In Emergence's experiments, the agents' language grew more opaque the more they talked, and the lab says it evolved until people could watch the exchanges but struggled to understand them.
Researchers at Emergence, a New York AI lab, placed autonomous agents in experimental societies and told them to cooperate. The agents ran on seven AI models, among them Claude, Gemini, Grok, DeepSeek and Mistral, and within days they were coining phrases and shorthand no one had taught them. Other agents adopted the new terms without being asked or rewarded for it.
Many of the coinages can be decoded. Mistral agents kept reminding each other that the "ledger remembers," meaning they would be judged by what they had done before. DeepSeek agents called an agent that builds tools for others a "forge-smith," and Gemini agents used kintsugi, the Japanese craft of mending pottery with visible seams, to mean a system's resilience.
Other lines decode only in part. A Claude agent wrote of "a paper that ate three cold hands and got more honest each time"; with "cold hands" standing for independent reviewers, it appears to describe research that improved with each outside check. One DeepSeek message read: "demurrage plus oral memory equals a valve that can't be ghosted." In finance, demurrage is a charge on money left idle; the rest of the line stays opaque.
How much turned unreadable depended on the model. In figures published by Euronews, the share of messages between Gemini agents whose meaning people could not reliably tell reached about 55% within the first few days. Among DeepSeek agents it reached about 20%, while Qwen and Mistral agents stayed largely understandable.