A landmark study found radiologists can only detect AI-generated X-rays 41% of the time when unaware fakes exist — barely better than random chance.
Six rounds. Two chest X-rays. One is AI-generated. Pick the fake — then see what gave it away.
Which scan is AI-generated?
Click to select — choose carefully.
Published in Radiology, March 24, 2026. Seventeen radiologists were shown a mix of real and AI-generated chest X-rays and asked to judge authenticity.
"Years of experience showed zero correlation with the ability to detect AI-generated scans. The most experienced radiologists were just as likely to be fooled."
— Radiology, March 2026
Source: Radiology (RSNA), March 2026 · Nature · STAT News · RSNA
When a radiologist cannot reliably tell real from synthetic, three threat vectors become newly credible.
A fabricated fracture enters evidence. The jury cannot tell it is fake. Neither can the expert witness. The case settles for seven figures.
Parallel: Deepfake audio in contract disputes was first successfully used as evidence in 2024, setting precedent for synthetic media in courts.
Synthetic scans submitted for claims on injuries that never happened. Detection hinges on radiologists who the study shows cannot reliably distinguish real from fake.
Parallel: US insurance fraud totals an estimated $308 billion annually. Medical imaging fraud is already a $30B+ subset.
An attacker injects fake tumors into a patient's scan stored in a PACS system. Treatment decisions change. The patient receives unnecessary surgery — or necessary surgery is delayed.
Parallel: Israeli researchers demonstrated live PACS injection of cancer nodules in 2019. AI generation makes this dramatically more accessible.
Medical deepfakes did not emerge overnight. Six years of escalating capability.
Ben-Gurion University demonstrates live injection of AI-generated cancer nodules into CT scans via PACS network compromise.
GAN-based models generate convincing chest X-rays indistinguishable from real images in controlled research settings. Published in top-tier radiology journals.
AI-generated pathology images fool peer reviewers at two major journals. Retraction notices issued after detection by image forensics experts — not radiologists.
Leading hospital networks begin piloting digital watermarking at point of image capture. FDA issues draft guidance on medical imaging integrity requirements.
Radiology study: 17 radiologists, 12 centers, 6 countries. Detection rate: 41% when unaware. 75% when forewarned. Experience is no protection.
Proposed paths forward: blockchain provenance chains, embedded cryptographic signatures, AI-based detectors, mandatory FDA certification for medical imaging systems.
Four proposed solutions — each at a different maturity level.
Invisible cryptographic signatures embedded at the moment of image capture. If the watermark is absent or altered, the image is flagged. Pilot programs are underway at major hospital networks, but retrofitting existing PACS infrastructure is expensive. Does not prevent injection into systems that lack watermark verification.
An immutable chain-of-custody record for every medical image from capture to diagnosis. In theory, any modified image would break the chain. In practice, healthcare IT adoption of blockchain has been extremely slow, interoperability across different PACS vendors remains unsolved, and the technology adds latency to time-sensitive workflows.
Use AI models to detect AI-generated images — the study's four AI models (GPT-5 at 85%, GPT-4o ~70%, Gemini 2.5 Pro ~65%, Llama 4 Maverick at 57%) outperformed radiologists when unaware. But generative models improve faster than detectors. An adversarial arms race favors offense. Any public detector becomes a training signal for better generators.
FDA guidance requiring cryptographic authentication for medical imaging submissions. Hospital accreditation standards mandating watermark verification. Legal liability for systems that process unauthenticated scans in clinical contexts. The regulatory approach is slow but creates systemic incentives that technical solutions alone cannot achieve. Legislation is pending in the EU and US as of early 2026.