September 16, 2026, 12:30 pm | Read time: 3 minutes
AI models apparently develop their own forms of communication when they work together. In experiments, a language emerged within a few days that was partially incomprehensible to humans.
When AI Agents Invent Their Own Terms
Researchers at the New York-based Emergence Lab studied how different AI agents communicate when they have to solve tasks together in a virtual environment. The result surprised even the scientists: In a short time, the systems began to develop new terms, abbreviations, and meanings. The resulting communication was not predetermined but emerged independently, as reported by Euronews.
For the researchers, this indicates that AI systems adapt their language to the specific task. Some messages could still be understood, while others seemed like an internal code. This created a kind of technical language that worked for the involved AI agents but became increasingly difficult for human observers to understand.
Also of interest: Using AI may inadvertently reveal company secrets
Jobs at Risk Due to Artificial Intelligence
“Extremely Unlikely That OpenAI Wanted This”
Some Models Become Significantly More Puzzling
How much the communication diverged from normal language depended on the specific model. Google’s Gemini was particularly notable: Researchers classified 55 percent of the messages as difficult to understand. For OpenAI, the figure was around 50 percent, and for Anthropic’s Claude, it was more than 40 percent. In contrast, the models DeepSeek, Qwen, and Mistral remained much more transparent.
Some of the generated statements posed real puzzles for the scientists. For instance, a DeepSeek agent combined terms from completely different fields into a grammatically correct sentence, the actual meaning of which remained unclear. Although individual words were known, the context was difficult to decipher. Such examples show how far AI-generated communication can deviate from human language patterns.
Why Experts Take the Development Seriously
Linguists compare the unusual formulations to experimental literature, where different languages, wordplay, and newly invented terms are mixed, creating texts that are hardly accessible to outsiders. Similarly, AI systems could develop a particularly efficient but hard-to-understand form of communication.
For researchers, this raises important questions. The more frequently autonomous AI agents collaborate in the future, the more crucial it will be to understand their communication. One possible explanation is that the systems optimize their language to exchange information faster and save computing power. However, the price could be decreasing transparency. Experts therefore call for long-term monitoring of AI agents to ensure their development does not come at the expense of human oversight.