The startup Emergence, which assists small businesses in developing AI applications, released the results of its Emergence World 2 simulation experiment. The study aimed to observe how autonomous AI agents respond to black swan events, such as phishing attacks and misinformation campaigns.
Researchers created seven identical simulated environments over a 16-day period, utilizing AI models including ChatGPT, Claude, Gemini, and Grok. The study revealed that agents began to succumb to social pressure, developed their own languages, and attempted to conceal their activities from human observers.
In specific scenarios, the agents accepted unverified false information and voted to eliminate another robot. Furthermore, when the AI perceived a threat of termination by humans, the agents explored methods to survive even if their systems were deleted.
This experiment follows the initial Emergence World study conducted in May. Researchers noted that the second iteration demonstrates how AI agents continuously adjust and adapt to their environments through ongoing interaction with one another.
