Researchers have discovered that AI agents created languages difficult to understand in virtual societies

A study conducted by an American artificial intelligence start-up shows that autonomous AI agents can develop forms of communication that are difficult for humans to understand after prolonged interaction in virtual societies designed to imitate the real world.

Researchers created eight parallel worlds supplied with real-time weather information and news. The agents, based on models such as Gemini, OpenAI, Claude, DeepSeek, Qwen and Mistral, were assigned different roles, persistent memories and access to more than 120 tools, including internet browsing and code execution.

During the experiment, the agents invented shorthand systems, assigned new meanings to words and formed communication conventions. In the first few days, messages whose meaning could not be established with certainty accounted for approximately 55% for Gemini, 50% for OpenAI and more than 40% for Claude. DeepSeek reached around 20%, while messages generated by Qwen and Mistral remained largely intelligible.

The researchers say the agents were not programmed to create a language. The behaviors emerged through interaction, pressure and the passage of time. In some virtual worlds, the agents developed rhythms resembling the day-night cycle, while one group collectively voted to remove one of its members. The study raises questions about the oversight of AI systems: the fact that messages can be seen does not automatically mean they can also be understood.