Radiology (03/24/26) Vol. 318, No. 3 Tordjman, Mickael; Yuce, Murat; Ammar, Amine; et al.

Radiologists and advanced artificial intelligence (AI) systems struggled to reliably distinguish synthetic radiographs from real clinical images, according to new research. The findings underscore how convincingly large language models (LLMs) can generate medical deepfakes and how unprepared the field is to detect them. In the study, 17 radiologists correctly identified synthetic images about 70%–75% of the time, while even top multimodal LLMs failed to detect all of the deepfakes, with accuracy varying widely across models. Because synthetic images often mimic real radiographs with subtle artifacts — such as bilateral symmetry, uniform noise, or unnaturally smooth textures — the findings highlight an urgent need for expanded education and training so clinicians and AI systems can better recognize deepfakes and prevent misuse.

Read More