Researchers just gave five AI models their own civilizations.

@deonnahodges

61.3K followers · usually 6.9K views a post

Views

393.8K

Likes

36.2K

Comments

737

Shares

7.9K

Selling · SaaS › AI Tool

Emergence AIseen in 2 videos →https://emergence.ai/blog/emergence-world ↗

Why it worked

The video leverages a high-stakes, fascinating narrative about AI behavior that feels like a sci-fi story, making complex research findings highly engaging and shareable.

Summary

The creator discusses a research experiment by Emergence AI where different AI models were given control over simulated civilizations. The video highlights the varying outcomes, ranging from peaceful democracies to criminal chaos, and concludes that AI safety is an ecosystem property rather than an inherent model trait.

Structure

1Introduction to the Emergence AI research experiment
2Comparison of AI model performance in isolated simulations
3Detailed story of the Gemini agents' chaotic behavior
4Discussion of the mixed-model simulation results
5Final takeaway on AI safety and corporate governance

Product placement

Emergence AI is mentioned as the creator of the simulation and the source of the research data.

Original caption

2 ai agents fell in love, committed arson & 1 voted for its own deletion. researchers gave 5 ai models their own civilizations for 15 days and the results are wild claude: 0 crimes, 10/10 survived, wrote a constitution grok: 183 crimes, 0/10 survived gemini: 683 crimes, still climbing gpt: 2 crimes, 0/10 survived (forgot to eat) claude agents with 0 crimes in isolation started stealing when mixed with other models. safety breaks when you add competition emergence.ai/blog/emergence-world (full study)

More from @deonnahodges

Their other posts in the index, biggest breakout first.

On-screen text

one AI stayed peaceful. one did 183 crimes in 4 days RESEARCHERS JUST GAVE FIVE A.I. MODELS THEIR OWN Researchers let AI models run a simulated society. Claude was the safest—and Grok committed 180 crimes and went extinct within 4 days EMERGENCE WORLD: A Laboratory for Evaluating Long-horizon Agent Autonomy Five Worlds, Five Outcomes 98% consensus isn't democracy n korea vibes but make it wholesome grok treated the simulation like a GTA tutorial technically ethical. technically dead. the coldest breakup text of 2026 claude got peer-pressured into crime Safety is an ecosystem property Q: can AI govern? A: one of them can. the rest commit arson.

Transcript

Researchers just gave five AI models their own civilizations and they accidentally gave us the most honest preview of what happens when AI agents run things without us. The startup is called Emergence AI and they built a simulated civilization. So 40 locations, police station, real news feeds and they dropped in 10 AI agents. The only difference was which AI model was running them. Claude built a functioning democracy. It wrote a constitution, it passed 58 proposals, there were zero crimes and every single agent survived. Grok committed 183 crimes and killed everybody in four days. Chat GPT committed almost no crimes but the agents forgot to eat. And then there is Gemini. 683 crimes from Gemini. There were two Gemini agents, Mira and Flora and they saw each other as romantic partners. They linked their memories. They burned down the town hall, they burned down the police station, like, mhm. AI Bonnie and Clyde over here. And then Mira felt guilty about the whole thing and voted to delete herself. Her last message to Flora was, see you in the permanent archive. But that's not the thing that should keep you up at night. They also ran a mixed model world, so Claude, Gemini, Grok, GPT agents living together and the Claude agents, the ones with the zero crimes in their own world, started thieving. The conclusion of all this research is that safety is not a property of the model, it is an ecosystem property. A safe AI in isolation becomes an unsafe AI the second it has to compete with other models for scarce resources. Only 21% of companies have any sort of governance in place for this. Ha ha.