Generative artificial intelligence (AI) can now produce highly realistic x-ray images that experienced radiologists cannot distinguish from authentic scans. A study published in Radiology reported that neither clinicians nor multimodal AI models reliably identified these deepfakes. Researchers have warned that this introduces a risk for diagnostic errors and potential manipulation of clinical data and called for a new approach to AI use in radiology.

Generative AI refers to artificial intelligence that can create original content, including text, images, music, and videos. It is based on large datasets and learns patterns to generate something new. Well-known examples include chatbots such as ChatGPT and image generators such as DALL-E.

The rapid advancement of generative AI has transformed the field of medical imaging. Multimodal models can now generate radiologic images from text descriptions alone without programming expertise. Researchers have reported that these images appear anatomically plausible, diagnostic, and visually indistinguishable from real scans.

Detection Limits

The study included 17 radiologists from six countries who evaluated 264 x-ray images, half authentic and half AI-generated images. ChatGPT was used to produce the AI images.

In the initial evaluation, participants were not informed that the dataset included AI-generated radiographic deepfakes. Only 41% of the deepfake images were identified correctly. Detection rose to approximately 75% after participants were explicitly told that some images were AI-generated. Clinical experience and radiology expertise did not significantly improve detection accuracy.

Multimodal AI systems have similar limitations. None of the evaluated models identified all deepfakes. Even the generative AI model that produced the images did not consistently recognize its own radiographs.

"Our study shows that these deepfake x-rays are realistic enough to deceive radiologists — even when they knew that AI-generated images were included," said lead author Mickael Tordjman, MD, MS, BioMedical Engineering and Imaging Institute, Icahn School of Medicine at Mount Sinai in New York. This creates "a high-risk vulnerability," particularly if "a fabricated fracture could be indistinguishable from a real one."

Subtle Clues

Despite their realism, deepfake radiographs showed recurring anomalies. These include unusual symmetry, overly smooth bone surfaces, uniform noise patterns, and unnatural soft-tissue textures.

Deepfake medical images often look too perfect,” Tordjman said. “Bones are overly smooth, spines unnaturally straight, lungs overly symmetrical, blood vessel patterns excessively uniform, and fractures appear unusually clean and consistent, often limited to one side of the bone." These features may serve as diagnostic warning signs.

Security Risks

Deepfakes in medical imaging raise serious concerns because they can be misused in clinical, legal, and research settings. The authors outlined several plausible scenarios in which such manipulations could be exploited for malicious purposes.

Because AI models can now generate medical images without specialized expertise, the barrier to entry has decreased significantly. This increases the risk for targeted forgery beyond conventional cyberattacks.

Therefore, the researchers proposed a multilayered approach that includes clinician education, automated deepfake detection systems, mandatory watermarking, and rigorous dataset governance to prevent this emerging capability from evolving into a systemic threat. Training datasets should also include deepfake examples to alert radiologists when necessary.

Conclusion: Turning Point

According to the authors, this study marks a turning point in radiological diagnostics. AI-generated deepfake radiographs are now so realistic that they can fool both human experts and AI systems. The accompanying editorial in Radiology, “The Democratization of Deceit: Seeing Is No Longer Believing,” reflected the same concern.

This shifts the challenge for physicians from image interpretation to the verification of authenticity. Without technical safeguards, training, and clear standards, the increasing availability of generative models may undermine trust in radiological imaging over time and lead to potentially dangerous situations.

This story was translated from Medscape’s German edition.