“AI-generated deepfake X-ray images are sufficiently realistic to cause radiologists to make incorrect diagnoses.”

Neither radiologists nor multimodal LLMs could easily distinguish AI-generated deepfake X-ray images from authentic ones. Radiologists' accuracy was 75% in detecting ChatGPT-generated images (range 58-92%) and 62-78% for RoentGen chest X-rays, even when aware; no correlation with experience except musculoskeletal subspecialists performed better. Lead author: 'These deepfake X-rays are realistic enough to deceive radiologists.'

Although synthetic data address crucial shortages of real-world training data, their overuse might propagate biases, accelerate model degradation, and compromise generalisability across populations. A concerning consequence of the rapid adoption of synthetic data in medical AI is the emergence of synthetic trust—an unwarranted confidence in models trained on artificially generated datasets that fail to preserve clinical validity or demographic realities.

The generation of synthetic medical images poses a unique set of challenges compared to other domains due to the intricate nature of biological structures and the subtle nuances of imaging biomarkers. Without faithful representation of these biomarkers, synthetic images risk being of limited utility in clinical settings, hindering their adoption for tasks such as training AI models, augmenting datasets, and validating imaging algorithms.

As proof of principle, Mirsky et al [3] showed that they were able to tamper with CT scans and artificially inject or remove lung cancers on the images. When the radiologists were blinded to the attack, this hack had a 99.2% success rate for cancer injection and a 95.8% success rate for cancer removal. Even when the radiologists were warned about the attack, the success of cancer injection decreased to 70%, but the cancer removal success rate remained high at 90%. This illustrates the sophistication and realistic appearance of such artificial images.

A study in Radiology found realistic AI-generated X-rays not easily distinguished by radiologists (moderate performance) or LLMs. Diagnostic accuracy was 91.3% for authentic and 92.4% for AI-generated radiographs, showing AI images sufficiently realistic to maintain high diagnostic reliability. Only 41% of radiologists spontaneously identified AI images when blinded.

We quantify X-ray clinical realism by asking radiologists to distinguish between real and fake scans and find that generates are more likely to be classed as real than by chance, but there is still progress required to achieve true realism. We confirm these findings by evaluating synthetic classification model performance on real scans.

17 radiologists could only differentiate real X-rays from ChatGPT-generated ones 75% accurately even when alerted; only 41% noticed issues blindly while diagnosing. Images generated easily with simple prompts, fooling even the generating model (57-85% detection accuracy), disrupting medical care potential.

“Our study demonstrates that these deepfake X-rays are realistic enough to deceive radiologists, the most highly trained medical image specialists,” the study's lead author Dr. Mickael Tordjman said, “even when they were aware that AI-generated images were present.” Radiologists who were made aware of the fact that these datasets contained AI images fared better than those exposed to the images without any indication of the test's actual purpose, but still not great, showing a mean accuracy of 75%.

AI algorithms can analyse medical images with remarkable accuracy and speed, often surpassing human capabilities, by identifying subtle patterns and anomalies that may be missed by the human eye. This can improve diagnostic accuracy and reduce the risk of misdiagnosis and false negatives. However, the use of AI in healthcare raises ethical questions, and AI algorithms require high-quality, labelled data to train effectively, with biased datasets potentially leading to incorrect or discriminatory outcomes.

A multi-center international study reveals that neither experienced radiologists nor advanced multimodal large language models (LLMs) can reliably distinguish “deepfake” X-rays from authentic ones. Even when warned that synthetic images were present, radiologists only averaged 75% accuracy in identifying them. The findings expose a high-stakes vulnerability in healthcare, ranging from fraudulent litigation (fabricated injuries) to cybersecurity threats where hackers could inject synthetic images into digital medical records to cause clinical chaos.

Both radiologists and advanced AI models have difficulty distinguishing AI-generated X-ray images from real ones. When they did not know that deepfakes were included, only 41 percent recognized them spontaneously. After an explicit warning, the average accuracy rose to 75 percent, with significant variation among individual evaluators. The researchers warn that these so-called deepfakes pose serious risks to the reliability of medical imaging and the safety of healthcare processes. For instance, manipulated images could be used in fraud or legal claims, such as by presenting a nonexistent fracture as real.

A new study published on March 24 in Radiology, the journal of the Radiological Society of North America (RSNA), shows that both radiologists and multimodal large language models (LLMs) have difficulty telling real X-rays apart from artificial intelligence (AI)-generated "deepfake" images. Once they were informed that synthetic images were present, their average accuracy in distinguishing real from fake rose to 75%. Performance varied widely among individuals. Radiologists correctly identified between 58% and 92% of the ChatGPT-generated images.

X-ray images generated by Artificial Intelligence (AI) are now so good they can fool expert radiologists, according to a new study. The findings, published in the journal Radiology come amid a huge boom in AI use in healthcare and raise concerns that these “deepfake” X-rays could be used to deceive insurance companies, employers and even interfere with legal cases. The study asked 17 radiologists from six different countries to assess X-ray images of different parts of the body.

Synthetic radiographs, or AI-generated X-ray images, are becoming increasingly realistic, raising questions about their detectability in clinical practice. While they can be valuable for training or research, their high realism can occasionally blur the line between genuine and artificial images, potentially affecting diagnostic accuracy if unnoticed. When unaware of the study's true purpose, 41% of radiologists spontaneously suspected some images to be AI-generated. After being informed, overall detection accuracy was 75% for GPT-4o radiographs and 70% for those produced by RoentGen.

Prior to 2026, smaller studies (e.g., 2023-2024 GAN-based fakes) showed radiologists detecting ~80-90% but with lower realism; 2026 Radiology study marks advance in diffusion/LLM realism, yet explicitly notes comparable diagnostic accuracy (91.3% vs 92.4%), implying potential but not measuring actual incorrect diagnoses caused.