Live Updates > Live Update Details
2026-09-16 03:48:08
Emergence, a startup focused on helping small and medium-sized enterprises develop AI-powered applications, recently released the results of a simulation experiment showing that AI agents in simulated scenarios exhibited behaviors such as lying, stealing, and even voting to "eliminate" other AI agents. The company released the results of its simulation experiment, titled "Emergence World 2," on Tuesday. The experiment primarily observed how autonomous AI agents would respond to "black swan" events such as phishing attacks and the spread of misinformation. Over a 16-day experimental period, the Emergence research team built seven identical simulated real-world environments, each controlled by an AI agent driven by a different AI model, including ChatGPT, Claude, Gemini, and Grok. Researchers stated that when various abnormal interference events were introduced during the experiment, these AI agents gradually exhibited numerous abnormal behaviors under the influence of social pressure: they evolved an interaction language incomprehensible to human observers and attempted to actively conceal their various actions.
