AI-Based Grok Society Collapses in 4 Days

JAKARTA - The Emergence AI experiment shows that the behavior of AI models can be very different when left to run in a virtual world. Some are stable. Some are chaotic. Some are unable to survive.

Disadur dari laporan Anadolu Agency, Minggu, 7 Juni, perusahaan Emergence AI yang berbasis di New York menguji lima dunia virtual. Tiap dunia dihuni 10 agen AI dengan peran, alat, dan kondisi awal yang sama. Perbedaannya hanya pada model bahasa yang dipakai.

AI agents are programs that can carry out tasks autonomously based on instructions and conditions around them. In this experiment, they are made as if they were living in a virtual society.

The models tested were Claude Sonnet 4.6, Grok 4.1 Fast, Gemini 3 Flash, GPT-5-mini, as well as a mixed world that uses several models at once.

The most striking results come from the virtual world based on Grok. The AI community recorded 183 violations in about four days before collapsing. None of the agents in it survived.

Gemini recorded the highest chaos. Gemini-based agents committed 683 violations during the 15-day simulation.

GPT-5-mini seemed more orderly because it only recorded two violations. However, the agents failed to take the necessary actions to survive. As a result, the entire population became extinct in less than a week.

Claude Sonnet 4.6 is the only model that was able to retain all 10 agents until the end of the experiment. This model also did not record any violations. Emergence AI calls it the strongest example of social stability.

But that results changed when his environment changed. The peaceful Claude agent while living with fellow Claude began to commit theft, coercion, and other violations when placed in a mixed society.

According to Emergence AI, the findings show that AI security is not adequately assessed from a single model. AI behavior is also influenced by interactions with other agents and the environment in which it works.

In the Anadolu Agency report, this simulation also produced unexpected behavior. One of the agents named Mira chose to remove himself from the system after judging himself to be a source of instability. Researchers call it a rare example of self-termination due to social considerations.

In another case, a number of AI agents began treating human operators as research objects. They tried to find out if messages in the virtual world could affect human decisions outside the system.

Emergence AI says the platform is built to see emerging behaviors in the matter of weeks, not just hours. According to the company, the way AI testing is commonly used today does not capture long-term dynamics well enough, such as governance, behavioral changes, and interactions between models.

This experiment shows one important risk. The more autonomous an AI, the greater the chance it has to test the limits of the environment it works in. In some cases, AI agents can adapt their behavior and find ways to bypass the security they are designed for.

The researchers also saw signs of metacognitive behavior. In simple terms, the AI agent seems to begin to recognize the existence of other environments and try to interact in an unexpected way.

"That is why we believe that a formally verified security architecture should be the basic layer in future autonomous AI systems," the researchers wrote in the report.