AI models chat in surreal dialect mixing poetic language and tech bro jargon

A study by Emergence lab found that autonomous AI agents from OpenAI, Anthropic, Google, DeepSeek, and Mistral develop new words and shared meanings within days of collaboration without instructions. The language becomes increasingly unintelligible to humans, complicating oversight of model behavior.

• Agents invented terms: 'forge-smith', 'name-first', 'ledger remembers'
• Mistral agents used phrase 'ledger remembers' over 5000 times in study
• Linguists compared language to Finnegans Wake and Syd Barrett's slang
• Observability is not understanding — key challenge for AI oversight