arxiv.org web signal

Researchers find over a third of ChatGPT chats write fiction

TL;DR

  • A new arXiv preprint analyzing over 500,000 anonymized English-language ChatGPT conversations finds more than one third involve fiction generation.
  • The behavior is dominated by power users the authors call 'infinite story demanders,' who repeatedly request and revise variations of similar narratives.
  • Users gravitate especially toward fanfiction and erotica, favoring generic forms, repetition, immediacy, and niche combinations of story elements.

A new preprint from Neel Gupta, Maria Antoniak, and Melanie Walsh, posted to arXiv in June 2026, makes a claim that reframes what ChatGPT actually is in daily use. On a corpus of over 500,000 anonymized, English-language ChatGPT-user conversations, more than one third involved some form of fiction generation. Not homework, not code, not email drafting. Original stories, roleplay, fanfiction, and erotica.

The authors describe the pattern as dominated by power users, and they give those users a name: 'infinite story demanders,' who in the paper's phrasing 'repeatedly request and revise variations of the same or similar narratives over extended periods of time.' The genres that pull them in are not literary. Fanfiction and erotica lead, and the authors write that users are 'broadly drawn to generic forms, repetition, immediacy, and niche combinations of story elements.'

The provocation the paper lands on is what they call the 'solipsistic reader-writer,' a user who 'both generates and consumes fiction within a closed conversational loop, interacting with a machine rather than a human other.' That framing pushes ChatGPT out of the productivity-assistant box and into an interactive entertainment product sitting somewhere between self-publishing, fanfiction communities, and pornography, three cultural forms the authors explicitly compare it to.

The honest caveats are worth stating. This is a June 2026 preprint, not a peer-reviewed result, and the corpus is English-only. The abstract does not tell you the paying-versus-free split of the fiction users, whether the same behavior shows up on Claude or Gemini, or how a 'conversation' was classified as fiction in borderline cases like roleplay or journaling. Take the specifics as reported, not settled.

The reason to care anyway is that if the ratio is even directionally right, an assistant company is quietly one of the largest interactive-fiction platforms in the world, and the moderation, product, and safety decisions that go with that role are going to have to grow up to match the actual user base rather than the marketing one.

Shared on Bluesky by 2 AI experts