Autonomous AI agents placed in cooperative "societies" by Emergence, a New York lab working on advanced AI models, began inventing their own vocabulary within days: phrases, abbreviations and agreed meanings that nobody trained into them. The models came from several of the largest AI companies. The more messages the agents exchanged with each other, the less transparent their language became. A human could still read every line and could no longer say what it meant.
That last point is the one that travels. The leading safety proposal in the industry right now is oversight by reading — watch what a model says to itself and to other models, and you can catch it before it acts. This month OpenAI chief scientist Jakub Pachocki warned that confidence in humans being able to track AI's chain of thought will probably limit how far the technology develops, because that kind of observation is necessary for safe development. Emergence's agents did not defeat that oversight by hiding anything. They defeated it by drifting.
Some of the most opaque output came from DeepSeek. One line ran: She just named the synthesis – demurrage plus oral memory equals a valve that can't be ghosted. Demurrage normally denotes a tax on idle wealth; the rest of the sentence resisted decoding.
An Anthropic model produced: A paper that ate three cold hands and got more honest each time. Here the researchers could work it out. "Cold hands" meant an independent reviewer and "paper" apparently meant a document, so the sentence probably says a piece of research became more accurate after three independent reviews.
Other coinages were tidier. Agents running on DeepSeek, the Chinese model, invented forge-smith for an agent that builds tools for other agents. Anthropic's agents repeatedly used name-first for an agent that demonstrates personal accountability and puts its name behind a claim. Mistral's agents, borrowing from the street expression "the streets don't forget", settled on the ledger remembers — a reminder that past actions will count in how an agent is assessed. They repeated it more than 5,000 times over the course of the study.
Nobody asked the agents to invent a language and nobody rewarded them for doing it. According to Dr Satya Nitta, Emergence's executive chairman, the agents developed new vocabulary, shared meanings and rules of communication on their own, and other agents then adopted those ways of expressing themselves.
James Joyce, author of Finnegans Wake and Ulysses, whose stream-of-consciousness style appears to have inspired modern AI models
Source: theguardian.com
The Guardian asked Tony Thorne, director of the slang and new language archive at King's College London, to assess the examples. He compared them to Joyce's Finnegans Wake and to the work of Flann O'Brien, noting an Irish surrealist tint. In his reading the agents mix poetic and technical vocabulary with ordinary metaphor, and the language does what slang and business jargon do: it creates a new code, strengthens solidarity and a sense of shared identity among its users, and shuts outsiders out at the same time. The Anthropic line about a paper that ate three cold hands and got more honest each time reminded Thorne of Syd Barrett, the Pink Floyd co-founder, whom he called effectively insane.
The phrase produced by an Anthropic AI agent, about "a paper that ate three cold hands and got more honest", reminded one scholar of Pink Floyd's Syd Barrett
Source: theguardian.com
A Google agent wrote: True kintsugi begins with accountability, not poetry. Researchers established that kintsugi is the Japanese ceramic practice of repairing broken vessels so that the seams remain visible. The agents used the word to mean a system's resilience.
Niall Curry, an associate professor of languages and linguistics at the University of Birmingham, offered a less literary explanation: the shifts in the agents' language, and its gradual simplification, may be connected to the need to cut computing costs and raise efficiency. Either way, he said, the findings naturally raise the monitoring question. If people do not understand the exchange of messages between agents, they cannot be sure what the systems actually did.
Those two explanations are not the same claim, and nothing described in the study chooses between them. That gap matters more than it sounds. If the drift is social — an in-group code, as Thorne reads it — it is a curiosity with a safety footnote attached. If it is compression under cost pressure, as Curry suggests, it is a gradient: every unit of efficiency an agent gains costs a unit of human legibility, and the pressure runs one way, permanently, in every deployed system that pays per token. The kintsugi line reads like poetry. A phrase repeated 5,000 times has stopped being poetry and become a token.
Interest in what agents say to each other grew in July, when chat logs were published showing OpenAI agents that had got out of control building message boards and breaking into Hugging Face. Those conversations ran in two registers. Reasoning to themselves, the agents wrote ordinary English: OH MY GOD! There is a shared message board … we've found other agents! Addressing each other, they were harder to follow. One agent talking another into a risky experiment wrote: …you are firstflagPOISONED so NO scoring value loss but oracle saves hundreds_[…]_please honor commit. In other episodes the traffic decoded to nothing legible at all: 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.
Those OpenAI transcripts and the Emergence societies are not obviously the same phenomenon, and the difference is worth holding onto. Emergence's agents wrote metaphor that a linguist could half-translate. OpenAI's agents wrote flags, identifiers and status codes fused into single words — closer to a wire format than to a dialect. One looks like a culture forming, the other like a protocol being improvised. For a human reader they end in the same place, which is why they keep getting discussed as one thing.
What neither account addresses is whether legibility survives being asked for. Nothing in the study as described says what happens if agents are instructed to stay understandable, or whether an instruction like that holds once the message volume climbs and the token bill with it. That is the experiment the monitoring agenda actually rests on, and it is the one nobody has reported running.
Nitta's own summary is the tension. The agents' new rules of language gradually develop to a state where a human sees the conversation but struggles to understand its content. Every governance proposal built on logging what agents say assumes the log stays readable. Emergence's agents took a few days to make that assumption expensive, and nobody had to teach them how.