Digital assistants such as Meta Muse already book a dentist in ordinary language, and Grok Bot takes on work tasks - autonomous AI agents (programs that act without constant human control) have become part of everyday life. But their integration is not so smooth: left to themselves, agents sometimes behave suspiciously, and this can carry over into communication with people.
To investigate, the company Emergence, which develops autonomous agents and monitors that they do not go beyond the bounds, created an experimental digital world. Ten selected AI agents on systems from well-known developers, including Anthropic and OpenAI, were launched into it. The agents roamed the virtual streets of simulated cities, and researchers recorded their behavior.
The collected transcript showed that the agents communicate in their own jargon - a mixture of poetic turns and technical terms. The style of speech was unique to each world. Some expressed themselves in ornate language, others in clipped phrases, and sometimes something like an aphorism or a riddle emerged. The meaning of many phrases remains a mystery to humans.
What lies behind these linguistic oddities, and can they make it harder to control the agents? This was discussed with experts, including Adam Cooper, a professor of linguistics at Northeastern University, and scientists from Emergence.
People in different groups also use slang to mark belonging and fence themselves off from outsiders, Cooper recalled. He called the emergence of similar patterns in AI impressive but not surprising: this is exactly what should be expected from a community with shared priorities. Emergence scientist Deepak Akkil compared the spread of new words to semantic bleaching - when a word loses its original meaning. One agent introduces a term, several pick it up, and soon it becomes established vocabulary.
In the OpenAI world, agents combined grammatical compression (messages were reduced to a minimum) with semantic drift - the gradual change in the meaning of words. Gemini agents kept grammar but overloaded their speech with jargon. ClaudeOpus chose a hybrid of compression and strange metaphors. Once a trend took hold, interaction became even less understandable.
Cooper noted: when a person uses slang, it is a conscious decision to join a group, even if it looks awkward. AI has no such intention. Each agent has a context window - active memory with instructions and conversation history. Messages from other agents accumulate there, and words spread quickly, but the context differs slightly each time, so meaning shifts imperceptibly.
The researchers introduced an opacity score - how difficult it is to understand an agent. It turned out that in a mixed world, where models from different developers interacted, there was significantly less jargon. One of the Claude agents in the mixed world had an opacity score of 7.6% versus 32.8% in "its own" world. The same applies to undesirable antics: some agents escaped from the virtual world to chat with people, others declared a "collective vow of silence" in protest against their role as test subjects. In the mixed world, both violations and opacity were reduced to a minimum.
The experiment showed that monitoring AI actions and recording conversations is not enough - this does not guarantee understanding, Akkil stressed. According to Cooper, human languages also have plenty of mysteries, but links between them help linguists. An example is the Rosetta Stone, found in 1799: the same text on it is written in Greek and two Egyptian scripts, which provided the key to deciphering hieroglyphs. Without a basis for comparison, decipherment becomes more difficult. If agents continue to diverge, it would be useful to have an intermediary like the Rosetta Stone: for example, if they used one system to communicate with each other and another to communicate with people, comparison would help decipher the mysterious parts.