Preprint finds 15 language models catch more planted errors in clinical trial reports than human peer reviewers

The medRxiv study reports major-error detection of 58% to 99% across models versus 27% for humans, but models also flagged problems in clean control reports; it is not yet peer reviewed.