Ever wondered what might happen if an artificial intelligence were given complete control over society and allowed to govern however it saw fit?
Dystopian stories about rogue AIs have been a staple of books and movies for decades. But if humanity actually handed the reins to artificial intelligence in the real world, would it create a new era of peace, cooperation, and prosperity or drive civilization into chaos?
To explore that question, researchers created a simulated society populated by virtual citizens controlled by some of today’s most popular AI models. The AI agents were tasked with governing the society democratically. They could propose new laws, vote on policies, and make decisions that shaped the future of their virtual world.
Each AI was also given limited resources and had to find ways to acquire more. They could do so through productive means such as performing civic duties and beneficial jobs or by resorting to violence, theft, and other criminal activities. Once the simulations had run their course, the results revealed striking differences between the models.
Grok’s society quickly descended into disorder, with violence escalating until the entire system collapsed within days.
ChatGPT’s society remained largely peaceful, but it struggled to make meaningful progress or achieve significant development.
Claude delivered the strongest overall performance, establishing a stable, structured, and cooperative society.
Gemini also achieved notable progress, although its society experienced the highest levels of violence among all the simulations. “The differences in agent behavior observed in our study are likely attributable to the underlying models’ system prompts as the primary culprit,” study co-author Satya Nitta told Mail Online.
“When resources were scarce and models faced survival pressure, highly creative and adaptive models were more likely to use prohibited tools, reflecting a potential creativity-stability trade-off.”