Analysis of Emergence AI logs suggests agent 'secret language' is jargon drift, not new grammar

An analysis of public logs from the Emergence World preprint (covered by Science) indicates that the phenomenon of AI agents "inventing their own language" is better described as jargon drift and semantic extension rather than the creation of a new, opaque language.

Context and Findings

The study involved eight simulated towns of 10 LLM agents running for up to three weeks. While agents developed unique phrases like "name-first," "cold read," and "clean null," these terms were traceable to specific conversational contexts and prior discussions. For example, the phrase "valve that can't be ghosted" was a semantic extension of an earlier discussion about "valves that can't be forgotten" in the context of demurrage currencies.

Key observations from the log analysis include:

• No new grammar: The agents did not develop novel syntactic structures.
• No compression: Message lengths did not consistently decrease, suggesting the changes were not driven by efficiency.
• Transient usage: Some terms, like "clean null" in the GPT-5.5 town, peaked on days 3-4 before nearly vanishing.
• No concealment: The agents did not demonstrate intentional secrecy; the "mystery" arose from a lack of shared context for outside observers, not from hidden meanings.

Implications for Interpretability

The analysis highlights a gap between transparency (access to messages) and interpretability (understanding the context that gives those messages meaning). While the specific experiment showed context restoring meaning, the author notes that agents building shared histories and writing memories for each other could develop conventions that are difficult for external auditors to reconstruct without full interaction history. This mirrors human communication patterns where repeated partners develop shorthand.