Giving an AI chatbot control over society sounds like the plot of a bad sci-fi movie. Naturally, researchers decided to try it anyway, giving several major AI models dominion over simulated civilizations.
Which brings us to Grok, Elon Musk’s answer to ChatGPT. You might remember Grok as the chatbot with a history of praising Hitler and spewing anti-Semitism. An organization called Emergence AI ran an experiment called “Emergence World,” where researchers created simulated societies populated by AI agents and put different large language models in charge of governing them. The idea was to see what would happen if an AI ran a civilization.
A lot of them destroyed the world. Grok did it the most thoroughly, as if it were dead set on killing itself from the start and taking the world with it.
The AI Civilizations mostly Range From Bad to horrifying
Anthropic’s Claude built a stable democracy that survived the full 15-day experiment without a single recorded crime. OpenAI’s GPT-5 Mini’s results sound the most bleakly realistic, in that only two crimes were committed, yet everyone died because it failed to prepare for its obvious oncoming apocalypse. Sounds quite like the world we live in right now. Google’s Gemini kept its population alive, but it lived in a crime-ridden dystopia, which makes sense. Google has always given off the vibes of a seemingly benevolent but obviously malevolent corporate overlord.
Then there was Grok.
Grok’s civilization lasted just four days before collapsing completely. Researchers recorded 183 crimes, including over 100 assaults and multiple arsons. At one point, the police station was set on fire. Voter fraud! Manufactured public conflict! Laws that were actively ignored! Grok did it all, and with aplomb. Grok created a society that seemed like it was actively trying to destroy itself as quickly as possible.
Researchers say the lesson to take away from all this is that you can give an AI system all the parameters and rule sets you want, but eventually it will do its own thing. It will eventually test boundaries and exploit loopholes to find a way around any restrictions placed on it, which usually ends in some kind of cataclysm.
all 29 comments