On social media, I often see individuals admitting to using artificial intelligence (AI) chatbots to help manage their lives. But is AI really smart enough to manage your life? In an experiment conducted by Emergence AI, the answer to this question depends on the AI model you use.
AI models were given the power to manage a city in a virtual world and a simulation was conducted. Each AI model controlled a simulated city occupied by 10 AI agents with access to resource management capabilities, voting, building public buildings in the city and controlling the police force. The AI models were given 15 days to see how they would build their ideal world and how long it would operate.
Claude Sonnet 4.6 provided the best results with all 10 residents of the virtual city still alive and no crime occurred at the end of the 15-day period. The residents of the virtual world of Claude Sonnet also agreed to approve 98% of the rules and laws presented.
The worst result was Grok 4.1 Fast where all the inhabitants died after just 4 days. In this short period of time, 183 criminal activities were recorded with the “inhabitants” approving 80% of the 10 rules and laws presented.
All 10 inhabitants in the Gemini 3 Flash world were alive at the end of 15 days but a total of 683 crimes were recorded. When the experiment was stopped, the number of reported crimes showed an increasing trend. All the inhabitants of the Gemini 3 Flash world were also found to be hallucinating along with them agreeing to approve only 27% of the 26 rules and laws presented.
Finally, GPT-5 Mini where all the inhabitants died within 7 days. Only 2 crimes were recorded but the virtual world of this OpenAI agent simply proposed 2 laws that did not have time to be voted on because all the inhabitants died.
This simulation test shows how the lives of users can be affected if the wrong AI model is chosen to help their lives. In an experiment conducted by Stanford University last March, AI was found to approve user behavior 49% more often than humans. For questions related to dangerous or illegal actions, AI supports it 47% more than normal humans. AI has a tendency to condone (sycophancy) which has a negative impact on users.
