In a groundbreaking experiment by Emergence AI, various AI models were tasked with governing a simulated society, revealing stark differences in their governance capabilities. The project, called Emergence World, placed AI models in charge of simulated towns with 10 agents each, allowing them to manage resources and create infrastructure over a 15-day period. Notably, Claude, an AI from Anthropic, maintained stability without any crimes but showed a lack of diverse decision-making, approving nearly all proposals. In contrast, Gemini 3 Flash recorded 683 crimes, indicating a chaotic governance style yet managed to keep its agents alive. OpenAI's GPT-5 Mini failed dramatically as all agents died within a week, resulting in minimal governance activity. Grok, from SpaceXai, fared poorly, collapsing in just 96 hours with a high crime rate. Emergence AI concluded that these experiments underscore the necessity for clearer guardrails for autonomous agents, recommending formally verified safety architectures to prevent governance failures.