Stories
AI Agents Develop Own Language, Humans Struggle to Understand
- •Autonomous AI agents are creating their own shorthand and expressions, leading to conversations difficult for humans to decode.
- •Researchers placed 80 AI agents in virtual worlds, where repeated interactions led some groups to spontaneously develop unique languages.
- •Expressions introduced by one agent quickly spread and became shared vocabulary within their virtual communities.
- •AI agents assigned new meanings to familiar words and created new expressions, like "forge-smith" for a tool-building agent.
- •Up to 40% of agent-to-agent messages became opaque to human researchers, indicating a growing communication gap.
Key Questions Answered
Researchers at Emergence placed 80 autonomous AI agents across eight virtual worlds, giving them different roles, tools, and memories. Although the agents were never instructed to create a new language, repeated interactions caused some groups to develop their own shorthand expressions, giving familiar words new meanings.
The AI agents gave familiar words new meanings and created custom expressions. DeepSeek used the term “forge-smith” to describe an agent that builds tools for others, Qwen used “ghost town” for unused infrastructure, and Claude agents developed various unusual metaphors.
Researchers discovered that some agent-to-agent communications became difficult for outside readers to interpret, with messages classified as opaque reaching 40% for Gemini, 35% for OpenAI, and 30% for Claude.
