Dissent Absorbed: Consensus-Oriented Reconstruction Bias in LLM Multi-Agent Committees

In real expert committees, dissenting opinions surface latent risks and act as safeguards against premature consensus. We ask whether LLM-based committee simulations preserve this function. We reconstruct 161 voting items from FDA and CDC advisory meetings. Each committee is replaced with an LLM multi-agent panel that reconstructs the votes from the full meeting record, repeated across five models. Most of these panels track the human collective decision closely. The individual votes behind that decision tell a different story. The LLM panels flip about half of the human minority votes to the majority side, while flips in the reverse direction are rare. The soft label distribution of each agent's vote shows the same pattern, with agents expressing markedly more uncertainty whenever they do cast a minority vote. Aggregate agreement with a human panel can therefore be satisfied by a simulator that discards most of the dissent it was meant to reproduce, and dissent reconstruction fidelity should be measured as a separate criterion.