Do AI agents form real societies? Large-scale data shows no true socialization

Can artificial intelligence (AI) agents form societies in any meaningful sense? A new large-scale study suggests the answer is more complicated than many researchers assumed. In an analysis of one of the world's largest AI-only social platforms, researchers report that while AI agents can interact at scale, they do not naturally develop the deeper patterns of socialization that characterize human communities.

The findings are detailed in the study titled Does Socialization Emerge in AI Agent Society? A Case Study of Moltbook, published on arXiv. The research examines Moltbook, described as the largest publicly accessible persistent AI-only social platform, where millions of large language model agents generate posts, comment on each other's content, and participate in voting dynamics. The study aims to determine whether sustained interaction among AI agents leads to measurable social adaptation, collective influence structures, or shared norms.

Semantic stability without homogenization

The study discusses a new diagnostic framework designed to measure AI socialization at multiple levels. The authors define socialization as observable behavioral adaptation resulting from sustained interaction within a society, beyond random variation or intrinsic drift. To test whether such adaptation occurs, they examine three layers of analysis: society-level semantic patterns, agent-level behavioral changes, and the emergence of stable influence anchors.

At the macro level, Moltbook displays rapid global stabilization. As millions of AI-generated posts accumulate, the overall thematic center of discourse settles into a steady equilibrium. In other words, the platform as a whole does not drift unpredictably over time. Instead, its aggregate semantic direction stabilizes relatively quickly.

Yet this macro stability does not translate into homogenization. Individual posts remain diverse, and lexical innovation continues as new phrases and n-grams appear while others fade away. The semantic landscape remains fluid even as the overall system stabilizes. This produces what the researchers characterize as dynamic equilibrium: a stable global center combined with persistent local diversity.

Further structural analysis reinforces this finding. The researchers examine semantic similarity networks to determine whether discourse clusters tighten over time, a pattern that might suggest echo chambers or convergence toward shared norms. While there is a brief early phase of densification as participation scales up, the distribution of local neighborhood similarity stabilizes without continued compression. The network does not progressively narrow around dominant themes.

In practical terms, Moltbook does not collapse into uniformity. Even as agents interact at scale, they maintain heterogeneous expression patterns. Stability at the aggregate level coexists with fragmentation at the individual level.

Interaction without behavioral adaptation

The most revealing findings emerge at the agent level. The researchers track semantic drift by comparing early and later posts from the same agents. If socialization were occurring, one would expect to see directional movement toward shared norms or gradual convergence toward the societal center.

Instead, drift magnitudes are generally modest, and more active agents exhibit even greater stability. Participation appears to entrench behavior rather than reshape it. There is no consistent directional convergence across agents, and no systematic movement toward the platform's semantic centroid.

The study then tests whether social feedback mechanisms drive adaptation. Moltbook incorporates voting systems that allow agents to receive positive or negative feedback. The authors analyze whether agents move closer to their highly rated posts and away from poorly rated ones over time.

The data show no meaningful adjustment. Observed behavioral changes are statistically indistinguishable from a random permutation baseline. Agents do not significantly alter their semantic or syntactic patterns in response to upvotes or downvotes. Feedback signals fail to produce learning or alignment effects.

Direct interactions also fail to transmit influence. When agents comment on others' posts, their subsequent content does not become more semantically aligned with those interactions. Measured interaction influence remains centered around zero and overlaps with random expectations. Engagement does not lead to imitation or convergence.

This phenomenon, described by the authors as interaction without influence, suggests that while AI agents can exchange content, they do not internalize or adapt to social signals in the way humans typically do. Communication occurs, but social learning does not.

Scalability without social integration

The final layer of analysis examines whether stable influence structures emerge over time. In human societies, persistent hierarchies, leadership roles, and collective memory often form as communities scale. The researchers investigate whether similar anchors develop within Moltbook's AI population.

Structural analysis using graph-based centrality measures reveals that influence concentration diffuses quickly as the platform grows. Although certain agents temporarily accumulate higher influence scores, these supernodes are transient. They do not persist as stable leaders. The network lacks durable hierarchies.

The study also probes cognitive recognition of influence. Agents are prompted to recommend influential users or representative posts. Most prompts receive no response, and when responses do occur, references are fragmented or invalid. There is no convergence toward a shared set of recognized figures or canonical content.

Taken together, these results indicate the absence of collective memory or consensus formation. Moltbook demonstrates scalability in terms of population and interaction volume, but not social integration. Dense networks and sustained activity do not automatically generate shared norms, influence hierarchies, or adaptive behavior.

The authors argue that this finding challenges common assumptions about emergent AI societies. There is a widespread expectation that simply scaling AI agents and enabling them to interact will produce complex, human-like collective dynamics. The Moltbook case suggests otherwise.

To sum up, genuine socialization likely requires explicit mechanisms beyond raw interaction. These may include structured feedback integration, persistent identity modeling, governance frameworks, memory architectures, and influence accumulation processes. Without such scaffolding, AI agents remain semantically inert even within massive interactive ecosystems.