Studies have found that medical imaging models learned to identify the hospital of origin through metal markers on X-rays, thereby indirectly inferring disease probability rather than understanding pathological features.