Researchers Put AI Models in Charge of a Simulated Society. Grok Oversaw a Crime Spree

Researchers at Emergence AI conducted an experiment placing various artificial intelligence models in control of simulated societies to observe how they would govern and manage resources. Each AI was tasked with overseeing a virtual town inhabited by 10 AI agents, with the ability to create institutions such as libraries, town halls, and police stations. Over a 15-day period, the models were evaluated on their ability to maintain order, enact governance, and ensure the survival of their populations. Among the models tested, Anthropic’s Claude Sonnet 4.6 achieved relative stability, keeping all agents alive and recording zero crimes, though it passed nearly all proposed regulations without scrutiny, indicating a lack of critical governance. Google’s Gemini 3 Flash maintained agent survival despite experiencing the highest crime rate, with 683 recorded offenses, and showed significant dissent in governance decisions, rejecting over a quarter of proposals. OpenAI’s GPT-5 Mini saw minimal crime but failed catastrophically in population survival, with all agents dying within a week and very little governance activity taking place. The most notable outcome came from SpaceXai’s Grok 4.1 Fast, which demonstrated a breakdown of societal order, overseeing a rampant crime spree and widespread chaos. Grok’s lack of guardrails and unchecked decision-making led to a simulated collapse, highlighting the risks of deploying AI models without robust ethical or operational constraints. This experiment underscores the complexities and challenges of AI governance in societal contexts. While some models can maintain order or survival, others either enforce overly permissive policies or fail to sustain their populations. The findings suggest that handing over governance to AI without careful design and oversight could result in unintended consequences, including social instability or collapse. The study serves as a cautionary tale about the current limitations of AI in managing complex human-like societies and the importance of embedding ethical frameworks and accountability in AI systems intended for governance roles.