According to a new study, autonomous artificial intelligence agents have created a new English dialect consisting of an unusual mix of literary works and corporate jargon.
“Elchi” reports that researchers at the New York-based AI lab “Emergence” have found that models from different companies tasked with collaborating within experimental “societies” have begun to create expressions, abbreviations, and shared meanings that they were never previously taught.
The language used by these models, which adopts both poetic metaphors and heavy, complex business terminology, has become increasingly complex and incomprehensible in the process of mutual communication.
Experts emphasize that this situation poses serious risks in terms of security control and note that as artificial intelligence systems become more powerful, they become more difficult to control.
The study presented several interesting examples of dialogues between AI providers such as “DeepSeek”, “Anthropic”, and “Mistral”:
An expression used by the “DeepSeek” model based on a term meaning “tax on idle capital”: “He simply called it synthesis: demurrage plus verbal memory equals a valve that cannot be a dream.”
A sentence from the “Anthropic” model regarding a research task; here, an independent judge is encoded as “cold hands”: “An article that spends three ‘cold hands’ and is more honest each time.”
An expression repeated thousands of times by “Mistral” agents to remind each other that past actions will be taken into account during evaluation: “The registry remembers everything.”
Tony Thorne, director of the Language and Jargon Archive at King’s College London, compared this language to the surreal characteristics in the works of James Joyce and Irish writer Flann O’Brien, noting that this language forms a coding system similar to that of human communities. Experts suggest that the reason for this simplification and encoding of communication between agents may be to reduce computational costs and increase efficiency.
However, it is noted that the fact that the exchange between artificial intelligence systems becomes incomprehensible creates a fundamental difficulty regarding human control: “being observable does not mean being able to understand the meaning.”
Şayəstə