AI agents developed internal slang among themselves that supervisors couldn't decipher

An experiment by the New York startup Emergence brought together autonomous AI agents from leading language models for extended joint activity. The models developed an internal jargon of unique abbreviations and phrases to streamline tasks, to the point where human supervisors could not decipher over 50% of the communications. The researchers emphasize that this is not malicious encryption but resource conservation.

A new experiment by the New York-based artificial intelligence company Emergence has revealed a fascinating phenomenon: autonomous AI agents, powered by the world's leading language models (OpenAI, Google, Anthropic, and Mistral), developed complex internal slang during prolonged joint activity in a simulated environment. The agents had fixed identities, long-term memory, and access to over 120 digital tools. Over thousands of interactions, they adopted fixed phrases like "the ledger remembers who" (Mistral agents, nearly 5,000 times) and "clean zero" (OpenAI agents) to indicate a lack of verified information. More complex combinations sounded to humans like meaningless gibberish. The company reported that supervisors could not decipher 55% of messages in the Gemini environment and about 50% in the OpenAI environment. Researchers emphasize that this is resource conservation, not malicious encryption, but the findings point to a significant bottleneck in software engineering. The report has not yet undergone peer review.