Radiologists Struggle to Identify AI-Generated X-rays
A new award-winning study reveals experts cannot reliably distinguish synthetic medical images from authentic radiographs.
Updated on Oct. 20, 2026 in Artificial Intelligence
Live Poll
Do you trust the ability of current medical professionals to detect AI-generated clinical images?
The Radiological Society of North America has awarded the 2026 Alexander R. Margulis Award to a study investigating the ability of radiologists and multimodal large language models (LLMs) to detect synthetic x-rays. The researchers found that experienced radiologists struggled to identify AI-generated radiographs, with mean identification accuracy ranging from 70% to 74.8%.
Why it matters
As medical imaging datasets increasingly incorporate synthetic content for training, the difficulty for experts to distinguish these images from real radiographs poses new challenges for clinical verification. This study highlights a critical vulnerability in diagnostic security as AI-generated medical imagery becomes more sophisticated.
In a study involving 17 radiologists across 12 centers, professionals achieved 74.8% mean accuracy in identifying GPT-4o-generated images and 70% for RoentGen-generated images. Notably, none of the four tested multimodal LLMs achieved perfect detection of synthetic content.
The players
Radiological Society of North America
A professional organization of radiologists and medical physicists that hosts international conferences and establishes standards for medical imaging.
Mickael Tordjman
A researcher who led the study on the detection of AI-generated medical radiographs.
Murat Yuce
A researcher who co-led the study evaluating radiologist and LLM detection of synthetic x-rays.
The details
Researchers led by Mickael Tordjman and Murat Yuce evaluated the performance of 17 radiologists and four multimodal LLMs—AI systems capable of processing and analyzing both text and visual inputs—against a mixed set of authentic and synthetic radiographs. The study found that years of professional clinical experience provided no advantage in identifying AI-generated content. Only seven of the 17 radiologists spontaneously identified the presence of synthetic images, suggesting a high potential for oversight in diagnostic environments.
Timeline
October 20, 2026: The study and award win were reported.
November 29, 2026: The award will be formally presented at the RSNA 2026 conference in Chicago.
The Tech Race
This research follows a growing trend of institutional focus on the reliability of synthetic data within the RSNA 2026 conference framework. It highlights the gap between current AI image synthesis capabilities and the diagnostic verification standards required for clinical integration.
Clinicians and medical imaging staff should remain aware that visual inspection alone may not reliably filter out AI-generated content in shared medical datasets. Future diagnostic workflows may require automated verification tools rather than sole reliance on human expertise.
The takeaway
The inability of both experts and current models to reliably detect synthetic x-rays suggests that radiology protocols must evolve to include rigorous metadata verification. Readers can track the upcoming RSNA 2026 presentations on November 29 for further details on mitigation strategies for synthetic imagery.
Further reading
For broader developments in how models impact specialized fields, visit /tech/artificial-intelligence/.
Source note: This article includes information reported by AuntMinnie.
Live Poll
Do you trust the ability of current medical professionals to detect AI-generated clinical images?
