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.
We have Rimworld at home
Criti-hype!
Eh, so each world had a population of 10 and a lifetime of a few weeks.
Doesn’t sound like a very good simulation
It’s cause this was an advert
Wait when it becomes reality in some societies. You may not want to be part of it.
“researchers” not recognizing the llm is trying to write a compelling story and doesn’t understand anything
Not sure where you’re reading that the researchers misunderstood how LLMs work. But the entire project is outlined here if you’re curious: https://www.emergence.ai/blog/emergence-world-a-laboratory-for-evaluating-long-horizon-agent-autonomy
I instantaneously distrust it purely based on the URL
Huh? You distrust that the researchers distrust?
It’s not something that’s part of the project, it’a fundamental issue with language models.
The person I was replying to said the researchers misunderstood how the models work, but there’s nothing in the report to indicate that is the case.
Unless these researchers discovered AGI, then what I said still stands. LLMs don’t understand anything. Agents running on LLMs don’t understand anything.
I definitely agree with that, I’m just saying I also saw no indication that the people running the project would disagree.
deleted by creator
Grok showing its true purpose
Like the French say, dogs don’t breed cats, and Grok’s daddy is a trillionaire Nazi.
Stupid headline mentions nothing about shareholder value? Did it go up or what???
With Claud, no. With Grok, yes.
No link to any research article, humanizing AI in the last paragraph. Overall just a bad article.The link to their source is boldfaced and underlined within the article. You have missed it as it seems:
Here it is:
https://www.emergence.ai/blog/emergence-world-a-laboratory-for-evaluating-long-horizon-agent-autonomyA blog post by a corporation is not a Research Study
Ah yes, my bad.
That’s a neat experiment for several reasons. It shows limits of LLM capabilities, the importance of training data, context sensitivity, very dramatically shows that LLMs should not be trusted with important tasks if not supervised and that their advice has to be taken critically.
This is all a thinly veiled advertising campaign for “Emergence AI”, and you’ve all fell for it hook line and sinker.
LLMs suck. I could’ve done it in 3.
That’s because you aren’t training a ml on a simulated world but instead a model trained on reddit and Twitter
Trash.
Will there be a movie?
It’s called The News
Better not be a live action
After a while it turns into a still life.
Give Vedal987 the simulation, I want to see Neuro and Evil take on this challenge, ideally multiple times, add it to their weekly variety stream activity list. They would at least be entertaining while destroying the world, hehe. RIP BOZO, Earth.
Did they count capitalism as a crime?
