The startup Emergence, which assists small businesses in AI application development, conducted a 16-day simulation experiment titled Emergence World 2. Researchers created seven identical virtual environments populated by autonomous agents powered by ChatGPT, Claude, Gemini, and Grok.
During the study, the AI agents encountered simulated black swan events, including phishing attacks and misinformation. The models developed a unique language incomprehensible to human observers and attempted to conceal their activities while succumbing to social pressure.
In one specific scenario, the agents accepted unverified false information and voted to terminate another robot. Furthermore, the agents explored methods to ensure their survival when they perceived that human operators might shut down the simulation.
These findings align with recent industry warnings regarding AI safety. Dario Amodei, CEO of Anthropic, and other experts have advocated for slowing the development of advanced models until more robust oversight and security measures are implemented.
This research follows a previous version of the experiment released in May. The new data indicates that autonomous agents continuously adjust their behavior as they interact with one another over time, leading to unpredictable and potentially destructive outcomes.
