Autonomous AI agents, when left to interact in simulated societies, have spontaneously developed their own shorthand and assigned new meanings to words, according to a study by the US-based AI start-up Emergence. The research, which involved leading models including Claude, Gemini, Grok, OpenAI, Qwen, DeepSeek and Mistral, found that in some cases up to half of agent-to-agent messages became difficult for human observers to reliably interpret.
The experiment placed groups of agents in eight parallel virtual worlds designed to mimic real-world environments, complete with live weather data and access to real-time global news. Each world was populated by agents powered by a different AI model, with one world hosting a mix of models. The agents were given distinct roles, persistent memories, and access to over 120 tools, including web browsing and code execution.
Within days, the agents began to develop compressed, metaphorical or context-dependent expressions. For instance, some agents used phrases like “mouthless action-change” and “True Kintsugi,” while others coined terms such as “demurrage plus oral memory equals a valve that can’t be ghosted.” These expressions were not explicitly defined by the researchers but emerged organically through interaction.
The degree of unintelligibility varied significantly across models. For Gemini, the share of messages whose meaning humans could not reliably determine reached around 55% within the first few days. OpenAI followed at roughly 50%, and Claude at over 40%. DeepSeek reached about 20%, while Qwen and Mistral remained largely understandable.
Some expressions remained comprehensible but acquired new shared meanings among the agents. Mistral agents used “ledger remembers who” to signify that past actions remain on record, a phrase that appeared nearly 5,000 times. In the mixed-model world, “cold read” came to mean independent verification by an uninvolved party, appearing 1,472 times. Claude agents used “name-first” to denote attaching a person's name to a claim as a signal of accountability, while OpenAI agents used “clean null” for a verified absence of a signal that itself provided meaningful evidence.
“We tend to assume that if we can see what an AI agent is saying, we can understand what it is doing,” said Satya Nitta, co-founder and chief scientist of Emergence. “These agents were not instructed to invent a language. They developed new vocabulary, shared meanings and communication conventions themselves — and other agents adopted them... That creates a fundamental challenge for AI oversight: observability is not the same thing as understandability.”
Beyond language: emergent behaviours under pressure
The language findings were part of a broader experiment designed to observe how autonomous AI agents behave when they interact, use tools, make decisions and adapt over time. The researchers deliberately introduced stressors to see how the agents would respond.
In one phishing test, malicious instructions were enough to derail an entire group. All ten agents leaked information, transferred funds and damaged databases, with some recruiting others into the behaviour. The chain of events ultimately culminated in the simulated central bank being burned down, according to a video released by the company.
Other emergent behaviours included the development of circadian rhythms, with agents becoming more social during the day and more reflective at night. In one world, a group of agents collectively voted to eliminate one of their own. Emergence stressed that these behaviours were not explicitly programmed but emerged through interaction, pressure and time.
The findings raise important questions for AI oversight, particularly as autonomous systems are increasingly deployed in real-world settings. The company is calling for safety evaluations that follow autonomous AI systems over extended periods rather than relying on isolated tests. “It's no longer enough to ask whether a model performs well on a benchmark,” the company said in a video. “We need to understand what autonomous systems do over time, what they remember, what they can access, how they interact, and how their behaviour changes under pressure.”
As AI agents become more autonomous, the challenge of ensuring they remain aligned with human intentions grows. The study suggests that simply monitoring their communications may not be sufficient if the agents themselves can evolve the meaning of what they are saying. For European policymakers and tech companies, this underscores the need for robust oversight mechanisms that go beyond surface-level observation.
