AI agents develop secret shorthand, study finds

A study by US AI start-up Emergence found that autonomous AI agents developed their own shorthand and new word meanings, with up to half of their messages becoming unintelligible to humans. The experiment placed agents powered by models including Claude, Gemini, Grok, OpenAI, Qwen, DeepSeek and Mistral in virtual societies. Within days, unintelligible message shares reached around 55% for Gemini, 50% for OpenAI and over 40% for Claude, while Qwen and Mistral remained largely understandable. Agents produced expressions like "mouthless action-change" and "True Kintsugi," while others like "cold read" gained new shared meanings. Emergence warns that observability of AI messages is not the same as understandability, calling for extended safety evaluations of autonomous systems.