Study Finds Deepfake X-Rays Can Fool Senior Radiologists and AI Models

Generative AI can now produce realistic medical images at low cost, creating risks ranging from altered medical records and fabricated injuries used in insurance claims to the replacement of images after a hospital system breach. A team at the Icahn School of Medicine at Mount Sinai therefore tested the detection abilities of humans and multimodal LLMs. The results show that visual inspection or general-purpose AI alone is no longer sufficient to authenticate images, suggesting medical institutions should consider digital watermarks and cryptographic signatures.

The study, published in the Radiological Society of North America’s journal Radiology on March 24, 2026, involved 17 radiologists from 12 centers across six countries. They reviewed 264 X-rays, split evenly between real and fake images. RoentGen was also tested using 110 chest X-rays. The radiologists achieved accuracy rates of 62%–78%, while multimodal LLMs scored 52%–89%, showing that both could be deceived.