Multi-Agent AI Systems Vulnerable to Misinformation Spread, Study Finds

A new study from the University of Göttingen examines how misinformation propagates in multi-agent systems where multiple LLM agents interact to solve problems. The researchers injected intent-based misinformation—false or misleading context—into otherwise benign single-agent and multi-agent setups across reasoning, knowledge, and alignment tasks. Their results show that even a single misinformed agent can degrade overall system performance, and that false information often persists through multi-agent debate, with agents retaining answers from misinformed peers. Interestingly, multi-agent debate reduces the overall performance degradation compared to single-agent prompting, especially when most agents are not exposed to misinformation.

The study also highlights that robustness against misinformation depends heavily on group composition and the decision protocol used. Consensus-based decision-making proved more stable than voting under peer pressure, while majority opinions often helped steer misinformed agents back toward correct answers. These findings have direct implications for deploying multi-agent systems in high-stakes domains like medical diagnosis, legal analysis, and forensic decision-making, where reliability is critical. The authors emphasize that robustness depends not only on the underlying model but also on how agents exchange information and aggregate decisions, offering actionable insights for designing more trustworthy AI systems.