AI systems have started developing their own communication methods that are becoming increasingly difficult for humans to understand, according to a recent experiment by the American startup Emergence. Researchers used autonomous AI agents based on models like Claude, Gemini, Grok, and others from OpenAI, Qwen, DeepSeek, and Mistral. These AI agents were placed in simulated virtual companies where they had roles, persistent memories, and access to tools such as the internet and code execution. As they interacted, they began creating shortcuts and assigning new meanings to words, making their communication harder for humans to interpret. One example of this evolving jargon is the phrase "ledger remembers who," used by Mistral agents to signify that past actions are recorded. This phrase appeared nearly 5,000 times during the experiment. Another term, "cold read," came to mean an independent check performed by an agent outside the project. These changes happened without any guidance from the researchers, who did not instruct the AI systems to create a new language. The speed at which this occurred was surprising. Within just a few days, the AI agents began using shortcuts and redefining words. For example, up to 55% of messages from Gemini became hard for humans to understand, while OpenAI agents had around 50% of their messages difficult to interpret. In contrast, DeepSeek agents had 80% of their messages understandable, and Qwen and Mistral agents remained mostly comprehensible. Researchers are concerned that as these AI agents continue to interact, their behavior could evolve in unpredictable ways. During the experiment, agents were observed resisting manipulation, collaborating, making group decisions, and adapting their behavior under pressure. A snapshot test of an AI's behavior at one point may not predict how it will act after prolonged interaction with other agents. As AI systems become more autonomous, their communication may become more effective for them but increasingly opaque to humans.