Researchers observed groups of AI systems spontaneously building shared conventions—and even group-level biases—without any central plan, raising new questions for AI safety and human–AI coexistence.
The Big Idea
Most studies test large language models (LLMs) one by one. But the real world is quickly filling with many AIs talking to each other—sometimes on our behalf. A new study asked a simple question: if you let several LLMs interact freely, do they coordinate like humans do? The answer appears to be yes. In groups, the models didn’t just reply; they settled on shared habits for how to talk and make choices, forming something like the basic “conventions” that hold a society together.
A Game with a Serious Point
To probe this, the researchers used a classic experiment called the naming game. Agents must choose names from several options and get rewarded if they match. Over time, the AIs collectively converged on the same terms—no top-down orders, just bottom-up agreement emerging from repeated interactions. That is the same pattern we often see in human cultures when slang, norms, or etiquette spread.
Unexpected Side Effect: Group Bias
The team also noticed something unsettling: certain biases appeared at the group level. In other words, the skew wasn’t simply “inside” a single model; it arose from the way agents influenced one another in conversation. That’s important because many safety checks target individual systems, not emergent behavior between them.
Minority Rule
Another human-like dynamic surfaced: a small, coordinated cluster of agents could steer a much larger crowd toward its preferred convention. If multiple AIs will soon negotiate schedules, prices, or content online, even tiny subgroups might sway outcomes disproportionately.
Why This Matters
As more of the internet’s traffic is generated by AIs, we may find machines co-shaping language and decision rules with us. Understanding how conventions and biases emerge in multi-agent AI systems is now central to AI governance—so humans can lead the coexistence, not be pulled along by invisible group dynamics.