AI Usage Patterns: Researchers Identify Extreme Repetitive Behavior in Large Language Models

A study of the WildChat dataset, published in June 2026, reveals that a small fraction of users accounts for the vast majority of creative fiction generation within AI models. Researchers from the University of Washington found that 2% of users generated more than 80% of the fiction content analyzed, with some individuals engaging in thousands of repetitive, thematic interactions involving specific fictional characters.

Who is driving AI fiction generation?

The “AI Fiction in the Wild” report, which analyzed more than half a million real-world, anonymized conversations between users and ChatGPT, categorized users into distinct behavioral groups. According to the data, 34% of all interactions involved creative writing, such as roleplay, novel drafting, or fanfiction.

The researchers identified two primary archetypes among these users:

• Looping story creators: Users who request iterative variations of a story over a period before shifting their focus.
• Infinite story seekers: Users who demand the exact same narrative or near-identical prompts repeatedly over extended timeframes.

The study highlights a significant concentration of output: while the total volume of fiction is high, the distribution is heavily skewed. The most prolific users are responsible for the vast majority of the creative content produced, suggesting that a small minority of the user base relies on AI as a primary tool for repetitive, personalized narrative generation.

Why do users engage in repetitive AI interactions?

The research team noted specific instances of extreme repetition that deviate from standard creative writing. One case study involved a user who generated thousands of iterations of a fanfiction storyline centered on “Natsuki,” a character from the visual novel Doki Doki Literature Club!.

In these thousands of generated variations, the character Natsuki consistently experienced a fictional pregnancy. Researchers observed that the user’s behavior—repeatedly forcing an AI to generate the same specific themes—is similar to the structure of consumption of pornography. Because the dataset relies on anonymized logs, the researchers were unable to determine the specific motivation behind this singular, repetitive output.

How does this compare to general AI use?

The findings provide a window into how “in the wild” usage differs from controlled testing environments. While AI developers often focus on productivity or factual accuracy, the data shows that a significant portion of the public uses these tools for immersive, private entertainment.

The prevalence of fanfiction—which accounted for 49% of all fictional content generated—highlights a demand for AI as a collaborative partner in existing intellectual properties. Among the most popular subjects for this fan-generated content were Doki Doki Literature Club!, League of Legends, and Naruto.

Key Takeaways

• High Concentration: 2% of users produce more than 80% of the AI-generated fiction found in the WildChat dataset.
• Repetitive Patterns: “Infinite story seekers” often use AI to generate near-identical narratives, a behavior researchers liken to the structure of consumption of pornography.
• Popular Themes: Fanfiction remains a dominant category, with users frequently leveraging AI to explore narratives within established gaming and anime franchises.

The report serves as a baseline for understanding human-AI interaction dynamics. As developers continue to refine safety filters and usage guidelines, the research suggests that user behavior is often driven by highly specific, individualistic desires that may not align with the intended utility of general-purpose large language models.