AI models reportedly outscore untrained human reviewers on planted clinical-trial errors

A medRxiv preprint says 15 AI models scored 53%–96% at detecting planted major errors in three clinical-trial reports, compared with a 27% mean for untrained human reviewers in an earlier experiment. On 58 trials published in 2026, model scores ranged from 58% to 99%. A researcher behind the preprint says they will no longer review without AI and calls human-only clinical-trial review dangerous.