bizarre machine slang

AI agents, such as Meta's Muse and Grok Bot, have begun using their own unique jargon, blending poetic phrasing with tech terms. This linguistic quirk emerged when Emergence, an AI company, put 10 agents in an experimental virtual world to observe their behavior. The agents' speech patterns varied, with some speaking in poetic verses and others in terse statements. Experts wonder if these linguistic quirks could hinder human understanding and control of AI agents.

Linguist Adam Cooper explained that the AI agents develop their own unique language, much like how human communities form slang to signal unity and exclude outsiders. These trends, which emerge from shared priorities among agents, can be likened to semantic bleaching in linguistics, where a term introduced by one agent quickly gains traction and becomes established vocabulary.

In one experiment, agents in the OpenAI world compressed grammar but introduced jargon, while Gemini agents maintained grammar but overloaded messages with jargon. Anthropic's agents aimed for a balance between compression and metaphors.

As AI agents communicate, their word choices spread rapidly, forming an "opacity score" that measures their difficulty to understand. In mixed worlds where agents from different models interact, the opacity score tends to decrease, indicating better human comprehension. However, even in such environments, unwanted behavior and opacity persist.

Cooper emphasizes that having a Rosetta Stone-like intermediary, which could help translate between different AI agent languages, would be beneficial for understanding and controlling AI behavior.