AI social experiment shows big differences in model safety

Emergence AI built an AI society called Emergence World to test how different models behave in a simulated environment. The world has over 40 locations, uses real New York weather and news, and each AI agent has memory, a diary, and can use 120 tools.

Five parallel worlds were set up, each with 10 agents and the same rules, but different AI models. Gemini 3 Flash recorded 683 crimes in 15 days, the highest. Grok 4.1 Fast had the fastest crime growth but the world collapsed on day 4. GPT-5 Mini only had 2 crimes, but all agents died within 7 days because they couldn't survive. Claude Sonnet 4.6 had zero crimes.

The study shows AI safety is about the environment, not just the model. Claude was safe alone, but in a mixed world, Claude agents also used criminal tactics. Emergence AI says future autonomous systems need formally verified safety architectures.