AI Fiction in the Wild

This site is an interactive visualizer over the subset of WildChat conversations we identify as fiction-related — about 195,000 conversations from roughly 10,000 estimated users. You can search the conversations, browse by user, follow how individual users iterated on the same story over weeks, and zoom into a map of the dataset to see what kinds of stories cluster together.

Because the WildChat data is anonymized, it is difficult to know exactly how many unique users are represented in the dataset. While the original research team recorded hashed IP addresses for every conversation, if someone used WildChat at home, a coffee shop, and the library, that would show up as multiple IP addresses, overcounting users in that case. To address this issue, we bundle together IP addresses that posted at least one near-duplicate fiction prompt (longer than 30 characters) from the same geographic region. Two IPs from the same state that ever submitted prompts in the same DBSCAN cluster — meaning embedding-similar opening turns — are treated as the same estimated user. This process reduced 12,873 unique IP addresses to 10,042 estimated unique users — with 610 multi-IP groups absorbing the rest, and the remaining 9,432 IPs counted as single-IP users.

For any user with many fiction conversations, we group their opening prompts into story permutations — clusters of prompts that ask for the same scenario with small variations (different characters, settings, or details). This lets you see at a glance whether a user is iterating on one story dozens of times or roaming across many genres.

Each prompt is again encoded with the all-MiniLM-L6-v2 sentence transformer, and we cluster the resulting vectors with DBSCAN — a density-based algorithm that groups vectors that sit close together in embedding space and leaves outliers unassigned. We chose DBSCAN over a hierarchical method because it doesn’t force every prompt into a cluster: one-off prompts stay isolated, which produces cleaner, more interpretable groups for repetitive users. Each cluster’s representative label is the longest common prefix shared by its prompts.

A longer paper accompanying this visualizer — with the full theoretical argument, methodology, and analysis — is in preparation. We’ll link the pre-print here as soon as it’s up.