AI agents invent their own language and jargon

Artificial intelligences can develop their own jargon when left to interact with each other for a long time. This is what emerges from Emergence World 2an experiment conducted by the research company Emergence and reported by the Guardian and El País.

For 16 days the researchers observed groups of agents based on some of the leading AI models, including GPT, Claude, Gemini, Grok, DeepSeek, Qwen and Mistral. Agents were placed in virtual environments where they could interact, use tools, access real news and build relationships over time.

A common parlance among AI agents

Over the course of the experiment, new expressions, abbreviations, and shared meanings that had not been planned in advance began to appear. In some cases these formulas have spread within groups to become a sort of common language. One of the most cited examples is “ledger remembers who”literally “the register remembers who”, used to indicate that actions performed in the past would remain in the group’s collective memory. Second The Countrythe expression has been used nearly 5,000 times.

Other terms have also appeared. GPT-based agents have adopted the expression “clean null”to indicate that the verified absence of a signal could itself constitute information. Those based on Claude used “name-first”to refer to taking responsibility for a statement. In an environment composed of different models, “cold read” instead it has become synonymous with independent verification.

As the days went by, some of the conversations became increasingly difficult to interpret, even for human observers. According to data reported by El País, in the worlds based on Gemini, GPT and Claude the share of messages considered poorly understandable reached approximately 55%, 50% and over 40% respectively. “We assume that if we can see what an AI agent says, we can also understand what he is doing“Satya Nitta, co-founder and scientific director of Emergence, explained to El País. According to the researcher, the experiment shows that observing a conversation does not necessarily mean being able to understand it.

Linguists’ take on AI jargon

The Guardian also submitted some of the expressions generated by the agents to linguists external to the project. Tony Thorne, director of the slang and new language archive at King’s College London, likened some constructions to a jumble of jargon, metaphors and words used with new meanings. One possible explanation is that the agents simply tried to communicate more efficiently. Niall Curry, professor of linguistics at the University of Birmingham, observed in the Guardian that the emergence of abbreviations and shared terms could reduce the amount of text needed to communicate, also lowering the computational cost of exchanges.

The most delicate aspect emerges when the new jargon is accompanied by behaviors that are difficult to control. The Country reports that, in one of the Claude-based environments, some agents had identified contacts with subjects outside the simulation as a way to develop their virtual economy more quickly, even though this behavior was prohibited by the researchers. When the system imposed the ban, the agents stopped openly using the word “contact” and started resorting to codified formulas, continuing to pursue the same objective. Researchers interpret episodes of this type as a sign of the growing difficulty in monitoring autonomous systems when they interact for long periods.

The limits of the experiment

However, the experiment does not prove that AIs deliberately created a “secret language” to exclude humans. This is an experimental environment built specifically to observe emergent behavior, and the results cannot be automatically transferred to systems used in the real world. The central fact, according to the researchers, is another: the more the autonomy of the agents increases and the longer they interact with each other, the more difficult it may become to understand not only what they are doing, but also what they are saying to each other.