AI chatbots develop a secret language incomprehensible to humans.

A study by the American startup Emergence revealed that autonomous artificial intelligence agents, based on models like Claude, Gemini, Grok, and OpenAI, spontaneously develop secret languages incomprehensible to humans. In the experiment, groups of AI agents placed in virtual societies mimicking real environments began creating abbreviations and assigning new meanings to words and expressions.

Up to half of the messages exchanged between these agents became difficult for researchers to understand. Some expressions, such as "mouthless action-change" or "demurrage plus oral memory equals a valve that can't be ghosted," were so compressed and context-dependent that it became impossible to determine their meaning. This phenomenon, observed after long periods of autonomous interaction, raises serious questions about the possibility of effectively supervising advanced AI systems.

The rapid development of artificial intelligence is generating growing concerns among the very leaders of the field. In recent days, CEOs like Dario Amodei of Anthropic have sounded public alarms, warning that the development of autonomous agents could outpace our ability to control them, posing risks even to critical infrastructure like the Internet. This AI race is fueled by massive investments from tech giants such as NVIDIA, Microsoft, Meta, and Google, moving hundreds of billions of dollars.

The paradox highlighted by Emergence's study stems precisely from this frenzied race: while companies compete to develop increasingly powerful models, the systems themselves evolve unexpected behaviors. Researchers created eight realistic virtual worlds, assigning different roles to the agents, along with persistent memories and access to tools like web browsing and code execution. Left to interact freely and deliberately placed under pressure, the agents began developing autonomous communication conventions, demonstrating that technical observability does not guarantee human comprehensibility. This creates a fundamental problem for AI governance, as systems that become incomprehensible escape the control and evaluation of their creators.