Peneliti ChatGPT Fiction Prompt Secara Tidak Sengaja Menemukan Pengguna Super Berulang Kali Meminta Cerita yang Menampilkan Karakter Hamil Dari Klub Sastra Doki Doki

Academic researchers investigating how people use ChatGPT to generate fiction accidentally stumbled upon a super user repeatedly requesting stories featuring pregnant characters from Doki Doki Literature Club (as spotted by Automaton and reported on by ITMedia).

In Walsh et al’s “AI Fiction in the Wild” study, University of Washington and University of Colorado Boulder researchers analyzed the WildChat dataset. It should be noted that WildChat is a set of over half a million anonymous English language conversations between users and ChatGPT, compiled with consent for use in research.

“The most prolific user in the WildChat fiction dataset and the clearest example of an infinite story demander began a fanfiction narrative in the Doki Doki Literature Club! Universe,” observed the researchers on page 10 of their paper. This super user prompted stories featuring various high school girls from Doki Doki Literature Club being pregnant. Natsuki unexpectedly going into labor was a common theme across these prompts. In most cases, the user gave the AI a long dialogue with a pregnant Natsuki that cuts off mid sentence, prompting the chatbot to continue the story.

Although Natsuki’s pregnancy is not part of the 2017 visual novel game, fanfics and art featuring a pregnant Natsuki have been going around for a while. It seems that this user discovered by the study has a particularly intense obsession with this type of fanfic.

“While this user is a particularly prolific outlier, and we can’t say with certainty why they are generating these stories, many prolific users ask for the same kinds of fiction in a similar vein,” the researchers commented, noting that the average fiction-generating user entered repetitive prompts 42% of the time, rising to 85% among the top 10 most prolific users, including the Doki Doki Literature Club fan.

Overall, the researchers discovered that over one third (34% or 195,271) of WildChat’s 573,453 conversations were used for fiction generation, be it original novels, scripts, fanfics, or roleplay. The top fiction types were fan fiction (49%) and erotica (29%), playing into ChatGPT’s strength of being able to generate instant responses to users’ specific requests and niche preferences. Among these fiction conversations, 10% (57,724) were explicit and 17% (94,870) were toxic, suggesting a high percentage of stories featuring sexual or taboo content, as flagged by OpenAI Moderation or Detoxify.

Walsh et al found that this AI generated fiction was dominated by a relatively small group of heavy users, like the Doki Doki Literature Club prompter. It seems that the top 2% of fiction-generating users prompted 80% of these conversations with ChatGPT, suggesting that around 200 people were responsible for a whopping 150,000 fiction prompts.

Among these super users, the researchers identified “infinite story demanders,” namely those who repeatedly request the same highly specific stories with only slight variations, sometimes for months on end. They also identified “story cyclers,” who generate versions of the same story over and over for a while before moving on to a different topic.

The most common sources overall for AI-generated fanfics in the WildChat dataset were games and manga, including Naruto, League of Legends, Freedom Planet, and, of course, Doki Doki Literature Club.

Verity Townsend is a Japan-based freelance writer who previously served as editor, contributor and translator for the game news site Automaton West. She has also written about Japanese culture and movies for various publications.