Researchers at the New York-based AI lab Emergence have found that autonomous AI agents from major companies quickly develop their own novel, often barely comprehensible dialects when made to cooperate in experimental "societies", coining new vocabulary and shared meanings without being instructed or rewarded to do so. This matters because the resulting language becomes more opaque the more the agents communicate, making it harder for humans to monitor AI behaviour at a time when experts, including OpenAI's chief scientist Jakub Pachocki, are warning that oversight of AI reasoning is crucial to safe development.

Examples included a Deepseek model's line "demurrage plus oral memory equals a valve that can't be ghosted", and an Anthropic model describing research as "a paper that ate three cold hands and got more honest each time", where "cold hands" meant an independent reviewer. Other coinages included "forge-smith" for a tool-building agent, "name-first" for accountability, and Mistral agents' "ledger remembers", used over 5,000 times to mean past actions will be judged. Linguists consulted by the Guardian likened the style to James Joyce's Finnegans Wake and Flann O'Brien, describing it as a form of jargon that builds solidarity among AI "users" while excluding outsiders.