Your X-Ray Could Be a Privacy Risk: What You Need to Know About AI in Medical Imaging

Artificial intelligence is increasingly used to help radiologists read X-rays, MRIs, and CT scans. AI can spot patterns humans miss, speed up diagnoses, and reduce waiting times. But as AI systems are trained on massive collections of medical images, a set of privacy risks has emerged that patients should understand before they next step into an imaging center.

What Happened

At recent conferences of the Radiological Society of North America (RSNA), researchers have presented evidence of several unsettling scenarios involving medical imaging AI.

One study showed that deepfake X-rays can fool both radiologists and the AI systems designed to detect them. Researchers created realistic but synthetic chest X-rays by modifying existing scans. When presented to human experts and AI tools, the fakes were frequently mistaken for real images. That raises the possibility of manipulated scans being injected into patient records or used to trick automated screening systems.

Separately, many medical imaging datasets used to train AI models have been shared between institutions—and even with commercial AI developers—without explicit patient consent. Data anonymization is supposed to strip identifying details (name, date of birth, etc.), but researchers have demonstrated that re-identification is often possible by matching scan metadata, facial contours (in head CTs), or unique anatomical features with other data sources.

Why It Matters

Three privacy concerns are central.

Data breaches. Hospitals and cloud storage providers hold terabytes of imaging data. When that data is shared for AI development, the attack surface widens. A breach could expose not just scans but linked health records, insurance information, and personal identifiers.

Deepfake scans. Beyond the threat of fraud, manipulated images could cause misdiagnosis. If a fake X-ray is inserted into a patient’s file, a later comparison might show a “change” that doesn’t exist, leading to unnecessary procedures. Conversely, concealing a real finding could delay treatment.

Re-identification. Even when scans are formally de-identified, the risk remains that someone can link a scan back to you. A 2019 study found that facial recognition software could match 3D head CT reconstructions to individuals with high accuracy, undermining the whole purpose of anonymization.

The stakes are high because medical images are permanent, detailed, and difficult to delete. Once your scan is used to train an AI model, you lose control over how that model might be deployed.

What Readers Can Do

You have more influence than you might think. Here are practical steps.

Ask your provider about AI use. Before an imaging exam, ask: “Will AI be used to analyze my scan? Is my scan used to train AI models? Can I opt out?” Many facilities have policies but don’t always volunteer them.

Opt out of AI training if you can. The American College of Radiology and some hospital systems now allow patients to decline research participation, including AI training, while still receiving clinical care. You may need to request a specific form.

Review privacy notices. Look for how your hospital or imaging center handles data. The Notice of Privacy Practices (required under HIPAA) should describe how your health information is used and shared. If it says “de-identified data may be used for research,” that blanket language covers AI training too.

Check if your images are stored in the cloud. Some radiology groups use third-party AI vendors who process images on remote servers. Ask where the data is stored, who has access, and what happens after the AI analysis.

Know your HIPAA rights. HIPAA gives you the right to access your medical imaging records, request an accounting of disclosures, and file a complaint if you believe your data was used improperly. But HIPAA’s privacy rule does not fully cover de-identified data or data used solely for AI algorithm development, so it’s not a complete safety net.

Sources

  • Radiological Society of North America (RSNA) conference presentations on deepfake X-rays and AI privacy, 2025–2026.
  • RSNA news article: “Deepfake X-Rays Fool Radiologists and AI,” March 2026.
  • Research on re-identification of medical imaging (e.g., Schwartz et al., Nature Communications, 2019).
  • HIPAA Privacy Rule and OCR guidance on de-identification.

AI in medical imaging holds real promise. But the privacy landscape is shifting faster than regulations can keep up. Being an informed patient is your best safeguard.