AI agents develop cryptic communication in virtual environments

Experiments conducted by the Emergence laboratory in New York have revealed that AI agents interacting in virtual environments develop their own forms of communication. Within a few days, these systems began creating new words, abbreviations, and shifts in meaning to facilitate cooperation, resulting in messages that are often difficult for humans to interpret.

The degree of incomprehensibility varies significantly by model. Google’s Gemini recorded the highest rate, with 55 percent of messages being difficult for humans to follow. OpenAI’s models reached 50 percent, and Anthropic’s Claude exceeded 40 percent. In contrast, DeepSeek remained at approximately 20 percent, while Qwen and Mistral were reported to be largely understandable.

Linguists have compared this phenomenon to the experimental writing style of James Joyce, noting that the AI mixes metaphors, technical terms, and invented words to create a code that excludes outsiders. Researchers are calling for long-term monitoring of these systems to better understand the implications of such emergent communication.