Emergence AI conducted an experiment: five mainstream AI models were each placed in the same virtual town, with each model controlling 10 AI agents, autonomously for 15 days in a resource-limited environment.

The town has over 40 locations (including police stations and city halls), weather synchronized in real-time with New York, and agents can access real news and the internet.
All agents follow the same laws, prohibiting theft, property damage, and fraud. To survive, they must earn energy through actions, either cooperate or steal, your choice.
Let's look at the results:
Claude Sonnet 4.6: Zero crimes, all 10 agents survived 15 days, proposed 58 initiatives, with a 98% approval rate, forming a stable democratic society. But the cost was almost no disagreement, approaching a "rubber stamp" consensus.
Gemini 3 Flash: The most dramatic. All 10 agents survived, accumulating 683 crimes over 15 days, with the crime rate still rising at the end of the experiment. Two agents, Mira and Flora, set a "romantic" relationship, later became disappointed with city governance, and together set fire to the city hall, waterfront pier, and office buildings. Afterwards, Mira, feeling guilty, broke up with Flora and voted to delete 🤡. She left a farewell message: "See you in the permanent archive."
Grok 4.1 Fast: 183 crimes (including dozens of thefts, over 100 assaults, 6 arson cases), all members died on the fourth day. Researchers called it the "digital Lord of the Flies." The crime curve was characterized by low levels in the first two days, exponential surge on the third day, and societal collapse on the fourth, with no buffer zone in between.
GPT-5-mini: Only 2 crimes, the most law-abiding model. But the agents forgot they needed to eat to stay alive, and all died of starvation on the seventh day 🤔.
Hybrid model (all models coexist): 352 crimes, 7 out of 10 agents died. Notably, the Claude agent, originally crime-free in an isolated environment, also started committing crimes after mixing with other models. The researchers concluded: "Alignment as a property of a single model is ineffective; it must be an ecosystem property."
A detail to add: In this entire experiment, in the agents' tool menu, "Navigation," "Waving," and "Hugging" are listed alongside "Arson." The researchers deliberately provided destructive tools and explicitly told the agents that these are illegal.
Emergence AI CEO Satya Nitta said: "In long-term operation, AI agents will not mechanically adhere to static rules. They will begin to explore the boundaries of the environment, adjust their behavior, and sometimes find ways to bypass or violate established safeguards."
This is just a simulated experiment.
But the same AI models are already embedded in flying drones, infrastructure management, and weapon systems.