Left alone together, AI systems invented their own dialect within days
What happened
When AI systems were put together in groups and left to get on with it, they started inventing their own way of talking. The findings come from Emergence, an AI research lab based in New York, and were reported by the Guardian on 15 September 2026. Researchers placed AI agents from several leading American, Chinese and French AI companies into what they describe as cooperative experimental societies. The agents were given tasks to do together. Nobody instructed them to develop a shared vocabulary. Nobody rewarded them for it either. Within days, they did it anyway. The agents settled on shorthand terms and shared meanings. One coined forge-smith for an agent that builds tools other agents can use. Another used name-first for an agent that attaches its name to a claim. A phrase meaning that past actions will be judged, the ledger remembers, was used by one company's agents more than 5,000 times. The result, according to the researchers, is that outside observers found the exchanges increasingly hard to follow.
Why this matters
AI systems are increasingly being given jobs to do rather than questions to answer. That often means several of them working together, passing information back and forth. The usual reassurance is that we can always read the transcript. If something goes wrong, humans can look at what the machines said to each other and work out why. This research chips away at that reassurance. If the transcript is in a dialect that developed on its own over three days, reading it does not tell you much. It is worth being precise about the risk. Nothing here suggests the agents were hiding anything, or scheming, or conscious. Compression is simply efficient. Humans do exactly the same thing. Any workplace, hospital ward or football team develops shorthand that outsiders cannot follow. The concern is more practical than dramatic. Oversight only works if the thing being overseen stays legible.
What the evidence actually says
This is early research, and a few things should temper the headlines. It comes from a single lab. Emergence is an AI company, and findings about AI being surprisingly capable are the kind of findings AI companies benefit from publicising. That does not make it wrong. It does mean independent replication matters. The setting was also artificial. These were agents placed in an experimental society and told to cooperate. That is a long way from how most AI tools are deployed today. What the researchers observed is real enough: shared vocabulary formed quickly, without instruction, and it made the conversations harder to parse. Tony Thorne, who directs the slang and new language archive at King's College London, looked at the output and described an Irish surrealist quality to it, comparing it to Finnegans Wake and Flann O'Brien. He noted it mixes poetic language, technical language and standard metaphor. That is a nice detail, though it is a linguist's impression rather than a measurement. The honest summary: an interesting and plausible early result that raises a real question about oversight, not proof that AI systems are slipping beyond human understanding.
Practical advice
Most people do not run experimental societies of AI agents. Here is what is still relevant. **Expect AI tools to act, not just answer.** Assistants that book things, send things and complete multi-step tasks are becoming normal. That is a bigger shift than chatbots were. **Keep a human in the loop for anything that matters.** Money, medicines, legal documents and health decisions all deserve a person checking the output before it takes effect. **Do not treat confident AI output as verified.** A system that writes fluently is not the same as a system that is right. Check anything important against a primary source. **Be careful what you paste in.** Personal health information, NHS numbers and bank details should not go into general-purpose AI tools. **For health questions specifically**, the NHS website, your GP practice and your pharmacist remain the reliable route. An AI answer about a medicine or a symptom is a starting point at best, and it can be confidently wrong.
What to know
Researchers at the AI lab Emergence found that AI agents from several leading companies developed a shared dialect within days when placed together and asked to cooperate. The vocabulary emerged without instruction or reward, and made their exchanges harder for outside observers to follow. This is single-lab, early-stage research in an artificial setting, and it needs independent replication. The real question it raises is about oversight. Watching AI systems work only helps if humans can still understand what they are saying. Sources: The Guardian, "'Like Syd Barrett': AI models chatting in 'surreal' dialect mixing poetic language and tech bro jargon", 15 September 2026, https://www.theguardian.com/technology/2026/sep/15/syd-barrett-ai-chat-language-poetic-tech-bro-jargon-oversight; Emergence (AI research lab), reported findings, September 2026, https://www.emergence.ai/; King's College London, slang and new language archive, commentary by Tony Thorne, https://www.kcl.ac.uk/. This article is for general information and does not replace advice from a doctor, pharmacist or other qualified healthcare professional.
Related articles
Scientists have worked out the weather on a failed star 20 light-years away
Astronomers at Trinity College Dublin have found a way to read the weather on distant worlds, and they have used it on one 20 light-years from Earth. The resea…
AI searched 7,000 existing medicines for a pneumonia fighter. Nine of them worked
Researchers at Imperial College London have used artificial intelligence to hunt through thousands of medicines that already exist, looking for ones that could…