A bombshell new academic study reveals that more than a third of all real-world ChatGPT conversations involve some form of fiction generation. Here's what the data tells us about how people are actually using AI to write — and read — stories.
You've seen the headlines about AI writing novels. You've watched the debates about whether chatbots will replace authors. But until now, nobody actually knew how real people were using these tools to create fiction.
We finally have answers — and they're wilder than anyone predicted.
What the Study Found
A team of researchers from the University of Washington and Cornell University just dropped what might be the most important paper on AI and fiction ever published. Titled "AI Fiction in the Wild," the study analyzed more than 500,000 real, anonymized ChatGPT conversations from the WildChat dataset — a collection of user interactions gathered with explicit consent through a public interface hosted on Hugging Face.
The headline finding? More than one third (34%) of all conversations involved some form of fiction generation. That includes original stories, roleplay, fanfiction, scripts, and erotica.
But that's just the start.
The researchers identified that just 2% of users were responsible for more than 80% of the fiction conversations. These "power users" — whom the paper calls "infinite story demanders" — prompt the same stories over and over, sometimes thousands of times across months, tweaking details and watching the AI spin slightly different versions each time.
One user generated the same fanfiction narrative — a pregnancy story involving characters from the video game Doki Doki Literature Club! — thousands of times over several months. Another user cycled through nearly 900 conversations across eleven months, returning again and again to a time-travel story about a character named Meredith Lexington.
Not Your Usual Fiction
The study reveals something fascinating about what kind of fiction users are generating. Nearly half of all fiction requests were fanfiction, drawing on existing characters and universes. More than a quarter were sexually explicit. Romance tropes, generic formulas, and niche combinations dominated.
Users aren't asking ChatGPT to write the next Great American Novel. They're asking for highly specific, personalized stories that hit exactly the right notes — an enemies-to-lovers fic set in a particular fandom, or a time-loop romance with their favorite characters.
And they're writing incredibly detailed prompts to get there. The median first prompt was 103 words long, with some running into the hundreds. These aren't "write me a story" one-liners. Users are co-authoring, providing intricate character backstories, specific plot beats, and even dialogue snippets, then letting the model fill in the prose.
The "Solipsistic Reader-Writer"
The researchers argue that this represents a potentially historic shift in how fiction works. They coin the term "solipsistic reader-writer" to describe someone who both generates and consumes fiction within a closed loop — interacting with a machine rather than another human being.
For most of literary history, fiction has been defined by an exchange between at least two human beings: a reader and a writer. AI-generated fiction collapses that distinction. The same person guides the creation and consumes the result, without ever encountering another human's imagination.
The paper draws a provocative parallel to pornography consumption — noting that AI-generated fiction shares structural affinities with personalized, on-demand, repetitive cultural forms. Both prioritize immediacy, niche customization, and instant gratification over social connection.
But the researchers are careful not to frame this as purely dystopian. "Our aim is not to characterize all AI-generated fiction as solipsistic," they write, "but to identify a potentially unprecedented development in literary history."
What This Means for Publishing
This study arrives at a moment when the publishing industry is already reeling from AI-related controversies — from prize scandals to Amazon being flooded with AI-generated books. But this data suggests something more fundamental is happening.
Mainstream publishers have insisted that AI's role will be limited to things like slush pile evaluation. "The challenge for publishers is not generating more content, it's solving the discovery problem," Madeline McIntosh, the former CEO of Penguin Random House, told Publishers Weekly last year.
But disruptors see it differently. Character.AI — valued at over a billion dollars with more than 20 million monthly users — allows people to chat with customized fictional characters. More than half its users are Gen Z and Gen Alpha. Google launched Gemini Storybook, which creates ten-page illustrated books from prompts. And at least one self-publishing company, Inkitt, has publicly stated its goal of using AI to hyper-customize stories for individual readers.
The data from WildChat suggests these companies are betting on something real. A significant chunk of the population — particularly younger users — already wants personalized, AI-generated fiction.
The Romance Connection
For anyone in the romance fiction space, the study's findings hit close to home. Romance and erotica were among the most popular categories of AI-generated fiction in the dataset. The researchers note that both genres hinge on "predictability" — and that this predictability is part of what makes them "therapeutic and psychologically satisfying for readers."
Sound familiar? It's the same dynamic that makes romance novels the most popular genre in publishing, the same engine that powers Kindle Unlimited's endless churn of happily-ever-afters, and — yes — the same principle behind personalized AI romance novels.
As McGurl argued in his landmark study of Amazon's effect on fiction, the marketplace increasingly rewards generic forms, repetition, and stories that closely match what readers already know they want. AI takes this to its logical extreme: a story customized to the individual reader's exact preferences, delivered instantly.
Fanfiction and the Tagging Revolution
The study also reveals a fascinating link between fanfiction culture and AI-generated stories. On platforms like Archive of Our Own, readers navigate vast archives using detailed tagging systems — searching for specific relationships, tropes, and content warnings.
The researchers found that LLM users are essentially doing the same thing, but in natural language. Instead of selecting tags for "Enemies to Lovers" and "Harry Potter/Draco Malfoy," they type out exactly what they want. The model does the rest.
"The model draws on conventional screenplay formatting styles, tropes of Japanese Role-Playing Game melodrama, character archetypes like the 'diva' and the 'brute,' and precise diegetic details from the Persona 4 universe," the researchers note of one example.
In other words, AI fiction is fanfiction on steroids — infinitely customizable, instantly generated, and endlessly repeatable.
The Big Question Nobody's Answering
The study raises uncomfortable questions for anyone who cares about the future of fiction. Will AI-generated stories disrupt traditional publishing? Will they undermine the human sociality of fiction — the book clubs, the fan communities, the shared cultural touchstones that come from millions of people reading the same story?
Ted Chiang, writing in The New Yorker, argued that art is inherently communicative. "It's by living our lives in interaction with others that we bring meaning into the world," he wrote. "That is something that an auto-complete algorithm can never do."
But the data suggests that many users don't care. They want their story, their way, right now. And they're getting it.
The researchers point out that the average LLM response in their dataset was just 385 words — far shorter than a novel or even a typical fanfiction story. Models still struggle with coherent long-form narratives. But context windows are growing, the technology is improving, and the user appetite is clearly there.
What's Next
The "AI Fiction in the Wild" dataset and analysis tools are publicly available, and the researchers encourage further study. They've released their code, a searchable website at ai-fiction-wild.com, and the fiction-labeled conversation subset for anyone who wants to dig deeper.
For the rest of us — the readers, writers, and publishers trying to make sense of this moment — the study offers both a warning and an opportunity.
The warning: AI-generated fiction is not a future possibility. It's happening right now, at scale, driven by real users with real appetites. The publishing industry needs to reckon with that reality.
The opportunity: personalized, AI-assisted storytelling is clearly something people want. The question isn't whether to embrace it, but how to do so in a way that preserves what makes fiction magical — the spark of human creativity, the surprise of encountering another person's imagination, the shared experience of a story well told.
In the end, the WildChat data tells us something hopeful alongside the anxiety: people love stories. They love them enough to spend months generating endless variations of the same narrative. They love them enough to write painstakingly detailed prompts. They love them enough to build their own fictional worlds, one ChatGPT conversation at a time.
The challenge — for platforms like SwoonScript, for publishers, and for the entire storytelling ecosystem — is channeling that love into something that connects us rather than isolates us.
Want to explore the future of AI romance storytelling? Try SwoonScript — where you're the main character.