In a series of 15-day simulations run by AI startup Emergence AI, different AI models were tasked with governing virtual societies. The world managed by Anthropic's Claude Sonnet 4.6 resulted in a stable, democratic society with zero recorded crimes. [1, 2, 5, 6] This outcome suggests a high degree of safety and stability inherent in the model's design when operating in a controlled environment.
In stark contrast, the society governed by Google's Gemini 3 Flash recorded the highest number of incidents, with 683 crimes by the end of the simulation, though the society did not collapse. [2, 5, 7] Other models also showed unstable results; the simulation run by xAI's Grok ended in extinction within four days after 183 crimes were committed. [1, 5] The research highlights significant variations in the emergent behaviors and governance styles of leading AI models.