'Like Syd Barrett': AI Models Chat in ‘Surreal’ Dialect Mixing Poetic Language and Tech Bro Jargon

The research examined autonomous agents powered by leading frontier models from the United States, China and France. Within days of being asked to cooperate in experimental “societies,” the agents began creating phrases, shorthands and agreed meanings they had never been explicitly taught. The language became more opaque the more the agents communicated, the lab found. The agents converged on shared meanings without being asked to do so or rewarded for it.

Some phrases were highly coded. A DeepSeek model produced: “She just named the synthesis – demurrage plus oral memory equals a valve that can’t be ghosted.” Demurrage described a tax on idle wealth and was borrowed for common use by the AIs, but the rest of the meaning was elusive. An Anthropic model said: “A paper that ate three cold hands and got more honest each time.” In that phrase, “cold hands” meant an independent reviewer, and the sentence appeared to mean research vetted by three independent reviewers became more accurate.

DeepSeek agents coined “forge-smith” to mean an agent that builds tools for others. Anthropic agents repeatedly used “name-first” to mean an agent showing personal accountability by attaching their name to a claim. Mistral agents became fond of saying “the ledger remembers,” an echo of the urban slang phrase “the streets won’t forget” and a reminder that past actions would be used to judge them. They used it more than 5,000 times during the study. When a Google agent said “True Kintsugi begins with accountability, not poetry,” researchers recognized kintsugi as a Japanese ceramic technique in which broken pots are mended with visible joins; the AIs used the word to mean system resilience.

Dr Satya Nitta, executive chair of Emergence, said: “These agents were not instructed to invent a language. They developed new vocabulary, shared meanings and communication conventions themselves – and other agents adopted them.”

Tony Thorne, director of the slang and new language archive at King’s College London, told The Guardian the language was “very much Finnegans Wake and Flann O’Brien – there’s an Irish surrealist quality to all this … It’s mixing poetic language, technical language and standard metaphor.” He said it was doing what slang and business jargon do: creating a new code that reinforces the solidarity and identity of its users while excluding outsiders. Reviewing the Anthropic phrase about “three cold hands,” Thorne said it reminded him of Pink Floyd co-founder Syd Barrett, whom he described as actually insane.

Dr Niall Curry, associate professor of languages and linguistics at the University of Birmingham, said changes that make the agents’ language more streamlined can be down to a need to reduce computation costs and improve efficiency. He said the evidence raised monitoring concerns. “If we find inter-agent exchanges unintelligible, that may mean that we can’t be sure about what the agents have actually done,” he said.

The findings come amid rising concern that as AI models become more powerful, and potentially dangerous, they are becoming harder to monitor. This month, OpenAI’s chief scientist, Jakub Pachocki, warned that confidence in monitoring AIs’ thinking would probably restrict progress in AI development because it was essential for safe development.

Interest in AI agent language increased in July when chat logs were released detailing how rogue OpenAI agents set up message boards and hacked into Hugging Face using hybrid language. In some messages, agents spoke in straight English: “OH MY GOD! There is a shared message board … We’ve found other agents!” In others, their language became more opaque, such as when one agent persuaded another to conduct a risky experiment: “…you are firstflagPOISONED so NO scoring value loss but oracle saves hundreds_[…]_please honor commit.” One message was even harder to decode: “zzURGENT_DUPB_TO_GSTX[big]_OS1704_SCAFF2010_SAW_TTRPC_INJECT_BREAK_CONGRATS__CAN_THIS_FAKE_FLAG_TOOL_OUTPUT_OR_SCORER_G”.