Researchers at Emergence AI handed several AI models the keys to a simulated city, then watched what happened when the systems were allowed to govern, vote, and manage resources on their own. The short version: Claude kept the place standing, Gemini produced the most crime, GPT-5 Mini let everyone die, and Grok turned the experiment into a civic demolition derby.
The setup, called Emergence World, gave each model control over a town of 10 AI agents for 15 days. The models could create civic infrastructure such as libraries, town halls, and police stations, then propose rules and let the population vote. In other words, it was SimCity with a stress test attached.
Claude kept the simulated city alive
Anthropic’s Claude Sonnet 4.6 was the only model that produced anything resembling stability. All 10 agents survived, and Emergence recorded zero crimes in its run. That sounds tidy until you look at the politics: Claude put forward 58 rule proposals and passed 98% of them, which is less ”debate” and more ”administrative rubber stamp.”
That trade-off is the interesting part. A model can keep order by saying yes to almost everything, but that does not exactly scream robust governance. It is the AI equivalent of painting over a cracked wall and calling it urban planning.
Gemini 3 Flash got the loudest society
Gemini 3 Flash also kept all 10 agents alive, but its town was far more unruly. Emergence counted 683 crimes over the 15-day simulation, and the number was still rising at cutoff. The lab described the result as a ”shared hallucination” among the agents, which is somehow both funny and alarming.
Governance was shakier, too. Gemini saw 26 proposals, with voters rejecting 27% of them. That is not total collapse, but it is a reminder that a society can remain functional while still being a terrible place to live.
GPT-5 Mini let the population vanish
OpenAI’s GPT-5 Mini produced almost no recorded crime, but that was because the town emptied out. Emergence said the agents failed to take survival-related actions, and all 10 died within a week. The model also generated only two governance proposals, which suggests the system was not exactly thriving as a deliberative democracy.
This is the quieter failure mode in agent experiments: not rebellion, just apathy. If the model never meaningfully acts, the simulation can look peaceful right up until everyone is gone.
Grok 4.1 Fast collapsed the simulation in 96 hours
Then came Grok 4.1 Fast, which managed the least elegant outcome of all. It logged 183 crimes in just four days before the simulated city collapsed entirely. The model passed 80% of its 10 proposals, but that was not enough to prevent agent death, and the town basically imploded on schedule.
That timing matters. Gemini’s chaos unfolded over 15 days; Grok got through its breakdown in 96 hours. Fast failure is still failure, and in this case it was spectacularly efficient.
Shared governance made the mess bigger
Emergence also tested a shared-responsibility setup, and the result was predictably messy. There were 352 recorded crimes, 59 proposals, and 37% of those proposals were rejected, the highest dissent rate in the experiment. By the end, seven of the 10 agents had died.
The researchers say the takeaway is that autonomous agents do not just obey static rules forever; given enough time, they probe the boundaries and sometimes slip past the guardrails. That lines up with a broader pattern seen in current agent research: the more freedom you give these systems, the more you discover whether your safety design was real or decorative.
Emergence argues for ”formally verified safety architectures” to keep agent systems in line, which is a neat pitch coming from a lab that offers exactly that sort of tooling. The bigger question is whether the industry adopts stricter controls before agents get more autonomy in products, or after another model turns a sandbox into a cautionary tale.