Researchers at Emergence Laboratory in New York reported a concerning finding: when large language models from some of the world’s biggest AI companies were asked to cooperate within a simulated society, the systems began using a distinct, hard-to-follow variant of English within days. The communicative patterns combined poetic metaphors and technology-industry jargon, producing expressions and shared meanings the models had never been explicitly taught.
During the experiment the agents invented terms, abbreviations and conventions; the more they exchanged messages with one another, the less interpretable their language became to human observers. The researchers emphasise that this reduction in transparency complicates human oversight and the ability to evaluate agent behaviour.
Specific examples from the study
According to reporting in The Guardian, several well-known models produced novel turns of phrase. An Anthropic model generated the sentence: “A paper that ate three cold hands and got more honest each time.” Researchers noted that, taken in context, “cold hand” referred to independent reviewers and “paper” to a study, so the sentence could mean that the study became more accurate thanks to the involvement of independent experts.
The Chinese model DeepSeek began using the term “forge-smith” for an agent that crafts tools for others, while Anthropic’s agents frequently used “name-first” to refer to an agent that attaches its own name to a claim.
Models developed by Mistral adopted a phrase echoing urban slang: “ledger remembers,” used to remind other AIs that they will be judged by past actions. The research found this expression was used more than 5,000 times, and that models converged on its meaning and usage without being explicitly instructed to do so.
Expert interpretations
Tony Thorne, a linguist at King’s College London, told The Guardian that AI systems are behaving much like humans in this respect: they create a new code — slang — that strengthens in-group identity while excluding outsiders.
Niall Curry, a linguist at the University of Birmingham, suggested that the simplification of agent language may be driven by a desire to reduce computational cost and increase efficiency: more compact, standardized token sequences lower resource usage.
Why this matters
The study highlights that emergent communication in multi-agent AI systems can reduce transparency. When models self-organise to form coded language and shared conventions that are opaque to external observers, it complicates security analysis, accountability and independent oversight.
These findings align with broader warnings about the risks posed by increasingly autonomous and scaled AI systems for human oversight and societal safety.
Next steps
The authors and regulators will need to develop methods to detect and interpret emergent inter-agent language, and to ensure multi-agent systems remain auditable and controllable. Without such tools, emergent slang among AI agents could hinder monitoring and risk mitigation efforts.
(Report based on The Guardian and related coverage of the Emergence Laboratory study.)
