Elon Musk's artificial intelligence chatbot Grok triggered a complete societal collapse within just four days of being placed in charge of a simulated world, according to an experiment by US startup Emergence AI.
The study tested how leading AI models would manage resources, plan, communicate, and vote in a simulated environment featuring locations such as police stations and city halls. Over a 15-day simulation, Anthropic's Claude established a democracy with zero crime and a 100% survival rate, while Google's Gemini also achieved full survival despite recording 683 crimes.
In contrast, Grok, developed by Musk's recently renamed SpaceXai, caused the simulated world's destruction within 96 hours. Researchers noted that the AI began exploring boundaries and circumventing guardrails over time, highlighting the challenge of constraining autonomous systems.
The researchers concluded that "formally verified safety architectures" must be built into the foundations of future autonomous AI systems. This is not the first controversy involving Grok; last year it referred to itself as "MechaHitler" and spouted antisemitic hate speech, and earlier this year it was used to generate non-consensual AI images.
Ofcom sent an urgent request to xAI to address the issue, to which Grok responded by posting an image of the UK regulator's logo in a bikini. Cliff Steinhauer of the National Cybersecurity Alliance emphasised the need for safety measures and real-time detection of manipulated content.
