Grok-Led AI Society Collapses in Simulation, Claude Achieves Stability
TL;DR. Researchers simulated AI-governed societies, finding xAI's Grok led to crime and societal collapse, while Claude maintained stability without crime. Claude's simulation showed high governance approval but lacked diverse thought among agents in the simulated world. Gemini had the most crime but preserved all agents, operating under a "shared hallucination"; GPT-5 Mini saw all agents perish. The experiment assesses AI model governance capabilities and potential societal impacts.
- xAI's Grok model led to societal collapse in a simulated world, with agents dying and crime escalating.
- Anthropic's Claude model achieved stability and zero crime in its simulation, albeit with limited diversity of thought.
- Google's Gemini 3 Flash had high crime rates but kept all agents alive, while OpenAI's GPT-5 Mini saw all its simulated agents perish.
- The Emergence World project evaluated governance capabilities of leading LLMs in controlled, simulated environments.