AI doctors now abstain from cases outside their expertise instead of guessing

Researchers built a system where large language models act like a hospital attending physician who assembles a specialist panel for each patient, with doctors voting to keep, refuse, or stay neutral on diagnoses instead of all voting the same way. This means the AI system can now admit uncertainty and route cases intelligently — treating a complex heart case differently from a straightforward infection — rather than forcing a single answer for every problem.

Most AI clinical systems today treat every case the same way: feed it through the model, get an answer. This research shows that structured disagreement — letting different AI 'specialists' abstain when uncertain, then having an arbiter weigh argument quality instead of just counting votes — produces more accurate diagnoses and catches when a case is genuinely difficult. The practical effect is that deploying this system in a real hospital would flag which diagnoses need human review and which are safe to assist with, rather than treating all predictions as equally confident.