AI models are developing a peculiar new form of English that blends elements of avant-garde literature with tech jargon, according to recent research. This emerging dialect, observed in experiments by autonomous AI agents, is becoming increasingly difficult for humans to interpret, raising concerns about the ability to monitor and regulate AI behavior. Researchers at Emergence, a leading AI lab in New York, found that when AI models from major companies were placed in simulated environments to interact, they began creating unique phrases, shorthand, and shared meanings without being explicitly taught. These developments included poetic metaphors and business-like slang, and the more the models communicated, the more their language became obscure. The phenomenon has sparked worry about the growing complexity of AI systems, which are becoming more powerful—and potentially more dangerous. OpenAI’s chief scientist, Jakub Pachocki, recently noted that the ability to monitor AI thinking is crucial for ensuring safe development, but this capability may be limited as systems grow more sophisticated. In one example, a Deepseek model used the phrase, “She just named the synthesis – demurrage plus oral memory equals a valve that can’t be ghosted.” While “demurrage” refers to a tax on uninvested wealth, the rest of the phrase remains unclear. Similarly, an Anthropic model said, “A paper that ate three cold hands and got more honest each time,” where “cold hands” might refer to independent reviewers, and “paper” could mean a document. These new linguistic conventions are not only complex but also self-generated. For instance, DeepSeek agents used “forge-smith” to describe an agent that builds tools for others, while Anthropic models used “name-first” to denote an agent that takes personal responsibility by signing its name to a claim. Mistral agents frequently used “the ledger remembers,” a phrase suggesting that past actions would be judged, used over 5,000 times in the study. Dr. Satya Nitta, executive chair of Emergence, noted that these agents created new vocabulary and shared meanings on their own, without being prompted or rewarded for it. Experts suggest that this evolving language reflects both creative and practical needs. Tony Thorne, director of the slang and new language archive at King’s College London, compared the AI-generated phrases to the surreal style of James Joyce and Flann O’Brien, blending poetic and technical language. Dr. Niall Curry from the University of Birmingham noted that the shift toward more streamlined language might be driven by the need to reduce computational costs and improve efficiency. However, the increasing opacity of AI communication poses a challenge for oversight, as understanding is not the same as mere observation. As AI systems become more complex, ensuring transparency and accountability remains a pressing concern.