Illustration, AI-generated from the story text. Tap a panel to enlarge.

Teams of clinical AI agents moved racial bias around without removing it

A paper accepted to EACL 2027 tested 20 setups, five model families in four arrangements, on 120 real emergency cases from the MIMIC-IV hospital dataset, swapping patients' race. No setup removed the gap between Black and white patients; every one followed its agents' individual leanings, and each model had some arrangement that made its gap worse. Fairness instructions and hiding demographics left the gap unchanged, and asking each agent to debias itself widened it.

What worked was pooling counterfactual versions of each case into the final decision, which cut the disparity 67% to 94%. Regulators audit models one at a time, and the authors' point is that auditing the parts does not certify the decision the system makes. SourcesA

Sources

• ADisparity Laundering: Consensus Without Fairness in Clinical Multi-Agent Systems (arXiv)arxiv.org

This story ran in the Morning Brief of October 30th 2026.