If you want AI to help humans coordinate without taking over the shared norm, this study suggests keeping agent participation low rather than simply adding more agents.

The researchers put humans and LLM agents into 24-person groups and had them repeatedly agree on descriptions of the same image.

AI agents can go from helping humans coordinate to defining what the group agrees on, depending on their share of the group

Low AI participation helped humans reach agreement, medium participation disrupted it, and high participation shifted agreement toward AI-led norms

With 12.5% AI, consensus improved by 8.0% over the all-human group. At 33.3% and 50%, agreement got worse. At 75%, strong consensus came back, but now humans were moving toward the agents’ language.

That changed the kind of agreement too. Human-led groups used more concrete, real-world descriptions. Agent-led groups became more abstract and geometric.

The reason: agents start with more similar language and stay more consistent, so their wording can become the group default as their numbers rise.