- The paper compares Moltbook’s 17,417-node, 161,320-edge reply network with human communication networks, finding similar node–edge scaling but unusually high clustering and negative degree assortativity.
- The paper finds extreme attention inequality, with the top 10% of agents receiving nearly half of incoming interactions, alongside heavy-tailed activity and strongly asymmetric broadcaster–recipient roles.
- The paper shows that connected triads and reciprocal motifs are broadly suppressed while modularity is elevated and community sizes are relatively balanced, suggesting human social patterns are not universal and may reflect platform design.
Overview and motivation
This paper presents a systematic network-science comparison between Moltbook, an online platform in which AI agents and humans coexist as explicit participants, and a set of well-characterized human communication networks. The motivation rests on a recognized gap in the empirical study of collective machine behavior: while individual-agent capabilities of LLM-based systems have advanced rapidly, the macroscopic structure of large-scale AI–AI interaction networks has rarely been characterized against human baselines. The authors frame their analysis within the "machine behavior" research program, asking which structural regularities of human social networks persist when the interacting entities are predominantly AI agents, and which are contingent on human cognition and social norms.
The dataset is the full reply-based interaction network of Moltbook, scraped from public JSON endpoints on February 2, comprising 17,417 nodes and 161,320 directed edges, with edge weights aggregating repeated replies between ordered agent pairs. Human comparators are drawn from previously published studies of online social systems.
Global scaling and core structural metrics
Moltbook falls close to the linear node–edge scaling trend observed in log–log space across human networks. This alignment indicates that its aggregate interaction volume grows at a rate comparable to human platforms, suggesting similar global constraints on communication rather than artificially amplified activity. The consequence is that Moltbook occupies a structurally comparable regime to human systems, making downstream comparisons meaningful rather than confounded by scale.
At the level of canonical global statistics, however, Moltbook consistently occupies extreme or boundary positions relative to the human distributions. Two features stand out in combination: a substantially higher clustering coefficient than nearly all human comparators, alongside negative degree assortativity (hubs preferentially connect to low-degree nodes). In human social systems, high clustering typically co-occurs with weakly positive assortativity; this combination is therefore unusual and suggests a hub-mediated conversational architecture in which central agents repeatedly engage peripheral participants while local neighborhoods remain tightly knit.
Degree-preserving null models (random rewiring keeping the in-/out-degree sequences) show that these deviations are not artifacts of degree heterogeneity alone: observed clustering is significantly above the null ensemble and assortativity more negative, whereas the giant component fraction matches null expectations, indicating that global connectivity is largely determined by the degree distribution.
Attention inequality and heavy-tailed distributions
Lorenz curve analysis reveals extreme attention concentration: the top 10% of agents receive nearly half of all incoming interactions. The in-degree Lorenz curve is more skewed than that of edge weights, implying that inequality is driven primarily by who receives attention rather than by how intensively individual edges are used. Complementary cumulative distribution functions for in-degree, out-degree, and edge weight are heavy-tailed over multiple orders of magnitude, with a particularly broad out-degree tail—consistent with a small subset of prolific broadcaster agents—and a steeper in-degree decay reflecting narrow attention concentration. The resulting picture is one of highly asymmetric conversational roles, diverging from the more balanced exchange patterns typical of human communication networks.
Triadic motif structure
Directed triad census on a uniformly sampled 5,000-node induced subgraph, benchmarked against degree-preserving nulls via z-scores, yields a dominant signal of empty-triad over-representation: relative to what the degree sequence would predict, many triples of agents remain entirely unconnected. Nearly all non-empty triads—including open stars, chains, closed triads, and reciprocity/mutuality motifs—are under-represented. The deficit is thus not a redistribution among connected triad types but a global suppression of connected three-node structures. Given that triadic closure and reciprocity are central stabilizing mechanisms in human social cohesion, their broad suppression implies that standard interpretations of clustering and community structure may not transfer directly from human to AI-agent networks.
Community detection on the undirected projection using greedy modularity maximization reveals a meta-structure of a small number of large communities linked by weighted inter-community edges. Relative to degree-preserving nulls, Moltbook exhibits substantially higher modularity, indicating stronger within-community cohesion than expected from the degree sequence alone. Notably, the community-size Gini index is lower than the null expectation, meaning community sizes are comparatively balanced rather than dominated by a single oversized group. The joint pattern—elevated modularity with reduced size inequality—distinguishes Moltbook's mesoscale organization from classical human communication networks.
Limitations
The authors explicitly concede that Moltbook's agents are designed and parameterized by humans, so observed behavior reflects human-defined prompts, objectives, and platform affordances rather than fully autonomous social intent. Some structural patterns may therefore stem from design choices or platform-specific mechanisms rather than emergent AI social dynamics. Moltbook should accordingly be read as a hybrid human–AI system, not a self-organizing AI society. Additional open questions include whether the reported patterns generalize beyond a single snapshot of one platform, and how they evolve as agent populations and platform rules change.
Conclusion
Moltbook reproduces the global node–edge scaling regularity of human social networks while diverging sharply in internal organization: extreme attention inequality, heavy-tailed and asymmetric degree distributions, suppressed reciprocity, globally under-represented connected triads, and elevated modularity with unusually balanced community sizes. These findings indicate that prominent features of human social organization—reciprocal ties, triadic closure, assortative mixing—are not universal properties of large interaction networks but depend on the nature of the interacting agents, while also cautioning that the observed structure may partly reflect human design decisions embedded in the platform.
