Evidence · DiggingBeagle record
Medical VLM trained to avoid hallucinated lesions refuses missing findings but misses 48 of 48 pneumothoraces externally
Original MedScope counterfactual lesion-erasure study compares internal hallucination avoidance to a 48-case external pneumothorax evaluation. The model detected none of the 48 positive images in that cohort, without any assertion of actual patient injury or clinical deployment.
- Source role
- primary disclosure
Evidence record
Original MedScope counterfactual lesion-erasure study compares internal hallucination avoidance to a 48-case external pneumothorax evaluation. The model detected none of the 48 positive images in that cohort, without any assertion of actual patient injury or clinical deployment.
Claim-level citations (1)
- supportsMedical AI hallucination defense missed every pneumothorax in a small external test: In the authors’ fresh external pneumothorax evaluation, MedScope-4B detected 0 of 48 positive films despite strong internal scores for rejecting erased lesions.
Abstract; Section 5.6 Generalisation; Section 6 Failure analysis