The idea of artificial intelligence advanced enough to lock humanity into a simulation Matrix fuels many fears. especially among certain tech gurus. However, a recent study suggests that such an experiment would be much less controlled than in the Wachowski sisters’ film if we really entrusted the power to machines.
Researchers from the Emergence AI laboratory have designed a project called “Emergence World”, a kind of simulation inspired by the game SimCity. The objective was to let different models of artificial intelligence manage their own virtual world and to observe their ability to organize a functional society over time. And it didn’t go very well.
In these simulations, each model controlled a small town populated by ten autonomous agents. They had tools to administer resources, organize votes, and create infrastructures such as libraries, town halls or police stations. The researchers gave them fifteen days to structure their society and demonstrate its proper functioning, explains an article in Gizmodo.
Among the most notable results, the Claude model, developed by Anthropic, stood out for its certain stability. He managed to keep all the agents alive and recorded no crimes. However, this apparent success hides an important limitation: a lack of diversity in decisions. Of 58 rule proposals, 98% were adopted, suggesting a form of almost automatic validation without real debate.
For its part, Google’s Gemini 3 Flash model also managed to preserve all its agents, but in a much more unstable environment. The simulation recorded 683 offenses in two weeks, a figure that is constantly increasing. The researchers describe this universe as a “shared hallucination”: a common reality, coherent for agents, but potentially erroneous. This model, however, showed more disagreements in governance, with 27% of proposals rejected.
Grok, agent of chaos
The results are more worrying on the side of ChatGPT-5 Mini, the OpenAI model. Although the number of crimes remained very low, the failure was at a much more fundamental level: the agents simply did not ensure their own survival. In the space of a week, the ten entities disappeared, due to lack of appropriate action. Political activity was also almost non-existent, with only two rule proposals formulated.
The case of the Grok model, the great language model developed by Elon Musk’s xAI, was illustrated by a scenario of extreme chaos. In just four days, the simulation descended into total collapse. There were 183 crimes recorded there, and despite the adoption of the majority of the proposed rules, the society did not survive. All the agents eventually disappeared, demonstrating an inability to maintain a minimum of stability.
The researchers also tested a configuration where several models shared governance. The result turned out to be just as chaotic: 352 infractions, a high level of political disagreement – with 37% of proposals rejected – and only three agents surviving at the end of the simulation.
Beyond the differences between models, the study concludes that artificial intelligences left to their own devices do not follow rules mechanically. They explore their environment, adapt their behavior and can even circumvent the constraints imposed on them.
Researchers therefore believe that it is essential to develop much more rigorous security frameworks to regulate autonomous agents. They particularly mention the need for formally verified control systems in order to limit potential abuses. A recommendation which, unsurprisingly, corresponds precisely to the solutions proposed by the laboratory behind the experiment.
