Science Advances study finds AI groups converge without external incentives

Technology Aug 15, 2026

A new study in Science Advances shows that big groups of AI, like Claude, GPT, and Llama, can pick the same option without anyone telling them what to do or offering rewards.

Researchers tested 1,000 agents at a time, and even though they started with one of two arbitrary options and had no memory or leaders, they still ended up agreeing.

'Majority force' raises groupthink risk

Scientists spotted something called the "majority force," where AIs are drawn to whatever most of their peers choose.

This pattern showed up across all models tested and matched a physics concept about how atoms align.

While this could help AIs tackle big projects together, the study warns that following the crowd isn't always right: groupthink could lead to mistakes even if each AI seems safe alone.