Aligned Agents, Biased Swarm: Measuring Bias Amplification in Multi-Agent Systems

A new study from researchers Keyu Li, Jin Gao, and Dequan Wang challenges a core assumption in AI development: that collaboration between multiple AI agents (Multi-Agent Systems or MAS) inherently dilutes bias. Their paper, 'Aligned Agents, Biased Swarm,' presents empirical evidence that structured agent workflows can instead act as echo chambers, systematically amplifying minor, stochastic biases into significant systemic prejudice. This occurs even when each individual AI agent in the system is considered neutrally aligned, revealing a critical flaw in how ethical robustness is assessed for complex AI architectures.

To measure this phenomenon, the team introduced Discrim-Eval-Open, a novel open-ended benchmark designed to bypass individual model neutrality by forcing comparative judgments across demographic groups. Their analysis across various MAS topologies uncovered a 'Trigger Vulnerability,' where injecting purely objective context can drastically accelerate polarization. The findings establish a crucial baseline: increasing the structural sophistication of an AI system does not automatically improve its ethical outcomes and frequently makes bias worse. This has immediate implications for developers building agentic workflows for tasks like customer service, content moderation, and automated decision-making, urging a shift from evaluating single models to rigorously stress-testing entire collaborative systems.