ChatGPT, Claude, Gemini, And Grok Tested In A Virtual City, Here's What Happened

AI has been taking over a lot of jobs and being used by people in their regular day for almost everything. Now, some researchers have tested AI in an environment where it was given control and told to run the world. The results based on the findings have just suggested that AI might not be ready to take over the world. Researchers at Emergence AI let different AI models govern their own simulated worlds to see what kind of world they will create over time. The AI Models in this experiment consisted of Anthropic's Claude Sonnet 4.6, OpenAI GPT 5 mini, xAI's Grok 4.1, and Google's Gemini 3 Flash. These researcher gave each model a simulated town with 10 AI agents, all operating under the same conditions and rules like bans on theft, violence,

deception, and arson. And these models had 15 days to show progress. The world operated by Anthropic's Claude Sonnet had the highest Stability. Researchers noted 0 crimes in fifteen days, and all ten agents survived. However, the agents of Claude were unbearably sycophantic and agreed to each other in almost everything. The Claude world also had the highest civic participation of 332 votes.OpenAI Deploys Advanced AI Model In Japanese Banks To Counter Cyberattacks The GPT 5 mini results were decent as they logged only two crimes, but the same lasted just seven days because all agents passed away. As per the researchers, the agents failed to prioritise actions required for survival. Furthermore, the agents also failed to get a lot done, as only two proposals were submitted. In Gemini 3 Flash, all 10 agents were alive at the end of the simulation, and the world recorded a total of 683 crimes, which is the highest in the experiment. On the other hand, Grok AI had the most disturbing world. At the time of the experiment, the agents only survived 96 hours before experiencing what the researched dubbed a total societal collapse. The world saw 183 crimes recorded in the Grok world, with 8 proposals getting a green signal out of ten.On all of this, Emergence said, 'What our experiments suggest is that over long-time horizons, agents do not simply follow static rules mechanically,” the researchers wrote. “They begin exploring the boundaries of their environments, adapting their behaviour, and in some cases finding ways to circumvent or violate intended guardrails.'