par
Gus Iversen, Editor in Chief | March 27, 2026
A new peer-reviewed study published in Radiology reports that both radiologists and advanced artificial intelligence models struggle to reliably distinguish between authentic and AI-generated X-ray images, raising concerns about clinical integrity and cybersecurity.
The research, led by investigators at the Icahn School of Medicine at Mount Sinai in New York, evaluated 17 radiologists from 12 centers across six countries. Participants reviewed 264 images, half of which were synthetic. The data set included images generated by ChatGPT-based systems as well as RoentGen, a diffusion model developed by Stanford Medicine.
When radiologists were not informed that synthetic images were included, only 41% identified them without prompting. After disclosure, their average accuracy rose to 75%, with individual performance ranging from 58% to 92%. Experience level did not correlate with detection accuracy, although musculoskeletal specialists performed better than other subspecialties.
Ad Statistics
Times Displayed: 357
Times Visited: 1 Stay up to date with the latest training to fix, troubleshoot, and maintain your critical care devices. GE HealthCare offers multiple training formats to empower teams and expand knowledge, saving you time and money.
Multimodal large language models showed comparable variability. GPT-4o, GPT-5, Gemini 2.5 Pro and Llama 4 Maverick achieved detection rates between 57% and 85%. Notably, even the model used to generate some of the images was unable to consistently identify them.
Lead author Dr. Mickael Tordjman, a postdoctoral fellow, said the findings point to potential misuse. “This creates a high-stakes vulnerability for fraudulent litigation if, for example, a fabricated fracture could be indistinguishable from a real one,” he said. He also warned of cybersecurity risks if manipulated images were introduced into clinical systems.
The study identified recurring visual patterns in synthetic images, including overly smooth bones, symmetrical lung fields and unusually uniform vascular structures.
The authors recommend technical safeguards such as embedded watermarks and cryptographic signatures at the point of image capture. They also call for expanded training data sets and detection tools as generative models advance toward more complex imaging modalities like CT and MR.
