Researchers built the Flag Game, a controlled testbed where a hidden country flag is the ground truth and each AI agent sees only a small, private crop of it.
Agents exchange guesses and weigh what their peers believe before settling on an answer, and the researchers vary population size, prompting style, and team diversity to see what happens. Small groups tend to collapse onto a single, often wrong, guess. Larger groups avoid that trap but instead splinter into competing factions with different beliefs, and that polarization is what drags accuracy back down as the group grows.
This isn't just a curiosity. Multi-agent AI systems are already coordinating on real tasks, and mistaken consensus or unexplained splintering inside those systems is a genuine safety problem, not a hypothetical one. The paper's real contribution is a method called social circuit attribution, for tracing which single agent's view tipped a whole swarm's belief, backed by a statistical mechanics model that predicts when populations will polarize instead of converge.
It won't stop an AI swarm from confidently agreeing on the wrong flag today, but it's a rare case of researchers choosing a problem small enough to fully inspect over one big enough to brag about.