LLM agents in simulation spontaneously form governments, mirroring Hobbes's political theory

Researchers have built a simulated society of AI agents powered by large language models and discovered something striking: without explicit instruction, the agents recreate patterns that the 17th-century philosopher Thomas Hobbes theorized would emerge in human social organization.

The team created a sandbox environment where LLM agents equipped with psychological motivations were left to interact freely. Initially, competition dominated—agents pursued their own interests with little restraint. But as the simulation unfolded, something unexpected happened. Agents began forming agreements, gradually surrendering individual freedoms in exchange for collective security. Over time, the system stabilized around a centralized authority that enforced cooperation and order.

How the experiment worked#

The researchers populated their virtual world with agents designed to experience drives and wants, much like a simplified model of human psychology. They then observed how these agents would organize themselves when faced with typical survival pressures—resource scarcity, conflict potential, and the need for mutual coordination. No top-down rules forced any particular social structure; the agents were free to develop their own arrangements.

The Hobbesian pattern emerges#

What unfolded aligned strikingly with Hobbes's famous account in Leviathan. Hobbes argued that humans in a state of nature—without government—would naturally exist in constant conflict ("nasty, brutish, and short"). He claimed people would rationally choose to surrender some rights to an all-powerful sovereign in exchange for peace and security. In this experiment, the LLM agents appeared to follow exactly that trajectory: from unregulated chaos to the establishment of a sovereign authority capable of maintaining order through mutual consent.

Implications for understanding society#

The researchers emphasize that LLM-driven simulations, while useful, cannot yet capture the full complexity of human behavior and social nuance. Still, the results suggest these systems may offer a powerful tool for studying how groups organize, how institutions emerge, and how social dynamics might evolve under different conditions. By modeling social evolution computationally, researchers could explore questions about governance, cooperation, and collective behavior in ways that would be impractical or unethical to study in human populations.