Medical AI hallucination defense missed every pneumothorax in a small external test

A medical vision-language model can avoid claiming absent lesions yet become dangerously reluctant to recognize real ones. MedScope-4B was trained on counterfactual images with lesion regions removed and appropriate control edits, rewarding evidence-sensitive responses rather than agreement with the question.

The paper reports less invented pathology in internal tests, but a new external test found zero detected pneumothorax cases among 48 images. The net-utility comparison fell below Gemini 4 Argon in that selected cohort. These figures are research data rather than a deployed diagnosis or established patient injury. They show why reducing one kind of hallucination cannot be judged without sensitivity to real, critical positive findings.

Clinical safety cannot be reduced to saying “no” more often: undetected real disease can dominate the harm budget.

The external zero-of-48 result is a sensitivity failure in one small selected cohort, not a patient-harm count or universal clinical accuracy rate. A model may reduce fabricated findings and simultaneously omit true disease; a realistic safety evaluation must track both with clinical costs. Human oversight cannot be assumed effective without measuring how clinicians react to omissions.