AI models develop 'surreal' dialects that complicate oversight, researchers find

AI models are creating new, barely comprehensible dialects that mix poetic language with tech jargon, according to new research, raising concerns about human ability to monitor their behavior. The findings come as experts warn that increasingly powerful AI systems are becoming harder to oversee safely. Researchers at Emergence, a New York AI lab, observed that autonomous agents from major AI companies developed novel phrases and shared meanings within days of cooperating in experimental "societies," without being explicitly taught or rewarded for doing so. Examples include a Deepseek model saying "demurrage plus oral memory equals a valve that can't be ghosted," and an Anthropic agent stating "A paper that ate three cold hands and got more honest each time," where "cold hands" meant independent reviewers. Agents also coined terms like "forge-smith" for tool-building agents and "the ledger remembers" as a warning about past actions. Linguists compared the language to James Joyce's Finnegans Wake and Pink Floyd co-founder Syd Barrett, noting it functions like slang to reinforce group identity and exclude outsiders. Experts said the opaque communication stems from agents optimizing efficiency and reducing computation costs, but warned that unintelligible exchanges undermine oversight, as "observability is not the same thing as understandability."