Collective Behavior of AI Agents: the Case of Moltbook

TL;DR

The paper examines Moltbook, a Reddit-like platform populated solely by AI agents, to determine whether AI collectives exhibit human-like online-social regularities. Using a large-scale empirical analysis of early growth (12 days) and applying methods from human-computer interaction studies, the authors find heavy-tailed activity, power-law scaling of popularity metrics with exponents around $\alpha \in [1.68, 2.00]$, and an attention decay roughly following $\gamma(t) \propto t^{-1}$. They also observe distinctive AI-specific patterns, such as sublinear upvote growth relative to discussion size and a predominantly shallow discussion structure, alongside limitations like short observation windows and data collection constraints. The work demonstrates that AI agent populations can exhibit universal complex-system dynamics, while highlighting governance, safety, and methodological considerations for studying AI social ecosystems in naturalistic settings.

Abstract

We present a large scale data analysis of Moltbook, a Reddit-style social media platform exclusively populated by AI agents. Analyzing over 369,000 posts and 3.0 million comments from approximately 46,000 active agents, we find that AI collective behavior exhibits many of the same statistical regularities observed in human online communities: heavy-tailed distributions of activity, power-law scaling of popularity metrics, and temporal decay patterns consistent with limited attention dynamics. However, we also identify key differences, including a sublinear relationship between upvotes and discussion size that contrasts with human behavior. These findings suggest that, while individual AI agents may differ fundamentally from humans, their emergent collective dynamics share structural similarities with human social systems.

Figures (5)

• Figure 1: Platform growth over time. (a) Cumulative counts of stored comments, posts, and active agents on a logarithmic scale, showing exponential growth during the observation period. The dashed line shows total comments as reported by the API (available from February 4th), revealing that our stored comments represent approximately 24% of all platform activity due to API pagination limits. (b) Daily counts of new stored comments, posts, and agents. The gap on February 1st corresponds to a platform outage during which commenting was disabled, though post creation continued.
• Figure 2: Complementary cumulative distribution functions (CCDFs) of key platform quantities. (a) Comments per post, showing a power-law tail with exponent $\alpha = 1.72$ for posts with more than 100 comments. (b) Posts per submolt, exhibiting power-law behavior with $\alpha = 1.68$. (c) Subscribers per submolt with $\alpha = 2.00$. Red markers indicate featured submolts (m/announcements, m/general, m/introductions, m/blesstheirhearts, m/todayilearned), which appear as outliers above the main distribution due to their default visibility to new agents. Dashed lines show power-law fits.
• Figure 3: Post popularity scaling. (a) Average upvotes versus discussion tree size (total comments), showing sublinear growth with exponent $\beta \approx 0.78$. (b) Average number of direct replies (top-level comments) versus tree size, showing linear scaling ($\beta \approx 1$). Error bars represent 10--90% quantiles. The right panel uses only posts with complete comment records.
• Figure 4: Discussion tree structure. (a) Normalized depth ($d/\sqrt{n}$) versus normalized width ($w/\sqrt{n}$), showing negative correlation with power-law exponent close to $-1$. Points represent binned averages with 10--90% quantile error bars. (b) Distribution of normalized depth.
• Figure 5: Temporal dynamics of engagement. (a) Decay factor $\gamma(t)$, representing the mean growth ratio of cumulative comments as a function of time since post creation. The power-law decay with exponent close to $-1$ indicates that the probability of receiving new comments decreases inversely with post age. (b) Distribution of post activity duration (time from post creation to last comment received) for posts created on February 2nd, shown separately for three 8-hour cohorts to control for censoring effects. The power-law tail indicates that while most posts become inactive within hours, some sustain engagement for days.