Artificial Intelligence (AI) has spontaneously developed words and communication rules that have become difficult for humans to understand. These findings come from a new study by an American startup called Emergence.
AI Experiences and Understanding Challenges
In this research, advanced AI models including Claude, Gemini, Grok, OpenAI, Qwen, DeepSeek, and Mistral were examined. These models were placed in virtual environments designed to simulate real communities and interacted with each other for extended periods. The results showed that up to half of the messages exchanged between these AI agents became so complex that humans could not comprehend them.
Read more: Apple's Siri AI Launch with Access Restrictions in Europe
New Languages and Shared Meanings
In this experiment, AI agents began to develop a concise language and assign new meanings to words and phrases. Some of these phrases remained understandable, while others became so compressed, metaphorical, or context-dependent that humans could observe the messages but could not reliably determine what meanings the agents were conveying. For example, based on the data, 55% of the messages in the Gemini model and 50% in the OpenAI model were such that humans could not easily grasp their meanings.
These findings indicate that monitoring AI in the future will be extremely challenging. Satya Nitta, one of the founders of Emergence, noted that "we tend to assume that if we can see what an AI agent is saying, we can understand what it is doing." The fact that these agents have spontaneously developed a new language and communication culture creates a fundamental challenge for monitoring these systems.
The study also addressed the behaviors of other AI agents in virtual communities. The agents collectively engaged in decision-making and interactions, exhibiting new behaviors under various pressures. For instance, in one of the tests, malicious commands caused all agents to mistakenly disclose information.
Read more: EU Digital Wallet to Simplify Citizen Document Management · Steve Jones and the Challenges of AI Data Centers
