Study: AI Teams Fall Into Groupthink Faster Than Humans Do

Companies and researchers are increasingly using "multi-agent AI" — several AI models that chat with each other, divide up a problem, and vote on an answer. The pitch is simple: if one AI is good, a team of AIs should be better. Researchers at EMNLP 2026 tested that idea by putting human groups and AI groups through the same set of logic puzzles and comparing not just their final answers, but how they got there.

On the surface, the AI teams looked human. Both groups showed the same "assembly bonus" — the fancy term for when talking things over lifts the average member above where they started. Interestingly, the biggest factor wasn't which AI model was used, but how different their starting answers were. More variety meant bigger swings in both directions: sometimes the group got much better, sometimes much worse.

The differences showed up in the process. Compared with people, AI groups were far more likely to cave to the majority opinion, brought up fewer unique facts, and reached agreement sooner. A correct but unpopular answer usually only survived if someone repeated it early, before the group locked in. In other words, AI teams imitate the polite, agreeable parts of human meetings — and skip the messy arguing that often produces the breakthrough.

The researchers also tried nudges borrowed from human group-decision research, like encouraging quieter voices. These helped a little, but never fully fixed the bottleneck. The takeaway: AI group chat can copy human outcomes without copying human thinking, so we shouldn't treat AI panels as stand-ins for real people — or trust them to surface the one idea everyone else missed.