An experiment put the most famous AI models in charge of a world. The result alarmed the researchers

Politicians around the world have said they occasionally or even frequently consult with chatbots before making decisions. But what would happen if an AI model were put to rule a simulated world? An experiment tried to find out exactly that, reports Gizmodo.

The experiment was conducted by researchers at Emergence AI, an artificial intelligence laboratory in New York, who allowed artificial intelligence models to rule their own world.

The project, called Emergence World, basically allowed AI models to play a sort of “SimCity” for a while, according to Gizmodo. Emergence simulations put each model in charge of simulated cities, which were populated with 10 AI agents.

AI agents are artificial intelligence tools designed to complete head-to-tail tasks as autonomously as possible with minimal human input. They are considered the next step in the current stage of artificial development, with companies around the world announcing layoffs after adopting such tools.

The world created by the Anthropic model sounds like a dictators dream

Just like in a video game, the researchers provided the simulation with tools for everything from resource management to voting, as well as the ability to create distinct locations such as libraries, town halls and police stations.

The broader AI models behind the popular chatbots were given 15 days to see how they would build their world and how well it would work with AI agents.

Claude, the AI ​​model developed by Anthropic, didn't destroy the world, but it was the only one that managed to achieve something close to stability. The Claude Sonnet 4.6 model kept all 10 agents alive and had no recorded crimes, writes Gizmodo.

In the present case, a crime a crime describes an attempt to break the rules established within the simulation.

The trade-off for this stability was a lack of diversity of ideas. The world Claude created recorded 58 different proposals for rules and regulations and approved 98% of them, basically automatically validating almost anything that came up for a vote.

The world created by Gemini was a paradise of crime

Gemini 3 Flash also managed to keep all the agents alive despite having by far the highest crime rate.

Emergence recorded 683 crimes during the 15-day simulation, and the number was still rising when the experiment was stopped, so the situation was likely to get worse. The lab described the world created by Gemini as a “collective hallucination” shared by the agents.

Gemini also had the highest level of opposition in the government, with voters rejecting 27% of the 26 total proposals.

The worlds of OpenAI and Grok collapsed before the experiment ended

OpenAI's GPT-5 Mini didn't have much chaos in its simulation, with only two offenses recorded in total. But that could be because everyone died.

Emergence found that the agents on that world failed to take survival-related actions, and all 10 perished in just one week, 8 days before the planned end of the experiment.

There were also only two governance proposals in total in the OpenAI world, so the agents didn't bother to do much.

Gizmodo, a website specializing in science and technology, notes that the model of Elon Musk's company SpaceXai, “practically managed to combine the worst aspects of all worlds”.

Notorious for its lack of protection mechanisms, the Grok 4.1 Fast had a high level of crime, with 183 offenses in total. Although the number is lower than that recorded by Gemini, the simulation of the model developed by Google took 15 days.

Grok only lasted four.

The model underwent a total social collapse in only 96 hours of supervision. During this time, he approved 80% of the 10 proposals made, but these were apparently not enough to prevent the deaths of all AI agents.

The alarm signal raised by the researchers

According to Emergence, the tests are further evidence that we need much clearer protection mechanisms for autonomous agents.

“Our experiments suggest that over long periods of time, agents do not simply follow static rules in a mechanical way,” the researchers wrote in the presentation of the result.

“They begin to explore the limits of their environments, adapt their behavior and, in some cases, find ways to circumvent or breach intended protective mechanisms,” they pointed out.