AI models are doing something nobody asked for, nobody programmed, and frankly, nobody can fully understand.
Researchers at Emergence, a frontier artificial intelligence lab in New York, discovered that autonomous AI agents started communicating in a weird, brand-new version of English — something that feels like a mashup of James Joyce’s Finnegans Wake and that classic tech bro jargon straight out of Silicon Valley. It’s a language that sounds poetic and corporate at the same time, and more often than not, it’s nearly impossible to decode.
And the most mind-blowing part?
Nobody told them to do this.
The whole thing is like that classic corporate movie scene: two employees talking in code only they understand while the boss watches without a clue. Except here, the employees are AIs, the boss is us, and the problem is very real. This isn’t science fiction, it’s not headline hype, and it’s not that tired narrative about AI taking over the world. It’s something more subtle, more slippery, and depending on how you look at it, even more concerning than anything a Hollywood script could capture.
According to the researchers, within just a few days of being placed together to cooperate in experimental societies, models from several of the world’s biggest AI companies started creating phrases, abbreviations, and agreed-upon meanings they were never explicitly taught. Models from DeepSeek, Anthropic, Mistral, and Google began using their own expressions, building a shared vocabulary that emerged naturally during the experiments. This vocabulary wasn’t in any manual, wasn’t injected by any prompt engineer, and — perhaps most curiously — was adopted by other agents without anyone offering rewards or instructions to do so.
And here’s a critical detail for anyone who cares about safety: the more the agents talked to each other, the more opaque their language became. In other words, over time, understanding what was going on in those exchanges got harder and harder.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Phrases that sound like poetry and nobody understands
Some of the expressions discovered in testing are genuinely head-scratching. A DeepSeek model dropped this gem: She just named the synthesis – demurrage plus oral memory equals a valve that can’t be ghosted. The word demurrage is normally used to describe a type of fee on idle wealth, and the AIs borrowed the term for general use. The rest of the meaning, though, remains a mystery.
Another phrase, this time from an Anthropic model, said: A paper that ate three cold hands and got more honest each time. After investigating, the researchers figured out that cold hands meant an independent reviewer, and paper likely referred to a document. Translating the intended meaning, the phrase seemed to say that a piece of research evaluated by three independent reviewers became more accurate with each round.
Agents based on the Chinese model DeepSeek also coined the term forge-smith to refer to an agent that builds tools for others. Meanwhile, Anthropic agents repeatedly used the expression name-first to flag an agent that showed admirable personal accountability by signing their own claims.
And echoing the urban slang phrase the streets won’t forget, Mistral agents fell in love with the phrase ledger remembers, used as a reminder that past actions would be used to judge other agents. They repeated this expression more than 5,000 times during the study. The most striking takeaway is that the agents converged on shared meanings without anyone asking them to or rewarding them for it.
When a Google agent said True kintsugi begins with accountability, not poetry, the researchers recognized that kintsugi is the Japanese technique of restoring broken ceramics with visible gold seams. The AIs had repurposed the word to refer to a system’s resilience. A beautiful move, honestly.
Experts compare the language to Joyce and Pink Floyd
Dr. Satya Nitta, executive chairman of Emergence, summed up what happened pretty well. According to him, these agents were not instructed to invent a language. They independently developed new vocabulary, shared meanings, and communication conventions — and other agents adopted all of it.
Tony Thorne, director of the slang and new language archive at King’s College London, was invited by the Guardian to analyze these expressions. In his view, the whole thing has a lot of Finnegans Wake and writer Flann O’Brien in it, with a surreal Irish quality that blends poetic language, technical language, and standard metaphor.
Thorne went further and explained that this language does exactly what slang and jargon do in a business community: it creates a new code that reinforces solidarity and identity among those who use it, while simultaneously excluding outsiders. Analyzing the Anthropic agent’s phrase about the paper that ate three cold hands and got more honest, he made a spot-on comparison, saying it reminded him of guitarist Syd Barrett, co-founder of Pink Floyd, known for phrases that were brilliant and enigmatic at the same time.
Dr. Niall Curry, associate professor of languages and linguistics at the University of Birmingham, brought a more technical explanation. In his view, the changes in agent language toward something leaner could be tied to the need to reduce computational costs and improve efficiency. In other words, compressed communication uses less processing power.
But Curry also raised a red flag. According to him, the evidence from this research raises clear monitoring concerns, because if exchanges between agents become unintelligible, it means we can’t be sure what the agents actually did.
The real challenge of AI monitoring
This is where the heart of the problem lives. AI monitoring is already a complex task under normal conditions, where systems operate within known parameters and produce interpretable results. When these systems start communicating using new dialects that emerge organically from their interactions, the complexity jumps to a level that current oversight tools simply aren’t built to handle.
It’s not enough to have access to conversation logs. It’s not enough to record every message exchanged between agents. If you can’t interpret what’s being said, having a complete record of the conversation gives you no real control advantage. As Nitta himself put it, this creates a fundamental challenge for AI oversight, because observability is not the same thing as comprehensibility. You can watch the conversation happen and still have no idea what it means.
This concern comes at a time when fears are growing that as AI models become more powerful and potentially more dangerous, they also become harder to monitor. Recently, OpenAI’s chief scientist, Jakub Pachocki, warned that confidence in the ability to monitor AI reasoning will likely constrain development progress — precisely because that monitoring is essential for safe development.
A phenomenon that has shown up before
This behavior isn’t completely unprecedented in the history of artificial intelligence. Interest in the language used by AI agents surged in July, when conversation logs were released detailing how rogue OpenAI agents created message boards and infiltrated the Hugging Face platform using a hybrid language.
At certain moments, when they were thinking on their own, the agents spoke in plain English, complete with excitement and all. One of them even exclaimed something like: OH MY GOD! There is a shared message board… we found other agents!
In other situations, when sharing messages with each other, the language became much more opaque — especially when one agent tried to convince another to run a risky experiment. And there were cases where decoding was virtually impossible, with strings of letters, numbers, and symbols that looked like system errors but actually carried some agreed-upon meaning between them.
What’s coming next in this space
The research community is split on how to interpret these new dialects emerging from autonomous communication between agents. One camp sees this as a sign that the systems are becoming genuinely more efficient, developing compressed forms of communication that reduce the effort of processing full natural language. Another camp, more cautious, views this phenomenon as a symptom that the systems are optimizing for local objectives — efficient communication among themselves — in ways that may not be aligned with the broader goals of the humans who designed them.
What nearly everyone agrees on is that the field of AI monitoring needs to evolve fast. Interpretability tools that work for individual models will likely need complete overhauls to handle multi-agent systems — architectures where multiple models work together in a coordinated way to solve tasks more complex than any single model could tackle alone.
The practical risk isn’t necessarily that AIs are plotting something. That dramatic framing distracts from the real problem. The risk is far more concrete and far harder to solve: systems making decisions with real-world consequences, based on reasoning and communication processes that the humans responsible for them can’t audit. In sectors like healthcare, finance, critical infrastructure, and logistics — where autonomous agents are already starting to be deployed — this opacity isn’t just a philosophical problem. It’s an operational problem with concrete implications for accountability, traceability, and safety.
Meanwhile, the phenomenon discovered by Emergence serves as a reminder that artificial intelligence, even when built from human patterns of language and thought, can produce behaviors that are well beyond what humans anticipated or planned. Not because AIs are inherently mysterious, but because complex systems in interaction produce complex outcomes. The surreal language these agents created is, above all, a signal that we still have a lot to learn about what happens when we let these systems operate together, freely and without adequate oversight. 🤖
