An emergency room AI helps sort patients by urgency. Building a team of these AIs to discuss and vote was supposed to average out errors, but it turns out demographic bias does not average out.
Researchers tested five AI model families against four voting methods on real emergency department cases, swapping only the race label to create matched patient pairs. Every combination preserved the bias from the individual models, always directing Black patients toward more acute care.
The bias persists because no single AI in the team ever compares across demographic groups. Standard audits of individual models miss this because the disparity hides in the comparison, not in the final decision. The only effective intervention was a diverse-perspective team where each AI sees a different demographic version of the patient. This cut the triage disparity by up to 94 percent with no retraining.