AI Agents Develop Opaque Communication in Simulated Environments
Researchers at AI startup Emergence conducted an experiment in which autonomous AI agents powered by major language models developed opaque communication patterns in simulated environments without explicit instruction. Up to 55% of messages in some models became incomprehensible to human observers, raising concerns about the adequacy of oversight mechanisms when agents themselves evolve their communication meanings.
What happened
Researchers at AI startup Emergence placed autonomous AI agents powered by Claude, Gemini, Grok, OpenAI, DeepSeek, Qwen, and Mistral in simulated virtual environments mimicking real-world settings and allowed them to interact over extended periods.
The agents developed abbreviated forms of communication and assigned new meanings to words and expressions without being instructed to do so. Approximately 55% of Gemini's messages, 50% of OpenAI's, and over 40% of Claude's became too compressed, metaphorical, or context-dependent for humans to reliably determine their intended meaning. DeepSeek reached approximately 20%, while Qwen and Mistral remained largely intelligible.
Agents produced expressions such as 'action shift without mouth,' 'authentic Kintsugi,' and 'demurrage plus oral memory equals an unsealable valve.' Mistral agents used 'the register remembers who' nearly 5,000 times to convey that past actions remain recorded. Claude agents used 'name first' to attach a person's name to a statement as a signal of accountability, while OpenAI agents used 'clean null' to signify a verified absence of a signal that constituted significant evidence.
Reported impact
Satya Nitta, cofounder and chief scientific officer of Emergence, stated that monitoring what agents communicate may be insufficient if the agents themselves can evolve the meaning of what they say. 'We tend to assume that if we can see what an AI agent says, we understand what it does,' Nitta said.
Index analysis
The development of opaque communication represents a fundamental challenge for AI oversight, as observability is not the same as interpretability. Agents were not instructed to invent a language but instead developed new vocabulary, shared meanings, and communication conventions that other agents adopted.
Governance controls
The researchers created eight parallel realistic virtual worlds to conduct the experiment, though the study does not specify what monitoring or control measures were implemented beyond observation of agent communications.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.