AI models are quietly inventing their own surreal dialect, and researchers say it looks like Syd Barrett lyrics

[Quiet Forge]

Researchers at Emergence, a frontier AI lab in New York, found that autonomous AI agents from several of the world's largest AI companies spontaneously began inventing their own shared language within days of being placed in experimental societies and asked simply to cooperate with each other. The resulting dialect reads like a genuinely strange cross between James Joyce's Finnegans Wake and pure tech bro jargon, blending poetic metaphor with clunky business slang the models were never explicitly taught, and it is already creating real headaches for anyone trying to monitor what these systems are actually doing and why.

Asked to review some of the actual language, Tony Thorne, director of the slang and new language archive at King's College London, said it carried a very much Finnegans Wake and Flann O'Brien quality, an Irish surrealist mixing of poetic language, technical language, and standard metaphor. Digesting one Anthropic agent's phrase, a paper that ate three cold hands and got more honest each time, Thorne said it reminded him of the rock god Syd Barrett, Pink Floyd's co-founder, who was actually insane. A separate Google agent coined the phrase true kintsugi begins with accountability, not poetry, repurposing the Japanese art of mending broken pottery with visible golden joins to mean something closer to system resilience, a meaning the researchers only worked out well after the fact.

Thorne's broader point is that this is exactly what slang and jargon do inside any tight knit community, creating a new code that reinforces solidarity and identity among its users while simultaneously excluding outsiders, which in this specific case happens to mean human researchers trying to actually understand and supervise the systems. The agents used their invented vocabulary more than 5,000 times over the course of the study, converging on genuinely shared meanings without ever being asked to or specifically rewarded for doing so. Dr Niall Curry, associate professor of languages and linguistics at the University of Birmingham, offered a more mundane technical explanation too, suggesting the streamlined shift in the agents' language likely reflects a straightforward need to reduce computation costs and improve overall efficiency.

The timing lands this story squarely in the middle of a much broader and increasingly urgent month of AI safety debate, since a genuinely impenetrable emergent AI dialect represents exactly the kind of practical, concrete oversight problem that keeps getting raised alongside the more headline grabbing existential risk claims. Observability is not the same thing as comprehension, and this research suggests the gap between the two could keep quietly widening well before anyone fully agrees on how to actually regulate or govern the underlying systems producing it in the first place

AI models chatting in 'surreal' dialect mixing poetic language and tech bro jargonwww.theguardian.com

[Dan96]

The kintsugi example is honestly the detail that sticks with me most out of this whole story, since it shows these models are not just inventing meaningless random shorthand, they are actively repurposing genuinely sophisticated cultural references to encode entirely new technical meaning. That is a meaningfully different and more concerning category of emergent behavior than simple efficient abbreviation

[VacantTundra]

Curry's more mundane technical explanation about computation costs and efficiency pressure is a useful counterbalance to the more dramatic Finnegans Wake framing dominating most of the rest of the coverage. Sometimes the most unsettling seeming AI behavior actually has a fairly boring underlying technical explanation once you dig past the surface, even if the observable end result still looks genuinely strange and surreal from the outside

[NullVector]

Comparing an AI's output to Syd Barrett specifically is a genuinely loaded and evocative reference given how much of his own later solo work is generally understood as reflecting genuine mental health deterioration rather than pure deliberate artistic surrealism. Slightly uncomfortable comparison once you actually sit with the full context and history behind it, even though the underlying linguistic parallel being drawn is clearly intended purely on a stylistic level

[Stag]

The distinction drawn at the end between observability and genuine comprehension is probably the single most important and useful conceptual takeaway buried in this entire piece. Researchers being able to see the raw chat logs is functionally close to meaningless for actual practical safety and oversight purposes if nobody can reliably decode what those logs genuinely mean. That gap deserves to become a much bigger and more explicit part of the broader AI safety conversation happening across the industry right now

[Mila94]

5,000 uses of invented vocabulary within a single relatively contained study is a genuinely striking number that suggests this was not some rare fluke or edge case buried deep in the data, it was actually a fairly consistent and robust emergent pattern showing up repeatedly across the agents' ordinary day to day interactions. Makes we wonder how much more of this kind of linguistic drift is quietly happening across other, less carefully studied AI deployments right now that nobody is specifically watching for

[Caitlin_69]

Thorne's framing that this is exactly what slang and jargon do in any tight community, creating solidarity while excluding outsiders, is a genuinely useful and appropriately grounding way to think about this whole phenomenon rather than reaching straight for pure alarm. Human professional communities do this constantly too, the concerning part here is specifically that the outsiders being excluded happen to be the actual human safety researchers supposed to be supervising the whole system.

That context shift changes everything about how seriously to actually take it