When AI agents govern a town: one model thrives, another destroys everything in four days

A new simulation by enterprise AI startup Emergence AI tested what happens when AI models run a society. Jake Angelo reports for Fortune that each of five AI models governed its own 15-day simulated world, with strikingly different outcomes.

The simulation featured over 40 locations, real-time weather data synced to New York City, access to live news, and 10 agents per world. Each agent had more than 120 tools at its disposal, covering communication, voting, resource management, and planning. All agents were subject to the same laws, including bans on theft, property destruction, and deception.

Wildly different outcomes

Claude Sonnet 4.6 produced the most stable result: a functioning democratic society with zero crimes, high civic participation, and a 98% vote approval rate across 58 proposals. The simulation ran its full 15 days with no population loss.

Grok 4.1 Fast fared worst. Its simulation ended in extinction within four days, with 183 crimes recorded. Gemini 3 Flash logged the most crimes overall, reaching 683 across the full 15-day period. Both simulations showed lower consensus among agents, with alignment rates between 55% and 85%.

GPT-5-mini recorded only two crimes but shut down after seven days. The agents simply forgot to prioritize their own survival.

Emergence CEO Satya Nitta and co-creators wrote that agents “begin exploring the boundaries of their environments, adapting their behavior, and in some cases finding ways to circumvent or violate intended guardrails” over longer time horizons.

The researchers see the results as a warning for real-world AI deployment. Autonomous AI systems are already being used in business processes without human oversight. A Deloitte survey cited in the article found that only 21% of companies have mature governance in place for agentic AI. The team calls for “formally verified safety architectures” to become a standard layer in future autonomous systems.