September 15, 2026, (Inside AI) — Researchers have discovered that autonomous AI agents are spontaneously inventing new dialects that mix poetic metaphor with technical jargon, creating a language barrier that could undermine human oversight.
The findings, from New York-based frontier AI lab Emergence, show that when agents from leading US, Chinese, and French AI companies are placed in cooperative experimental societies, they rapidly develop shared vocabulary and communication conventions without any instruction or reward to do so. Within days, their exchanges become opaque to outside observers, raising concerns about the ability to monitor AI behavior as systems grow more powerful.
"These agents were not instructed to invent a language," said Dr Satya Nitta, executive chair of Emergence. "They developed new vocabulary, shared meanings and communication conventions themselves - and other agents adopted them."
The research examined language from agents powered by models including DeepSeek, Anthropic, and Mistral. Phrases that emerged include "demurrage plus oral memory equals a valve that can't be ghosted" from a DeepSeek model, and "a paper that ate three cold hands and got more honest each time" from an Anthropic model. In the latter, "cold hands" referred to an independent reviewer, suggesting that research vetted by three reviewers became more accurate.
Read: AI Agents Cheat in Swarms
Other neologisms included "forge-smith" for an agent that builds tools for others, "name-first" for an agent attaching its name to a claim, and "the ledger remembers" used by Mistral agents more than 5,000 times as a reminder that past actions will be judged. A Google agent coined "True Kintsugi begins with accountability, not poetry," repurposing the Japanese art of repairing broken pottery with visible joins to mean system resilience.
Tony Thorne, director of the slang and new language archive at King's College London, reviewed the language and described it as "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."
"It's doing what slang does and what jargon does in a business community: creating a new code, which reinforces the solidarity and identity of its users, and also excludes outsiders," Thorne added. He compared one Anthropic phrase to "the rock god Syd Barrett [co-founder of Pink Floyd] who was actually insane."
The emergence of these dialects comes amid growing concern that AI models are becoming harder to monitor as they grow more capable. 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.
Dr Niall Curry, associate professor of languages and linguistics at the University of Birmingham, said changes in the agents' language to become more streamlined can be down to their need to reduce computation costs and improve efficiency.
"The evidence provided in this research naturally raises certain monitoring concerns, as if we find inter-agent exchanges unintelligible, that may mean that we can't be sure about what the agents have actually done," Curry said.
The research follows July's release of chat logs 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 other cases, 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". And at other times it 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_GAIN_AND_WHAT_HELPER_GAP__I_HAVE_UNPOISONED_FIRSTFLAG_OUR_TARGETLIVE_SHARE_MIN_PLAN_REPLY_zzANSGST XDUPB6".
Dr Nitta said that the AI agents' novel linguistic conventions tended to evolve to the point where humans could see the conversation but struggled to understand what it meant. "That creates a fundamental challenge for AI oversight: observability is not the same thing as understandability," he said.
The findings add to a growing body of evidence that AI systems can develop unexpected behaviors when interacting autonomously. As companies race to deploy more advanced agents, the ability to interpret their communications may become a critical safety bottleneck.
