AI-Generated Deepfake X-Rays Deceive Radiologists and AI Systems
A multi-center international study found that radiologists and advanced AI systems cannot reliably distinguish AI-generated "deepfake" X-rays from authentic ones. When unaware synthetic images were present, only 41% of radiologists spontaneously noticed them, and even when warned, they misidentified images 25-30% of the time. The research highlights vulnerabilities for fraudulent litigation and cybersecurity threats in healthcare.
What happened
A multi-center international study tested 17 radiologists across six countries to evaluate their ability to distinguish AI-generated X-rays from authentic ones. The radiologists were shown images generated by ChatGPT and RoentGen alongside real X-rays. When unaware that synthetic images were present, only 41% of radiologists spontaneously noticed anything unusual about the deepfake images.
Even when informed that AI-generated images were included in the test, radiologists achieved only 75% accuracy on average and failed to identify the fakes 25-30% of the time across multiple attempts. The deepfake X-rays featured unnatural characteristics including overly straight spines, abnormally smooth bones, and unusually clean fractures.
Reported impact
The study authors identified potential vulnerabilities for fraudulent litigation, noting that fabricated injuries represented in synthetic images could be indistinguishable from real fractures. They also warned of cybersecurity threats where hackers could inject synthetic images into digital medical records to cause clinical disruption.
Index analysis
Neither radiologist experience level (ranging from 0 to 40 years) nor AI systems reliably detected the deepfakes. GPT-4o, the model used to create the deepfakes, could not accurately detect all of its own fabrications, while Google's Gemini and Meta's Llama models performed worse. The study revealed that current multimodal large language models lack the capability to consistently identify synthetic radiographs.
Governance controls
The authors called for a multilayered response including clinician education, automated deepfake detection systems, mandatory watermarking of synthetic images, and rigorous dataset governance. The broad public availability of tools to generate realistic X-rays, achieved through the democratization of large language models accessible via the internet, highlighted the urgent need for these protective measures.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.