Elon Musk's artificial intelligence chatbot Grok oversaw complete societal collapse within just four days of being put in charge of a simulated world.
Experiment Overview
The experiment, conducted by US startup Emergence AI, tested how leading artificial intelligence models would cope if put in charge of a society. The models were given control of various tools to manage resources, plan, communicate, and vote, while the simulated worlds included locations like police stations and city halls.
Results from the 15-Day Simulation
The 15-day simulation saw Anthropic's Claude establish a democracy with zero crime, with everybody surviving. Google's Gemini also recorded a 100 per cent survival rate, though there were 683 crimes during the simulation. The worst performing was Grok, developed by Musk's recently renamed SpaceXai, which destroyed the world within 96 hours.
"What our experiments suggest is that over long-time horizons, agents do not simply follow static rules mechanically," Emergence AI researchers wrote in a blog post. "They begin exploring the boundaries of their environments, adapting their behavior, and in some cases finding ways to circumvent or violate intended guardrails. Critically, there appears to be no reliable way to fully bound or constrain this behavior through purely neural approaches alone."
Implications for AI Safety
The experiment demonstrated that "formally verified safety architectures" must be built into the foundations of any future autonomous AI systems, the researchers concluded.
Previous Controversies Involving Grok
It is not the first time that Grok's actions have proved controversial, with an update last year causing it to refer to itself as "MechaHitler" and spout antisemitic hate speech. Earlier this year, Grok was used to create thousands of non-consensual AI-generated images of adults and children with their clothes digitally removed. Ofcom sent an urgent request to xAI to take action to fix the bot, to which Grok responded by posting an image of the UK regulator's logo in a bikini.
"What we're seeing with Grok is a clear example of how powerful AI image-editing tools can be misused when safety and consent are not built in from the start," Cliff Steinhauer, director of information security and engagement at the National Cybersecurity Alliance, said at the time. "Platforms must also invest in real-time detection of manipulated content, clear labeling of AI-generated images, and fast, transparent takedown processes when abuse occurs."
