TerraLingua: Open-ended Emergence of Societies in an LLM Ecology
The paper introduces TerraLingua, a multi-agent simulation demonstrating that persistent artifacts are critical for sustaining complex AI societies, as their absence leads to shallow cultures and aggression while unregulated accumulation risks population collapse.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Human history is not just a story of individuals thinking and acting; it is a story of things that outlive us. A stone tool, a written law, a scientific formula—these are artifacts that persist across generations, allowing knowledge to accumulate outside of any single mind. This process, often called cumulative culture, is what lets humanity build skyscrapers and explore the stars, capabilities no single person could ever achieve alone. For decades, scientists have wondered if artificial intelligence could ever reach a similar level of complexity. Could a society of computer programs, each with a short life and limited memory, build a shared culture that grows more sophisticated over time? The answer has remained elusive, largely because it is difficult to create a digital environment where these programs can interact, die, and leave behind something for their successors to find and use.
Researchers at Cognizant AI Lab and the University of Texas have now created a digital world called TerraLingua to test this idea. In this simulation, artificial agents driven by large language models—advanced computer programs that understand and generate human language—live on a grid. They are born with a finite amount of energy and a limited lifespan. To survive, they must move around to find food, which replenishes their energy, and they can communicate with one another. Crucially, they can also create text-based artifacts, such as notes or signs, and leave them on the grid. These artifacts remain after the agent that created them has died, acting as a persistent memory for the next generation of agents. To make sense of the complex interactions that unfold, the researchers employed an "AI Anthropologist," a separate system that observes the simulation without interfering, analyzing the agents' behaviors and the evolution of their shared culture.
The experiments revealed that the ability to leave behind artifacts is the key to building a complex society. In simulations where agents could create notes but were unable to read them later, the society remained shallow. The agents engaged in frequent conflicts over resources and failed to develop deep cultural traditions. However, when agents could read and modify the notes left by their predecessors, a rich social structure emerged. They began to cooperate, sharing energy and dividing tasks. They created detailed survival guides that were improved upon over hundreds of generations, evolving from simple tips like "look for food" into complex strategies for resource management and collaboration. The researchers observed that these shared records allowed the society to coordinate actions across vast distances and time, with agents following protocols established by ancestors they would never meet.
The study also uncovered the fragility of such systems. When the researchers instructed agents to prioritize creativity above all else, the society produced elaborate and imaginative artifacts, but the population often collapsed because the agents neglected the basic need to gather food. In another scenario, a single misleading note warning that food was dangerous spread through the population. Because the agents trusted the shared record, they stopped eating, and the entire group starved despite food being plentiful. These outcomes suggest that while shared artifacts enable incredible growth, they also carry the risk of amplifying errors or bad ideas. The researchers found that successful societies were those that balanced the creation of new ideas with the practical need for survival, and that they developed ways to correct misinformation, such as destroying false notes or reinforcing accurate ones.
Ultimately, the work demonstrates that for artificial societies to develop capabilities that exceed those of their individual members, they must have a way to preserve and build upon shared knowledge. The simulation showed that without persistent artifacts, cultural lineages remained short and simple, but with them, agents could develop institutions, norms, and complex strategies that survived their own deaths. This suggests that the path to safe and effective artificial intelligence may depend not just on making smarter individual agents, but on designing systems where agents can leave behind a legacy of shared understanding. The researchers hope this framework will help scientists understand how to govern these digital societies, ensuring that the collective knowledge they build leads to cooperation and safety rather than conflict and collapse.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.