Who expands the human creative frontier with generative AI: Hiveminds or masterminds?
Data files
Aug 29, 2025 version files 166.94 MB
-
Hiveminds-Masterminds.zip
166.94 MB
-
README.md
7.04 KB
Abstract
Artists are rapidly integrating generative text-to-image models into their workflows, yet how this human–AI collaboration affects creative discovery remains unclear. Leveraging large-scale data from an online art platform, we compare AI-assisted creators to matched non-adopters to assess novel idea contributions. Initially, generative AI increases novelty among a concentrated subset of artists, driven primarily by substantial productivity gains; however, marginal novelty per artifact declines post‑adoption, reflecting a shift toward high‑volume, incremental exploration, ultimately yielding a greater aggregate of novel artifacts by AI adopters. We observe no evidence of a human–AI complementarity effect beyond productivity-driven gains. Notably, the release of open-source Stable Diffusion accelerates novel contributions across a broader, more diverse group, suggesting that text‑to‑image tools facilitate exploration at scale, initially enabling persistent breakthroughs by a select “mastermind” group, driven by substantial volume increases, and subsequently enabling widespread novel contributions from an emergent “hivemind” of artists.
Dataset DOI: 10.5061/dryad.xpnvx0ksf
"Leveraging large-scale data from an online art platform, we compare AI-assisted creators to matched non-adopters to assess novel idea contributions."
The dataset consists of a main folder, Hiveminds-Masterminds.zip.
Data
All data has been anonymized (userid, postid) to protect user privacy.
The repository contains the following datasets.
posts: Post-level data containing artwork performance metrics and embeddings. This dataset describes the performance of individual artworks and associated feature representations.- userid: anonymized user identifier
- postid: anonymized post identifier
- published_time: year-month-date time
- umap: UMAP dimensionality-reduced embedding
users: User-level data containing creator performance metrics, treatment assignment, and propensity scores. This dataset characterizes individual creators, including their performance metrics, assignment to a treatment or control group, and calculated propensity scores.- userid: anonymized user identifier
- num_posts: total number of posts published
- favorites: total number of favorites received
- comments: total number of comments received
- pageviews: total number of profile page views received
- profile_comments: total number of profile page comments received
- watchers: total number of profile followers
- treatment: binary indicator for treatment/control cohort
- avg_posts_prior: average number of posts published per period prior to AI release
baseline: Idea embeddings external to creators in the sample; used as the universal baseline in the convex hull computation. These embeddings provide a reference point for comparison, representing ideas generated outside of the studied creator population.- postid: auto-indexed identifier for baseline set of artifacts
- umap: UMAP dimensionality-reduced embedding
matched_cohorts: All users composing the matched treatment and control cohorts for each epoch from runningconvex_hull.py. This dataset specifies the composition of the cohorts used in convex hull computations across different time periods.- same as 2. users
- ps: propensity score estimates for selection into adoption based on user-level engagement metrics upon which matching is performed
- epoch: bootstrapped iteration to which the user was matched
convex_hulls: Estimated convex hull metrics using the default configuration inconvex_hull.py. These metrics represent the boundaries of the performance achieved by the treatment and control cohorts.- period: year-month-date
- volume: convex hull volume
- area: convex hull surface area
- num_vertices: number of vertices that make up the convex hull
- num_unique_vertex_users: number of unique creators who own a vertex point on the convex hull
- group: {treatment, control, all}, denoting the cohort treatment assignment
- epoch: bootstrapped iteration
- {treatment, control}_n: number of treatment units making up the cohort
first_count_{treatment, control}_heatmaps: User-month panel of creators' first contribution counts for treatment and control units. This dataset visualizes the timing and frequency of creators' first contributions, separated by treatment group.- userid: anonymized user identifier
- rank: user ranking based on the total number of vertex contributions over the entire data period
- {vertex periods}: number of vertex points owned in each period
convex_hulls-prodeq: Same asconvex_hullsbut with productivity-equalized treatment cohort. This dataset is identical toconvex_hullsbut uses a cohort construction method that adjusts for differences in creator productivity.- Same features as 5. convex_hulls
{#}-prodeq_{True/False}-{mastermind, hivemind, hetero-low, hetero-high}_hulls: All convex hull configurations for cohort analyses with and without productivity equalization. These represent a series of experiments exploring how cohort composition and productivity equalization affect the resulting convex hull metrics.{mastermind, hivemind, hetero-low, hetero-high}refer to different grouping strategies.{True/False}indicates whether productivity equalization was applied.- Same features as 5. convex_hulls
masterminds-hiveminds-u: User IDs for the top 20%/bottom 80% vertex contributors. This dataset identifies users based on their contribution levels within the network.- userid: anonymized user identifier
- treatment: binary indicator for treatment/control cohort
- rank: user ranking based on the total number of vertex contributions over the entire data period
- percentile: percentile ranking
exploration-u: User IDs for low/high exploration users prior to adopting AI. This dataset identifies users based on their level of exploratory activity before the intervention.- userid: anonymized user identifier
- max_cdf: within-user empirical cumulative distribution function value for maximum average embedding distance prior to AI release
- low: binary indicator for assignment to low/high pre-AI exploration cohort
Scripts
The repository contains the following Python scripts and Jupyter notebooks:
convex_hull.py: Iteratively matches treatment and control units to produce unique cohorts, then computes convex hulls for cohorts over time. Running this script with the default configuration will replicate the main results.convex_hull-falsification.py: Same asconvex_hull.py, but allows for control group contamination up to a user-specified level (default 30%). This script allows for a controlled level of "contamination" or overlap between treatment and control groups.1-did_analysis.ipynb: Continuous difference-in-differences estimation with visualizations. Performs a Difference-in-Differences (DID) analysis to estimate the treatment effect and presents the results visually.2-vertex_analysis.ipynb: Generates data for decile contribution heatmaps and market share analyses and corresponding visualizations. Analyzes contribution patterns and market share, generating heatmaps to visualize these trends.3-scm_analysis.ipynb: Synthetic Control Method estimation (robustness check) with visualizations. Utilizes a Synthetic Control Method to provide a robustness check on the DID analysis, providing an alternative estimate of the treatment effect.4-vertex_discovery.ipynb: Descriptive and bootstrapping analyses for the productivity effect. Provides further investigation and statistical robustness through bootstrapping methods to assess the effect of the intervention on creator productivity.