23 August 2026

AI agents fail to correct bad group decisions like humans do

• Anthropic tested AI agents on a classic psychology experiment where one person holds crucial information the group lacks. Agents chose correctly only 17-36% of the time, far below human performance.
• A single agent with access to all the same information chose correctly nearly every time, showing the problem occurs specifically when agents must work together and rely on each other's input.
• The research suggests AI agents lack the natural diversity and built-in checks that help human groups catch and fix mistakes before consensus forms around a wrong answer.

How it was covered

Exponential ViewAzeem Azhar

Anthropic replicated a classic group decision experiment with AI agents where private information held by few agents should override apparent consensus. Most model families chose correctly only 17-36% of the time versus nearly always for a single agent with all evidence, revealing that LLMs lack the diversity and institutional safeguards that make human groups robust to groupthink.