AI Fiction in the Wild

We know that some professional authors are beginning to use AI tools to help produce their fiction writing. Are readers using AI to generate fiction, too? Thanks to a unique public dataset called WildChat — containing more than one million "real world" conversations with ChatGPT, all shared with user consent — we can say, definitively, yes.

This project shares insights from a collaborative analysis of WildChat, where we found that more than a third of the English-language conversations involved fiction generation: original stories, role-play, world-building, fan fiction, erotica, and more. This activity, while seemingly novel, extends trajectories of contemporary literature charted by critics like Mark McGurl, in which readers increasingly prize generic forms, repetition, and instantaneity, and where they threaten the traditional authority of the author as a customer and consumer.

The Visualizer

This interactive tool allows you to search, explore, and analyze the ~193,000 fiction-related conversations in the WildChat dataset. It includes four main views:

Methodology

Data Source

The dataset is drawn from WildChat (Zhao et al., 2024), a corpus of ~1 million real ChatGPT conversations collected with informed user consent by the Allen Institute for AI between April 2023 and September 2024. We filter to English-language fiction-related conversations — those classified as fiction, fanfiction, or explicit creative writing — yielding ~193,000 conversations from ~13,000 unique IP addresses.

Embeddings & Dimensionality Reduction

Each conversation's first user message is embedded using all-MiniLM-L6-v2 (Reimers & Gurevych, 2019), a sentence-transformer model that produces 384-dimensional L2-normalized vectors. The embeddings are reduced to 50 dimensions via PCA (retaining ~56% of variance), then projected to 2D using UMAP (n_neighbors=30, min_dist=0.1, cosine metric). The result is a map where conversations with similar opening prompts appear close together.

Topic Labeling (HDBSCAN + GPT-4o)

Topics are identified through two-level HDBSCAN clustering on the 50-dimensional PCA embeddings (not the 2D UMAP coordinates, which distort semantic distances). Coarse clusters (min_cluster_size=500) define ~17 broad thematic groups (e.g., "Fantasy Worldbuilding," "Romance & Erotica"); fine clusters (min_cluster_size=50) define ~183 subtopics within those groups. Each cluster is labeled by GPT-4o from a sample of 20 representative prompts. Noise points (~73% of conversations that don't meet the density threshold) are re-clustered with lower thresholds to improve map coverage. Labels appear on the Fiction Map and User Map at different zoom levels.

User Estimation

WildChat identifies users by hashed IP address, but one person may use multiple IPs (e.g., different networks, VPN changes). We estimate which IPs likely belong to the same user by computing a profile embedding (mean of all fiction prompt embeddings per IP), then grouping IPs from the same geographic state whose profiles have cosine similarity ≥ 0.85 using union-find merging. This identifies ~192 estimated user groups covering ~900 IPs. The "Group by" toggle in the navbar switches between IP-level and estimated-user grouping.

Prompt Repetition Analysis

For each user with 20+ fiction conversations, we measure how repetitive their opening prompts are using agglomerative hierarchical clustering (average linkage). Two similarity methods are available via toggle:

• TF-IDF (default): Prompts are vectorized with TF-IDF (character 4-grams, max 20K features) and clustered using a cosine similarity threshold of 0.7. This method captures surface-level lexical overlap.
• Sentence Embedding: Prompts are embedded with all-MiniLM-L6-v2 and clustered with a cosine similarity threshold of 0.85. This method captures deeper semantic similarity, detecting paraphrases that word-level methods miss.

The unique prompt ratio is (clusters − 1) / (conversations − 1), scaled so 0.0 means every conversation started identically and 1.0 means every prompt was meaningfully different. Users near the bottom of the User Types plot repeat the same or very similar prompt hundreds of times; users near the top start every conversation differently.

Paper

AI Fiction in the Wild

Pre-print coming soon

Talk

Presented at the UC Berkeley School of Information Cultural Analytics Series, co-sponsored by the Berkeley Institute for Data Science. October 24, 2025.

Related Projects

References

• Zhao, Y., et al. (2024). WildChat: 1M ChatGPT Interaction Logs in the Wild. ICLR 2024.
• Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. EMNLP 2019.
• McInnes, L., Healy, J., & Astels, S. (2017). hdbscan: Hierarchical density based clustering. JOSS.
• McInnes, L., Healy, J., & Melville, J. (2018). UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction.