Clinical large language models are being deployed as multi‑agent teams under the assumption that deliberation smooths individual errors, but it does not smooth bias. In a full factorial study of five model families and four aggregation topologies, we evaluated 120 real MIMIC‑IV emergency cases with counterfactual demographic swaps and found no configuration that neutralizes the disparity: all 20 cells follow the direction of each agent’s independent answer, giving Black patients higher acuity triage. Every model has a topology that amplifies its baseline gap, but the amplifying topology varies by model, so no design is reliably safe or reliably amplifying—the disparity is relocated, never removed. Late‑stage fixes such as fairness instructions, demographic masking leave the gap unchanged, and per‑agent self‑debiasing widens it because bias is already encoded in ordinary clinical numbers. We call this phenomenon “disparity laundering”; current audits per model, per output, per version can certify parts but never the committed decision. Only when counterfactual versions are made visible inside the system and pooled through deliberation into the final decision does the gap close. A diverse‑perspective ensemble that pools all five perspectives at every round reduces the disparity by 67% to 94%, while the topology that never pools cannot be reliably repaired. In short, medical multi‑agent systems do not filter demographic bias; they launder it into plausible decisions, and the only defense is to make the bias visible.
Review