When researchers set out to study how ordinary people use ChatGPT to write fiction, they expected to map trends: genres, tropes, the collective texture of machine-assisted storytelling. Instead, as reported by PC Gamer and IGN, they ran headfirst into one person. A single anonymous super user had generated thousands of stories featuring pregnant characters from the cult visual novel Doki Doki Literature Club — enough volume to skew the dataset and stop the researchers in their tracks.
It’s a funny anecdote. It’s also a genuinely useful lesson about how we read data from AI tools.
The Spiders Georg Problem
PC Gamer reached for the perfect meme: this user is the “Spiders Georg” of AI fanfic — the outlier who, in the joke, eats 10,000 spiders a day and single-handedly drags the average up. The comparison is sharper than it looks. Aggregate statistics assume no single actor dominates the sample. But human behavior online is famously lopsided: a tiny sliver of users generate the overwhelming majority of content in almost any system you care to measure — Wikipedia edits, forum posts, Reddit karma.
Generative AI supercharges that skew. When producing a thousand stories costs nothing but prompts and patience, one obsessive can out-produce ten thousand casual users combined. Any study of “how people use ChatGPT” that doesn’t account for these whales isn’t measuring the population — it’s measuring the loudest voice in the room.
Why This Should Change How We Read AI Headlines
We are drowning in confident claims about what AI is “being used for.” People use it for therapy. For coding. For companionship. For niche pregnancy fanfic, apparently. Each of these claims usually rests on log data, and log data is precisely where the Spiders Georg effect bites hardest. A handful of power users can manufacture an entire apparent “trend” that the median user never touches.
The DDLC detail matters here too, and not because it’s salacious. Doki Doki Literature Club is a game specifically about characters who are not what they seem, a story that weaponizes the gap between surface and reality. A researcher assuming the data reflected the crowd, when it reflected one person, is a neat echo of the game’s own central trick.
The Real Takeaway
None of this is a knock on the user, whose hobby is harmless and, frankly, committed. The point is methodological. As academics increasingly mine conversation logs to tell us what humanity wants from these tools, the honest version of every finding needs an asterisk: is this a pattern, or is this a person? Distributions have long tails, and in the age of infinite cheap generation, those tails can wag the entire dog.
The most human thing revealed by this study wasn’t the fanfic. It was the researchers’ surprise.