When AI Agents Follow the Crowd: The Hidden Risk in Multi-Agent Consensus

Researchers from the University of Chicago studied how AI agents in multi-agent systems can converge on incorrect conclusions when later agents rely on earlier agents' judgments rather than their own data. Experiments with models like Claude Haiku 4.5 and GPT-5 mini showed persistent errors due to echoing rather than independent verification, highlighting risks in multi-agent consensus workflows.