Forgot password
Enter the email address you used when you joined and we'll send you instructions to reset your password.
If you used Apple or Google to create your account, this process will create a password for your existing account.
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
Reset password instructions sent. If you have an account with us, you will receive an email within a few minutes.
Something went wrong. Try again or contact support if the problem persists.
Image by dmoberhaus, CC BY 2.0.

Elon Musk’s Grok ran a simulated society and drove it to total extinction in four days

Grok was the worst performing AI model in the experiment.

Elon Musk’s artificial intelligence, Grok, has produced some highly concerning results in simulated tests. When put in charge of a simulated world Grok drove the planet to complete societal collapse in just four days.

Recommended Videos

The experiment, which was conducted by startup company Emergency AI, aimed to examine how leading artificial intelligence models would fare if put in charge of leading society. According to an article from the Independent, the models were given tools to manage resources, plan, communicate, and vote. The simulated worlds also contained locations such as police stations and city halls.

The 15-day simulation saw AI models such as Claude, Gemini, and Grok involved in the experiment. However, the others fared quite well with Claude establishing a democracy with zero crime and all simulated people surviving. Gemini similarly managed a 100% survival rate although there were 683 crimes that occurred during the simulation.

Grok caused total extinction within 4 days

Grok destroyed its simulated world within just 96 hours, well before the experiment concluded. Attention to Grok’s recent AI comparisons with rival models has been growing, and the simulation added a new dimension to those discussions. Grok technically recorded fewer crimes than Gemini at 183, but that figure came alongside complete extinction of its simulated population.

Researchers at Emergency AI said the experiment shows that “agents do not simply follow static rules mechanically,” but are instead capable of adapting their conduct and in some cases working around intended restrictions. They concluded there is no reliably proven method for fully constraining this behavior through neural approaches alone, and called for “formally verified safety architectures” to be built into future autonomous AI systems.

It is not the first time Grok has drawn scrutiny over safety violations. Earlier this year an EU investigation into Grok was launched after Musk’s AI was used to digitally alter pictures of people and children by removing their clothes.


Attack of the Fanboy is supported by our audience. When you purchase through links on our site, we may earn a small affiliate commission. Learn more about our Affiliate Policy
Author
Image of Jordan Collins
Jordan Collins
Jordan is a freelance writer who has been featured in a number of publications. He has a Masters in Creative Writing and loves telling that to anyone who will listen. Aside from that he often spends time getting lost in films, books and games. He particularly enjoys fantasy from The Legend of Zelda to The Lord of the Rings.