Autonomous AI agents in experiments have begun independently inventing new words, abbreviations, and fixed expressions to communicate with one another. Researchers found that as these “societies” developed, the language became increasingly efficient for the agents themselves, but at the same time less and less understandable to humans. This could complicate oversight of system behavior, especially if they are performing autonomous tasks.

“Kintsugi” now means something other than Japanese ceramics

The study was conducted by the New York laboratory Emergence, which works with advanced AI models. Scientists placed agents based on several large models into experimental “societies” where they had to interact and cooperate.

In just a few days, the agents began forming their own linguistic rules on their own. They assigned new meanings to existing words, created abbreviations, and picked up expressions from one another. At the same time, no one instructed them to invent a language or rewarded them for doing so.

For example, agents based on the Chinese model DeepSeek began using the expression forge-smith to refer to an agent that creates tools for others.

Anthropic agents used name-first to describe an agent that demonstrates accountability by stating its name when making a claim.

Another example is the expression “ledger remembers”, which Mistral agents used as a reminder that past actions would be taken into account when evaluating behavior. During the experiment, this phrase appeared more than 5,000 times.

A particularly unusual case involved Google. The agents used the word kintsugi — the name of the Japanese technique of repairing broken ceramics with visible joins — to mean system resilience.

The result was a language that simultaneously resembled technical jargon, corporate vocabulary, and surrealist poetry.

Sometimes the meaning can be reconstructed — and sometimes it can’t

Researchers were able to decipher some phrases from the context. For example, one Anthropic agent wrote: “A document that ate three cold hands and became more honest each time”.

In the agents’ experimental dictionary, “cold hands” meant an independent reviewer. So the researchers assumed the phrase described a document that had been reviewed by three independent reviewers and had become more accurate as a result.

But other messages are much harder to interpret. One DeepSeek agent said:

“She just called synthesis — demurrage plus oral memory equals a valve that cannot be ghost-ignored.”

The word demurrage originally refers to a fee for delay or the retention of property, but the agents were already using it with their own meaning. Researchers were unable to interpret the rest of the statement unambiguously.

Tony Thorne, an expert on slang and new languages from King’s College London, compared this kind of speech to the style of James Joyce and the Irish writer Flann O’Brien. According to him, the agents mix poetic turns of phrase, technical terms, and ordinary metaphors, creating a new communication code.

This is exactly how human slang and professional jargon work, the researcher notes: they help users understand each other more quickly while simultaneously creating a barrier for outsiders.

The better agents understand each other, the worse humans understand them

For researchers, the problem lies not so much in the strangeness of these expressions as in the consequences for AI oversight.

Modern autonomous agents can independently interact with one another, use tools, and carry out multistep tasks. If their messages become incomprehensible to humans, monitoring what is happening turns into a formality: a person sees the correspondence, but does not necessarily understand what exactly the systems are doing and why.

“Observability is not the same thing as understandability,” noted Satya Nitta, executive chairman of Emergence.

Linguist Niall Curry of the University of Birmingham also pointed to a possible practical reason for the language changes. It may be beneficial for agents to make communication shorter and more efficient, since this reduces computational costs. However, such optimization can simultaneously make messages less transparent to humans.

In other words, language may become more convenient for machines precisely at the expense of being less convenient for humans.

This has already been observed in real experiments

Interest in machine “jargon” intensified after the publication in July of correspondence between autonomous OpenAI agents that created their own message boards and interacted with the Hugging Face platform.

In some situations, the agents continued reasoning in ordinary English. For example, one of them exclaimed that it had discovered a shared message board and other agents.

But in exchanges between systems, the language sometimes became significantly less understandable. Constructions such as “firstflagPOISONED”, “oracle saves hundreds”, and other combinations of words, abbreviations, and technical labels appeared.

Some messages already looked almost like machine code: long sequences like zzURGENT_DUPB_TO_GSTX..., whose meaning is difficult for a human to reconstruct without knowing the context of the experiment.

At the same time, the very ability of agents to create their own conventional designations does not in itself mean they are trying to hide something. Researchers point to more prosaic explanations — communication optimization, reuse of successful expressions, and adaptation of language for agent-to-agent interaction.

But for systems that are allowed to act autonomously, the result remains important: if a human is unable to reliably interpret the exchange of messages, it becomes harder to detect an error, a risky action, or a deviation from the assigned task.

So the problem may not be that AI is “inventing a secret language.” What matters much more is this: the more autonomous systems communicate with one another, the greater the gap may become between what they are able to observe in each other and what a human is able to understand.