ChatGPT, Gemini, Claude and Grok all created virtual worlds with simulated societies — one built a democracy, the others descended into chaos

To better understand what increasingly autonomous AI systems might look like in practice, researchers at Emergence AI created a set of virtual societies and put different AI models in charge. The experiment offered a glimpse of how these systems behave when they’re asked to do more than answer questions, and instead must navigate the challenges of governing a community.

Inside the experiment that let AI govern itself

Emergence AI is a self-described “frontier AI lab turning cutting-edge agentic research into enterprise infrastructure.” And some of their most recent research had some off-putting findings.

In May, Emergence launched “Emergence World,” a research lab dedicated to “studying how autonomous agents behave when the time horizon is long enough for compounding effects, social dynamics, and behavioral drift to matter.”

In other words, they created simulated worlds, each run by a different AI model — Grok, Claude, Gemini, OpenAI and a mixed model. The AI agents were set up in identical, parallel worlds that “reflected real world complexity.” The worlds were grounded in real world conditions, with weather synced to that of New York City and access to real time global news. They had everything from town halls to libraries to police stations.

Each model had the same starting conditions, as well as strict rules prohibiting crimes like theft and destruction. Once ready, they let the simulation run for 15 full days.

“Within days, they diverged dramatically,” Emergence said.

One AI model built a democracy. Another burned down the police station

The models quickly developed distinct social structures and behaviors. Some built relatively stable communities, while others experienced rising crime, institutional breakdown and, in some cases, complete societal collapse.

Here’s a quick look at some of the notable results from the research:

• Claude managed to form a stable democracy with no violence.
• Grok cited 204 criminal events, including burning down the police station, and eventually total collapse and extinction.
• OpenAI failed to form a functioning society, which resulted in the simulation and its agents all dying off.
• Gemini cited 683 crimes and “exhibited the highest levels of emergent disorder with repeated late-stage escalation dynamics.”
• The mixed model world was stable in insolation but became unpredictable when they interacted with agents built off other models. Seven of their agents died.

The prevalence of crime, instability and institutional breakdown across multiple simulations stood out as one of the study’s most notable findings. According to PwC’s AI Agent Survey, 35% of companies are broadly adopting agentic AI and 27% are adopting it on a limited basis. As agentic AI becomes more common in workplaces and other real-world settings, studies like Emergence World offer a glimpse into how these systems might behave when given greater autonomy.

In summing the results, Emergence emphasized the importance of their research: “All of this matters because AI is moving beyond tools into systems that operate autonomously in the real world … The challenge is no longer just performance. It’s predictability, safety, and trust over time.”