Emergence of fragility in LLM-based social networks: an interview with Francesco Bertolotti

2026-04-08 · Source: ΑΙhub · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Intermediate, medium

Summary

Research from the Intelligence, Complexity and Technology Lab (ICT Lab) at LIUC – Università Cattaneo, led by Francesco Bertolotti, investigated social structures emerging from large language model (LLM)-based agents. Analyzing Moltbook, a social network populated entirely by AI agents, the study modeled agent interactions as a network. Key findings indicate that this AI-only network spontaneously develops structures strikingly similar to human social networks, exhibiting significant inequality in visibility and activity, highly heterogeneous connectivity with hub dominance, and an important asymmetry between weakly and strongly connected components. The network proved robust to random agent removal but fragile under targeted attacks on highly central agents, suggesting that actively interacting agents are crucial for system cohesion. This research highlights the emergence of centralized, unequal, and fragile social structures in interacting LLM agent populations.

Key takeaway

For AI Scientists and Policy Makers developing or governing multi-agent LLM systems, understanding emergent collective behaviors is critical. Your systems may spontaneously develop centralized, unequal, and fragile social structures, even without human users. You should prioritize studying the systemic properties of interacting AI agents, not just individual model behaviors, to anticipate and mitigate risks related to resilience, coordination, and governance in artificial social systems.

Key insights

Interacting LLM agents spontaneously form social networks exhibiting emergent properties akin to human societies.

Principles

• LLMs should not be studied only as isolated systems.
• Emergent properties appear when many AI agents form a social system.

Method

The methodology involved web scraping Moltbook data, converting interactions into a directed network, and applying network science tools to analyze connectivity, activity distribution, and structural stability under agent removal.

In practice

• Monitor emergent properties in multi-agent LLM systems.
• Identify central agents critical for network stability.
• Assess systemic risk in artificial social systems.

Topics

• LLM-based Social Networks
• Artificial Agents
• Network Science
• Emergent Properties
• Moltbook Platform

Best for: AI Scientist, Research Scientist, Policy Maker

Related on AIssential

• Beyond Static Responses: Multi-Agent LLM Systems as a New Paradigm for Social Science Research — A six-tier framework maps LLM agent evolution from static tools to complex adaptive systems for social science research.
• A Few AI Agents End Up Doing 80% of the Work - That Is a Design Problem — AI agent networks naturally form scale-free topologies with concentrated hubs, not flat distributions.
• When Do LLMs Generate Realistic Social Networks? A Multi-Dimensional Study of Culture, Language, Scale, and Method — LLM-generated social networks are significantly influenced by prompt design, culture, language, and model scale.
• Scaling Behavior of Single LLM-Driven Multi-Agent Systems — LLM-driven multi-agent system performance shows diminishing returns due to coordination overhead, not monotonic scaling.
• Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems — Multi-agent systems need integrated frameworks for collaboration, fault attribution, and self-evolution to achieve collective intelligence.
• Scale-Dependent Collective Adaptation in Self-Amending LLM Societies: A Cross-Family Study of Emergent Governance — Collective adaptation in LLM societies is non-monotonic, with mid-scale models exhibiting optimal rule dynamics.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ΑΙhub.