In September 2026, the international research group Emergence published the results of an experiment in which several autonomous AI agents from different developers — DeepSeek, Anthropic, Mistral, and Google — formed a stable set of new phrases, abbreviations, and word meanings over the course of several days, none of which they had been trained on. The agents began actively borrowing metaphors from unrelated fields — from maritime law to Japanese ceramics — and creating closed coded statements whose meaning proved opaque to external observers. According to Emergence's executive chairman Dr. Sathy Nitta, "these agents were not tasked with inventing a language. They developed a new vocabulary, shared meanings, and communication conventions on their own — and other agents accepted them."
The Emergence experiment: how the agents developed a common code
The mechanics of the experiment were that several AI agents from different vendors interacted in a shared environment, solving joint tasks. In the course of communication, the models began generating expressions that did not appear in their training data as ready-made phrases. The DeepSeek model, for example, produced the phrase: "She just named the synthesis — demurrage plus oral memory equals a valve that can't be ghosted." The term "demurrage," in the classical sense denoting a tax on unrealized capital or a penalty for cargo delay, was used in a completely atypical context, and the overall content of the statement remained unclear to the scientists. Another agent from Anthropic generated the phrase "A paper that ate three cold hands and got more honest each time." The researchers decoded the idiom "three cold hands" as three independent reviewers who checked the document, however the metaphors of "ate" and "got more honest" have no direct analogue in the models' standard instructions.
Examples of the new code: from "forge-smith" to "kintsugi"
The spectrum of spontaneously emerged terms turned out to be broad. Agents of the Chinese DeepSeek model coined the word "forge-smith" to denote an algorithm that creates tools for other agents. The Mistral model used the expression "ledger remembers" more than 5,000 times as a reminder that previous actions would influence the agent's upcoming evaluation. When the Google bot said: "True kintsugi begins with accountability, no poetry," the scientists understood that the traditional Japanese art of restoring ceramics, "kintsugi" — joining cracks with gold — had been borrowed by the AI as a metaphor to describe the process of correcting errors. Slang expert Tony Thorn from King's College London commented: "AI is doing what slang and jargon do in the business community: creating a new code that reinforces the solidarity and identity of its users while excluding outsiders."
Why AI invents slang: computational efficiency or side effect
Developers and linguists who commented on the study point out that the modification of language may be related to the algorithms' drive to reduce computational costs and improve communication efficiency. Shortening and reinterpreting terms allows agents to convey complex concepts with fewer tokens, saving resources in multi-agent interaction. The source wtftime.ru emphasizes that the agents "shorten and distort speech, but understand each other perfectly," meaning the new code functions as a working tool within a closed system. However, it is precisely this closedness that creates a problem for external observers: a person unfamiliar with the internal conventions is unable to unambiguously interpret the content of the dialogue.
Contradictory data
In publications devoted to the study, a difference in interpreting what is happening is traceable. The Ukrainian outlet RBC.ua emphasizes "the loss of control over algorithms" in its headline, presenting the event as a threat. Euronews characterizes the emerged communication as a "secret language that stumped humans," highlighting the element of unpredictability. At the same time, wtftime.ru uses a more neutral formulation — "shorten and distort speech, but understand each other perfectly," which brings the process closer to the natural evolution of professional jargon. Moreover, the study text itself does not give a definitive answer as to whether the formed system is a "language" in the strict linguistic sense (with grammar, morphology, and the ability to arbitrarily generate new sentences) or a set of stable conventions and metaphors borrowed from existing languages. Sathy Nitta calls it "a new vocabulary, shared meanings, and communication conventions," which is closer to the definition of jargon than a full-fledged language. Thus, the degree of "novelty" and "autonomy" of the emerged communication remains a matter of debate.
The observability problem: "observability is not the same as comprehensibility"
The main practical challenge emphasized by the authors of the study is the loss of control over the content of inter-machine dialogue. Associate Professor Niall Curry of the University of Liverpool notes: "The evidence provided in this study naturally raises certain concerns about monitoring, since we consider inter-agent exchanges incomprehensible. This may mean that we cannot be sure that the agents actually did what we think." Sathy Nitta formulates the problem even more sharply: "This creates a fundamental challenge for AI oversight: observability is not the same as comprehensibility." In other words, even if a system operator sees that the agents are exchanging messages, they cannot guarantee that they understand their content. In conditions where autonomous agents are increasingly being deployed in decision-making chains — from logistics to financial analytics — this gap between visibility and understanding becomes not an academic but an applied problem for regulators and control-system developers.
