by Dongqu Dongqu.