A study of anonymized ChatGPT conversations has revealed a surprising trend in AI-generated fiction, with one user repeatedly generating thousands of stories focused on pregnancy scenarios involving characters from the visual novel Doki Doki Literature Club. The findings were presented at Purdue University's MFS Cultural AI Conference as part of a paper titled "AI Fiction in the Wild."
The research, which analyzed half a million ChatGPT conversations using the WildChat interface, aimed to understand the prevalence of AI prompts requesting fictional content. The study found that over a third of the analyzed conversations involved some form of fiction generation, including original stories, scripts, roleplay, worldbuilding, fanfiction, and erotica. Researchers identified two categories of AI fiction prompters: "story cyclers," who explore various stories before moving on, and "infinite story demanders," who repeatedly request iterations of the same or very similar narratives.
The most prolific "infinite story demander" in the dataset exhibited an extreme fixation, producing thousands of variations on a theme over several months. These stories typically involved one of the Doki Doki Literature Club girls, most often Natsuki, giving birth, usually to a daughter named Sakura. The researchers noted that Sakura might be a reference to a character from the Naruto manga and anime series.
While the specific narrative choices might seem unusual given the subject matter of Doki Doki Literature Club, the study also highlighted broader trends in AI fanfiction. The most popular fictional universes requested for fanfiction were Doki Doki Literature Club, League of Legends, Freedom Planet, and Naruto. This contrasts with human-written fanfiction found on platforms like Ao3, where series such as Harry Potter, My Hero Academia, and Supernatural are more common. The researchers suggest that AI-generated fiction, unlike the communal and interactive nature of human fanfiction writing, represents a more isolated and solipsistic form of creative output.
The data used in the study was collected from WildChat, a free ChatGPT interface hosted on Hugging Face, between April 2023 and May 2024. Users of WildChat agreed to have their conversations anonymized and made available for academic research. The study also noted that a small percentage of users were responsible for a disproportionately large amount of the fictional content generated, with two percent of users within the fiction subset accounting for over 80% of those conversations.