Creative writing has become a primary workload for ChatGPT—fiction, fanfiction, and the economics of attention
A joint analysis by the University of Washington and the University of Colorado Boulder, drawing on millions of anonymized ChatGPT conversations, offers a revealing snapshot of how generative AI is being used in the wild—not merely as a productivity tool, but as a creative engine. The headline finding is unambiguous: fiction is a dominant use case, representing more than one-third of interactions examined. Within that fiction universe, fanfiction accounts for nearly half, and sexually explicit content exceeds a quarter of fanfiction-related exchanges.
For business and technology leaders, this is more than a cultural footnote. It signals that large language models (LLMs) are increasingly functioning as always-available narrative collaborators, and that the most “sticky” engagement may come not from office workflows, but from serial, emotionally driven entertainment creation. The data also underscores a platform reality: creative use is not evenly distributed. A small subset of users can shape both compute costs and product direction.
Key behavioral takeaways with strategic implications include:
- Fiction as a core engagement driver: sustained, repeat interactions are intrinsic to storytelling.
- Fanfiction as a high-volume genre: it is structurally suited to iterative remixing and variation.
- Adult content as a moderation stress test: explicit material is not fringe; it is statistically material within a major category.
Two archetypes define AI storytelling behavior: “story cyclers” vs. “infinite story demanders”
The researchers distinguish between two user patterns that, taken together, map neatly onto product segmentation and monetization strategy.
“Story cyclers” behave like rapid prototypers: they iterate on a plot a handful of times, harvest what they need, and move on. Their behavior resembles a lightweight creative workflow—high novelty, low persistence, and limited need for long-term memory.
“Infinite story demanders,” by contrast, treat the model as a long-running narrative machine. They request endless variations of a single story world over months or years, returning repeatedly to the same characters, themes, and emotional beats. The study’s illustrative outlier—the “most prolific user”—reportedly generated extensive pregnancy-themed fanfiction based on *Doki Doki Literature Club!* over an extended period. The specifics are less important than what the pattern represents: AI as a personalized serial-content generator, tuned to niche preferences that traditional media rarely serves at scale.
The concentration metrics are striking and commercially consequential:
- Only ~2% of fiction writers fall into the extreme “power user” category, yet they generate over 80% of fiction dialogue.
- Repeat prompting averages ~42% across fiction users, but rises to ~85% among power writers.
This is a dual signal. On one hand, it suggests deep engagement—the kind subscription businesses covet. On the other, it points to persistent dissatisfaction or unmet intent, where users must repeatedly steer the model to achieve desired continuity, tone, or character fidelity. In practical terms, the “infinite demander” is both a platform’s most valuable creative customer and one of its most expensive.
Iterative prompting exposes LLM limits—and highlights where model development can win
The study frames creative writing with ChatGPT as an iterative loop: prompt → response → refinement. That cycle resembles modern product design and A/B testing more than solitary authorship. Users are effectively using the model as:
- a brainstorming partner for plot and dialogue
- a simulation environment for “what if” narrative branches
- a drafting assistant that can instantly re-render scenes in new tones or structures
Yet the very need for constant refinement reveals friction points in current LLM architecture. “Infinite story demanders” embody a tension between novelty-seeking behavior and the model’s finite expressive register. When users push for countless endings, escalating twists, or hyper-specific emotional arcs, the system must juggle competing requirements:
- diversity vs. coherence (newness without breaking the story)
- character consistency vs. improvisation (agency without drift)
- emotional depth vs. safety constraints (intensity without policy violations)
For AI developers, power-user behavior is not noise—it is diagnostic data. High-frequency variation requests can pinpoint where models struggle most:
- maintaining long-range narrative consistency
- preserving stable character traits and motivations
- delivering distinctive voice without collapsing into generic prose
- handling sensitive or adult themes with contextual nuance and policy compliance
This is where targeted investment can yield disproportionate returns. Fine-tuning and product features optimized for high-engagement genres—fanfiction included—could improve satisfaction quickly, not because fanfiction is uniquely important, but because it is a stress test for memory, style control, and continuity, the same capabilities enterprise users increasingly demand in other domains.
Platform strategy: pricing, product design, and moderation in a fandom-powered AI economy
From a platform economics perspective, the study describes a familiar digital pattern: a small minority drives the majority of consumption. In AI-as-a-service, that translates directly into compute cost concentration. Without differentiated pricing, providers risk subsidizing heavy creative usage with revenue from lighter users—an imbalance that becomes more acute as context windows expand and generation volumes rise.
Several strategic responses emerge naturally from the observed archetypes:
- Tiered monetization aligned to resource intensity
– “Creative power” tiers with longer context windows, higher token limits, or advanced narrative controls
– consumption-based options such as pay-per-variation or narrative credits to curb runaway usage while preserving engagement
- Feature segmentation by creative intent
– for cyclers: fast templates, quick scene generators, “finish-the-story” tools
– for infinite demanders: branching logic, persistent story memory, character databases, versioning, and export pipelines
- Community and retention mechanics
– galleries, remix permissions, collaborative workspaces, and peer review can turn solitary prompting into a micro-community flywheel, similar to modding ecosystems and serial-content fandoms
Alongside opportunity sits brand and regulatory exposure. With sexually explicit content representing a substantial slice of fanfiction interactions, platforms face a delicate balancing act: enabling legitimate creative expression while managing legal risk, policy compliance, and reputational integrity. The operational requirement is clear: moderation systems must be fine-grained and context-aware, capable of distinguishing transformative fan works from harmful or prohibited material, and robust enough to handle edge cases without flattening creative utility.
What this research ultimately surfaces is a market truth: generative AI is not only automating tasks—it is manufacturing personalized entertainment at scale, with a small cadre of ultra-engaged users shaping both the cost curve and the roadmap. The platforms that thrive will be those that treat these behaviors not as anomalies to suppress, but as signals to productize—building durable narrative tooling, sustainable pricing, and governance that can withstand the intensity of human imagination when it finally has an infinite co-author.
By
By
By
By
By
