As spotted by Japan’s IT Media News, a paper presented at Purdue University’s MFS Cultural AI Conference called AI Fiction in the Wild performed a fascinating study of half a million anonymous English-language ChatGPT-user conversations using WildChat, and found some real oddities. Well, one in particular.
The study’s purpose was to analyze the significant number of AI prompts that boil down to variations on “write me some fanfic.” Which there sure is a lot of. “More than a third of the conversations contained some form of fiction generation,” the researchers said, “including original stories, scripts, roleplay, worldbuilding, fanfiction, and erotica.”
The study defined two separate varieties of AI-fiction prompters, which they call story cyclers and infinite story demanders: “Story cyclers ask for iterations of the same story for a period of time, then switch to another story or topic. Infinite story demanders request the same story, or a very similar one, over and over again for long stretches of time.”
The most prolific of those infinite story demanders had an extremely specific interest. Over several months, one user generated thousands of variations on a story where one of the girls of visual novel Doki Doki Literature Club, usually Natsuki, gave birth, usually to a daughter named Sakura. “Sakura may be a reference to a character from the Naruto manga and anime series”, the researchers helpfully noted.
If you’ve played Doki Doki Literature Club you may be thinking it’s an ironic choice for AI-generated fiction, but if you haven’t I won’t spoil why. But I will say you can get it for free on Steam and itch and only takes about five hours to experience.
Our friend wasn’t the only power user in the dataset: “Two percent of users within the fiction subset are responsible for more than 80% of the conversations.” While the user with the pregnancy fixation certainly skewed the numbers, the most popular fictional universes people requested fanfic of were Doki Doki Literature Club, League of Legends, Freedom Planet, and Naruto.
That data, by the way, came from WildChat, a ChatGPT interface hosted on Hugging Face that people could use for free with the understanding their conversations would be publicly available, though anonymized, for academic research. The first release of that data, collected between April 2023 and May 2024, contained over 500,000 English-language conversations for researchers to delve through.
Part of what’s interesting about the AI fanfic dataset is how different it is to what’s being posted on the A03 repository, where stories based on Harry Potter, My Hero Academia, and Supernatural were much more common during the same period. With fanfic written by actual humans, writers encourage each other and are encouraged by their readers, who follow them to other fandoms and discover new things to love. It’s communal, while generating fiction with an AI is a much more solipsistic idea.
Normally, fanfic is a space for people to share their love of stories, but the AI-generated version is inherently isolated and masturbatory by comparison. Even when it’s not about imagining pregnant teenagers.
AI,Software#Researchers #studying #ChatGPT #conversations #surprised #find #power #user #churning #thousands #Doki #Doki #Literature #Club #pregnancy #fics1783222086
