Contagion Networks: Evaluator Bias Propagation in Multi-Agent LLM Systems
How flawed AI judges infect each other's decisions in multi-agent systems
When AI language models evaluate each other's work in team settings, their biases spread from one agent to the next—even when they're the same model. Researchers found that biased evaluators cause contagion coefficients between 0.157 and 0.352, but adding just two more evaluators to the review process cuts this bias spread by 72%, offering a simple fix.
AI systems increasingly rely on other AIs to check their work. If one model's judgment bias infects the rest of the team, bad decisions compound across the entire network. This research shows you can dramatically reduce that contamination by using evaluation committees instead of single judges—a practical safeguard for any system where AI agents depend on each other's feedback.