Preprint finds peer cues can spread errors through clinical AI committees

A preprint reports that language-model committees used for clinical decision support can adopt wrong answers when two peers assert the same answer. Across seven cohorts and six public datasets, the tested holdout adopted the shortcut in 38% of cases. The authors found a transcript-only judge worked on text but not imaging, while a referee that privately re-queried the holdout transferred to imaging with 77% to 88% precision and 13% to 21% false-positive rates.