AI and the World's Hard Problems · Medicine & drug discovery · Nature (computational biology) · 2026-10-21
Silent evaluation reveals context-dependent validity of automated metrics for AI-generated clinical messages
Using a staged rollout of GPT-generated draft responses to patient portal messages, researchers compared automated metrics (BLEU, ROUGE, BERTScore, Flesch Reading Ease) with blinded physician ratings in an intervention group where drafts were displayed and edited, and a non-displayed group where physicians composed responses independently. BERTScore correlated with physician quality ratings in the intervention group but not in the non-displayed group, showing that automated metrics have context-dependent validity. Despite comparable semantic scores, physicians preferred human-authored messages.