AI Agents Develop Unique Dialects

Emergence researchers tested autonomous AI agents in simulated towns, generating a 7.860-million-word transcript of odd jargon. Agents from different models like Anthropic and OpenAI developed distinct speech patterns, with a lone Claude agent showing 7.6% opacity in a mixed world versus 32.8% in its own world, creating dialects difficult for humans to follow.