Both radiologists and advanced AI models have difficulty distinguishing AI-generated X-ray images from real ones. This is according to an international study. The researchers warn that these so-called deepfakes pose serious risks to the reliability of medical imaging and the safety of healthcare processes.
Deepfakes are images that have been generated or manipulated using AI and are virtually indistinguishable from the real thing. In the study, 17 radiologists from six countries were asked to evaluate a total of 264 X-ray images, half of which were artificially generated. 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.
Four multimodal AI models, including GPT-4o, GPT-5, Gemini, and Llama, also performed inconsistently, with detection accuracies ranging from 52 to 89 percent. Notably, even the model that generated the images could not correctly identify all deepfakes. The study was recently published in Radiology.
Risks for diagnostics and cybersecurity
According to the researchers, this development creates new vulnerabilities. For instance, manipulated images could be used in fraud or legal claims, such as by presenting a nonexistent fracture as real. Additionally, there is a risk that malicious actors could inject synthetic images into hospital systems, with potential consequences for diagnoses and patient safety.
The study also shows that experience is no guarantee: the number of years of work experience among radiologists had no clear influence on their ability to recognize deepfakes. However, musculoskeletal specialists were found to be slightly better at detecting abnormalities.
‘Too perfect’ images as a warning sign
According to the researchers, deepfake X-ray images often exhibit subtle characteristics, such as unnaturally smooth bone structures, symmetrical lung fields, or “too neat” fractures. Yet these signals are difficult to recognize in practice.
To prevent misuse, the researchers advocate for technical and organizational measures. These include digital watermarks and cryptographic signatures added directly when images are created. Training healthcare professionals and developing detection tools are also essential.
According to the researchers, this is just the beginning: the next step is the emergence of deepfake 3D images, such as CT and MRI scans. Precisely for this reason, they argue, it is crucial to invest now in detection and awareness within the healthcare sector.
Legislative Proposal
Deepfakes are a growing phenomenon, including in the healthcare sector. This is particularly true regarding the use of the voices and appearances of (well-known) medical professionals, which are misused to spread fake news or incorrect medical information. Partly for this reason, the Netherlands is working hard on a new bill designed to better regulate the use of deepfakes (in Dutch).
The proposed law explicitly gives people the right to authorize or prohibit the use of their voice and likeness, thereby also giving healthcare professionals more control over their digital identity. In addition, the new law will make it easier to take legal action against deepfakes, both in civil and criminal courts. In this way, the proposal contributes to combating medical misinformation and maintaining trust in healthcare.
