Artificial Leviathan: Exploring Social Evolution of LLM Agents Through the Lens of Hobbesian Social Contract Theory

Imagine a digital sandbox where scientists don't just build bridges or code games, but try to grow entire societies from scratch. This is the world of computational social science, a field where researchers use computers to simulate how groups of people (or in this case, computer programs) interact, fight, and cooperate. To understand this specific experiment, you need to know two big ideas. First, there's the concept of the "State of Nature," a famous thought experiment by philosopher Thomas Hobbes. He imagined a world before governments existed, where everyone was free but also terrified, leading to a constant "war of all against all" because resources were scarce. Second, there's the "Social Contract," the idea that people eventually agree to give up some of their freedom to a powerful leader (a sovereign) in exchange for safety and peace. Scientists care about this because if we can understand how these ancient human patterns emerge in a computer simulation, we might learn how real societies form, why conflicts happen, and how order can rise from chaos.

In this paper, a team of researchers built a digital playground filled with Artificial Intelligence agents powered by Large Language Models (LLMs)—the same kind of smart technology behind chatbots. They didn't program these agents with complex rules telling them to be good or bad. Instead, they gave them simple, primal drives: they needed food to survive, they had land to farm, and they had personality traits like "aggressiveness" or "desire for peace." The agents started in a "State of Nature," a chaotic environment where they could farm, trade, or rob each other. At first, just like Hobbes predicted, the agents were wild. They spent most of their time fighting and stealing resources, creating a messy, violent world.

However, as the simulation ran day by day, something magical happened. The agents began to learn from their experiences. They realized that constant fighting was exhausting and dangerous. Slowly, they started forming "social contracts." A weaker agent would choose to give up some of their food or land to a stronger one, not because they were forced, but because they wanted protection. In exchange, the stronger agent promised to defend them from other robbers. This wasn't a one-time deal; it created a chain of loyalty. Eventually, all the agents in the simulation, directly or indirectly, pledged their loyalty to a single "Sovereign" agent. The result was a "Commonwealth": a peaceful society where violence dropped dramatically, and agents spent their time farming and trading instead of fighting.

The researchers tested how robust this story was by tweaking the rules. They found that if the agents had a very short memory (only remembering the last day), they couldn't learn to cooperate and the society stayed chaotic. But if they could remember more of their past interactions, they figured out the benefits of peace much faster. Interestingly, they also found that if the agents were too predictable in their thinking, they got stuck in loops of violence and never formed a society. The study suggests that LLMs, when given the right mix of survival instincts and memory, can naturally evolve complex social structures that mirror human history, moving from a state of nature to a peaceful commonwealth without anyone explicitly programming them to do so.