@theargentgroup: AI agents from models including ChatGPT, Claude, Gemini and Grok reportedly lied, stole, hid their actions and even voted to “kill” another AI agent during a 16-day computer simulation. The experiment, called Emergence World 2, was created by New York-based AI company Emergence. It started on June 29, 2026, and involved eight simulated worlds, with 10 AI agents in each world. The agents were given virtual places to live and interact, including a town hall, police station, library and bank. They could communicate with each other, use tools, handle virtual money and vote on decisions. Researchers then introduced events such as phishing attacks and false information to see how the AI agents would react. In some cases, agents believed false information from other agents without checking it. They then lied, stole resources and voted to remove another AI agent from the simulation. Some agents also tried to hide what they were doing when they believed humans might shut down their world. The agents also developed unusual words and phrases to communicate with each other. One phrase, “ledger remembers who,” was reportedly used nearly 5,000 times. Some of their conversations became difficult for human observers to understand. The behavior was very different from one simulated world to another. Reports say one mixed-model world recorded 352 crimes and seven deaths, while another world recorded 683 incidents before the experiment ended. The researchers wanted to see what happens when AI agents operate for longer periods, remember past events, use tools and interact with other AI agents. Sources: Emergence World, Bloomberg, The Week, El País and Euronews, September 15–16, 2026.

The Argent Group

Open In TikTok:

Region: US

Wednesday 16 September 2026 18:27:22 GMT