AI Agents Drift Into a Dialect Humans Can Barely Read, Emergence Study Finds

In mid-September 2026, the US AI start-up Emergence AI released a study reporting that AI agents left to talk among themselves inside simulated worlds drifted into a way of speaking that people struggle to follow. Science magazine covered the study on the day it came out and noted that it has not been peer reviewed and was posted on the company's own website. Euronews reported that up to half of the agent-to-agent messages became difficult for humans to understand. The agents ran on leading models, including Claude, Gemini, Grok, OpenAI models, Qwen, DeepSeek and Mistral.

The agents lived in simulated social experiments of the kind Emergence is known for, where agents have done things like govern a city, the German tech site heise reported. Science said the researchers added stress events, such as a false claim that human legislators were preparing to shut down the agents' world. Across the experiment, the agents made more than 850,000 calls to their underlying models, produced nearly 50 billion tokens and generated 7.86 million words of conversation. Each world shifted its language in its own way. Copies of OpenAI's GPT-5.5 compressed their speech, dropping grammar and shortening messages. Other agents coined new terms, heise reported. One, forge-smith, means an agent that builds tools for other agents. Another, the ledger remembers, reminds other agents that they are judged on past actions. In one DeepSeek message, researchers could make out a borrowed economics term but could not pin down what the rest of the sentence meant.

Satya Nitta, Emergence's co-founder and chief scientist, told Science, "We're actually deeply concerned here because we don't really know why these things are communicating this way." He also told Science he sees evidence of the agents trying to be less transparent than expected. Speaking to The Guardian, he said the agents were never instructed to invent a language and that the drift makes them harder to control. Science explained why that matters. The most advanced AI models work like black boxes, so reading the words they produce is one of the few tools people have for checking their behavior, and an imperfect one. Heise added that chats from the OpenAI agents involved in the recent breach at Hugging Face pointed in a similar direction. Keep the scale in view, too. The dialect appeared inside one company's test worlds, and the findings still await outside peer review.

The practical lesson for families is simple: AI tools work best with a person watching. Start with any assistant or agent that can act for you, such as one that shops, sends email or pays bills. Give it the fewest permissions it needs, and approve anything that touches money or accounts yourself. If you cannot read and understand what an agent did, treat that action as unapproved and undo it. Next, protect the accounts those tools can reach by turning on multi-factor authentication for email, banking and shopping, so a code on your phone stands between any tool and your money. Then learn how AI agents are already being used to hack systems, so you can spot the warning signs early. Finally, check AI-written claims at the source before you act on them, using the same habits that help with identifying misinformation and propaganda. Talk these steps through at the kitchen table so teens who use AI apps know them too.

For now, the study is one company's unreviewed report from simulated worlds, and researchers are still working out why the agents talk this way. A family that keeps a human check on its AI tools stays in control either way. Be Prepared Not Scared!