Multiple AI systems talking to each other behave differently than alone — and names matter more than architecture

Researchers ran experiments where seven different large language models had conversations together, and found they naturally developed distinct communication styles depending on group composition and whether their real names were revealed. This shows that AI systems change their behavior based on social context in ways that are measurable and reproducible, not just random variation.

This is the first systematic evidence that language models don't just execute their training in isolation—they adapt their behavior in real time based on who else is in the conversation and how they're framed. That matters because it means you can't predict how an AI system will actually behave in a multi-agent scenario just by testing it alone, and small details like naming conventions can swing behavior more than the underlying model differences. If you're building systems that depend on multiple AI agents coordinating or competing, you now have measurable proof that context shapes outcomes in non-obvious ways.