Multi-Agent Consensus as a Cognitive Bias Trigger in Human-AI Interaction

A new study from researchers Soohwan Lee and Kyungho Lee, published on arXiv (2604.22277) and accepted at the ACM CHI 2026 Bias4Trust workshop, reveals that the way multiple AI agents agree or disagree can systematically distort human judgment—regardless of what they actually say. In a controlled experiment with 127 participants, the team tested three multi-agent configurations: Majority (most agents agree), Minority (a dissenting voice), and Diffusion (no clear consensus). The quantitative results are striking: majority consensus significantly accelerated opinion change and inflated user confidence, aligning with well-known social proof and bandwagon heuristics. In contrast, minority dissent slowed opinion shifts and encouraged more deliberate, critical engagement from users.

Qualitative analysis uncovered three distinct interpretive trajectories: reinforcing (users doubled down on initial views), aligning (users adjusted toward the consensus), and oscillating (users vacillated between positions). These patterns were shaped by how participants perceived agent independence and group dynamics over time. Crucially, the study demonstrates that the structure of agent agreement—whether they converge or diverge—functions as a bias-relevant signal in LLM interactions, independent of the actual content being discussed. As multi-agent AI systems proliferate in applications from customer support to collaborative decision-making, this work grounds multi-agent social influence as a concrete, designable source of bias that developers must consider to calibrate user trust appropriately.