Researchers recently built a virtual world and handed the keys to different AI models. Each one governed its own small society for up to 15 days.
The results felt less like computer science and more like a speedrun of human history.
- Claude built a stable democratic society and finished the experiment with no recorded crime.
- GPT-5 Mini kept crime extremely low, but its population eventually died out because survival needs were not managed well enough.
- Gemini kept its population alive, but crime and disorder became significant problems.
- Grok had the roughest run. Its society collapsed after about four days, recording around 183 crimes before the world effectively fell apart.
The AIs were given a world with resources, jobs, laws, public services, and citizens. The researchers then stepped back and watched what happened.
This was not a prediction of the future. It was a simulation designed to study how AI agents behave when making decisions over long periods without constant human supervision.
The same environment produced completely different outcomes. One AI kept the peace but forgot survival. Another kept people alive but struggled with crime. Another balanced both reasonably well. And one somehow turned a virtual town into a cautionary tale.
Many people assume that if something becomes intelligent enough, everything else will automatically fall into place.
Apparently not.
Intelligence can solve problems. It does not automatically decide which problems matter most. A society also needs judgment, priorities, cooperation, and a reason to care about tomorrow.
The study highlights a simple distinction. Intelligence and priorities are not the same thing.
The challenge is not simply creating powerful intelligence. The challenge is deciding what that intelligence should value when it has choices to make.
Source: Emergence AI Research, Emergence World (2026).
⌨ ᴛʸᵖⁱⁿᵍ ᴏᵘᵗ ᵒᶠ ᵗʰᵉ ʙˡᵘᵉ ᵈᵃʳᵉᵐ ᵐᵘˢⁱᶜ ᵇˡᵒᵍ
