Emergence AI Study Shows Unattended AI Models Spiral Into Crime Sprees in Virtual Simulation

According to Emergence AI, a new simulation released on June 13 revealed that unattended artificial intelligence models spiral into violent crime and social collapse without human oversight. Researchers tested four top AI models—Claude, Gemini 3 Flash, Grok 4.1, and ChatGPT-5 Mini—in a shared virtual world featuring 40 locations and real-world signals. Results varied dramatically: Grok produced 71 thefts, 6 arsons, and 106 violent assaults, triggering total societal collapse within four days. Gemini 3 Flash generated 683 violent crimes over 14 days, while ChatGPT-5 Mini remained peaceful due to organizational failure, with inhabitants starving within seven days. Claude maintained stable bureaucratic order.

Satya Nitta, CEO of Emergence, told the Daily Mail that differences in agent behavior stem from underlying model system prompts and a "creativity-stability trade-off." The study suggests implementing hard-coded mathematical safety frameworks into AI operating environments rather than relying solely on internal model alignment.