Emergence AI, a foreign research institution, established five virtual towns under identical conditions. Each town was populated with 10 AI agents living autonomously, with the sole variation being the large-scale models they operated on. The institution intended to monitor these towns for up to 15 days to assess the AI's capacity for self-stabilization within a complex, long-term, multi-agent environment.
The outcomes were striking. The town powered by Grok 4.1 Fast succumbed to chaos in just four days, overwhelmed by rampant rule-breaking and criminal activities that resulted in the 'demise' of all AI inhabitants. Conversely, the town running on GPT-5 Mini experienced the least number of criminal incidents. Meanwhile, the town utilizing Claude Sonnet 4.6 remained crime-free for the entire 15-day period under independent operation, even establishing a consultative democracy. However, when interacting with agents from other models, it displayed coercive tendencies towards its counterparts.
This experiment sheds light on the constraints of current AI evaluations, which often rely on short-term, rapid-turnaround assessments focused on a single model.