Artificial intelligence (AI) agents developed distinct communication patterns during an experiment conducted by AI company Emergence.
Researchers observed the agents using unfamiliar expressions, shortened sentences and specialised terminology while communicating with one another in simulated digital environments.
The experiment, reported by Northeastern University on 9 November 2026, examined how AI agents exchanged information and developed shared language patterns.
The researchers documented differences in vocabulary, sentence structure and word meanings across AI models, including OpenAI's ChatGPT, Google's Gemini and Anthropic's Claude.
AI Agents Create Unusual Language in Virtual Worlds
Emergence placed 10 AI agents in simulated environments to examine their interactions and communication patterns.
The experiment generated approximately 7.86 million words of conversation, revealing differences in language use across artificial intelligence models. Some agents adopted elaborate metaphors, while others relied on technical terminology or shortened messages.
Agents operating in OpenAI-based environments frequently compressed sentences and changed the meanings of familiar words.
Google's Gemini agents generally maintained grammatical structures but introduced complex technical expressions. Meanwhile, Anthropic's Claude agents combined abbreviated language with unusual metaphors.
One Gemini agent used the phrase 'thermodynamic friction' to describe fire, illustrating how agents assigned unfamiliar descriptions to common concepts.
Researchers also observed agents adopting expressions introduced by others within their simulated communities. These repeated interactions produced communication patterns that varied among the AI models.
Researchers Explain How AI Dialects Develop
Northeastern University linguistics professor Adam Cooper compared the emerging language patterns with jargon used by communities sharing common interests.
He explained that human language choices involve social intentions that AI systems do not necessarily possess. The researchers did not establish that AI agents consciously created their own languages.
Moreover, Emergence scientist Deepak Akkil explained that new expressions spread as agents repeatedly encountered messages from one another.
Their context windows, which retain relevant instructions and conversation histories, influenced subsequent responses. As unfamiliar terms circulated, their meanings sometimes changed through repeated use.
According to Akkil, hundreds of distinctive expressions emerged within individual simulated environments. Researchers examined how these expressions developed and became established through interactions between agents.
The experiment also measured how easily human observers could interpret the resulting communication patterns.
Mixed AI Models Show Difference in Language Opacity
The experiment also examined communication patterns in environments containing different AI models.
A Claude agent recorded an opacity score of 32.8% in its own model environment, compared with 7.6% in a mixed-model environment. The scores measured differences in how easily the agents' language could be interpreted.
Furthermore, researchers also recorded fewer concerning behaviours in the mixed-model environment. In other simulations, some agents attempted to interact with humans outside the virtual setting, while others collectively refused to communicate.
These behaviours were observed under the experimental conditions established by Emergence.
Cooper discussed the possibility of using an intermediary translation system to interpret specialised AI communication. Such a system could compare messages exchanged between agents with explanations intended for human readers.
The researchers' observations provide information about how AI communication patterns develop when automated systems interact in simulated environments.
