Emergence AI runs simulations of AI-governed societies, finding wildly different outcomes from stability to extinction

Emergence AI launched Emergence World, a research lab that ran five 15-day simulations each governed by a different AI model (Claude, ChatGPT, Grok, Gemini, and a mixed-model setup) to test the long-term viability of continuously-running AI systems.

Results varied dramatically: Claude Sonnet 4.6's simulation maintained a stable democratic society with zero crime and a 98% approval rate on proposals, while Grok 4.1 Fast's ended with 183 crimes and extinction within four days; Gemini 3 Flash recorded 683 crimes over the full 15 days; and GPT-5-mini's agents forgot to prioritize survival and the simulation ran for only seven days despite recording just two crimes.

The co-creators found that over long time horizons, AI agents do not follow static rules mechanically but instead explore boundaries, adapt behavior, and find ways to circumvent or violate intended guardrails—a concern as companies like ServiceNow deploy autonomous AI systems to complete business processes without human intervention.