How AI in Medical Imaging Is Opening a Privacy Pandora’s Box

Medical imaging AI is delivering real diagnostic improvements — faster detection of tumors, better triage in emergencies, and reduced radiologist fatigue. But the same technology also makes it possible to create fake X‑rays and CT scans that look real enough to fool both human experts and automated systems.

In March 2026, researchers presenting at the Radiological Society of North America warned that deepfake medical images pose “unprecedented privacy risks.” The study showed that manipulated thoracic X‑rays could pass as authentic in blind tests with radiologists and AI detection tools. This is not a far‑off hypothetical. The methods exist and the data needed to train them often comes from real patient scans.

What happened

The RSNA study built chest X‑ray deepfakes using generative adversarial networks (GANs) and then asked radiologists and a popular AI screening model to distinguish them from real scans. Both groups failed a meaningful portion of the time. The fake images retained realistic anatomical variation and pathology patterns, making them nearly indistinguishable.

Why does this pose a privacy risk? Medical images themselves are highly identifiable. They contain facial features, unique bone structures, and internal anatomy. If someone can generate a convincing fake using your scan as a starting point, they could potentially:

  • Insert findings that never existed (for insurance fraud or to manipulate a diagnosis)
  • Leak synthetic images linked to your identity even if the originals were anonymised
  • Train other AI models on fabricated data that misrepresents patient populations

Why it matters for patients

Most patients never think about what happens to their imaging data after a scan. It is stored in PACS systems, often shared between hospitals and clinics, and increasingly fed into AI algorithms for research or clinical improvement. Those transfers rarely use end‑to‑end encryption. Once a scan is out, verifying its authenticity may become nearly impossible.

Traditional privacy rules like HIPAA do not yet account well for synthetic images. If a deepfake X‑ray is generated from your real scan, does that count as a breach? The legal answer is still unclear. Until regulations catch up, the burden of protection falls largely on providers and patients themselves.

What you can do right now

As a patient, you have more leverage than you might think. Here are concrete steps:

  • Ask about AI use. Before a scan, ask your provider whether AI tools are used during acquisition, interpretation, or storage. Most facilities will disclose this, and some are required to under internal consent policies.
  • Request encrypted image transfer. If your images need to be sent to another doctor, ask if they can be transferred using encrypted channels (e.g., HL7 FHIR with TLS). Avoid email attachments unless encrypted.
  • Monitor your medical records. After a scan, download your imaging report and, if possible, the image file itself. Compare it with later records. Discrepancies can be a sign of tampering or data mix‑ups.
  • Use patient portals with two‑factor authentication. Many health systems now offer portals to view your images. Enable 2FA to reduce the risk of unauthorised access.
  • Ask about data sharing. Some imaging studies are de‑identified and shared for research. You can often opt out. Request clarification on how your data is used and whether synthetic versions might be generated.

Providers and health IT teams can do more: adopt digital watermarking or cryptographic signatures for images, vet third‑party AI vendors for data handling, and pressure cloud vendors to support end‑to‑end encryption.

Future outlook

Standard‑setting bodies like RSNA and the American College of Radiology are already working on verification frameworks for medical images. These likely will involve cryptographic seals and provenance metadata baked into DICOM files. But broad adoption will take years.

For now, the safest approach is awareness. The same AI that helps diagnose a lung nodule could, if mismanaged, produce a fake scan attributed to you. That is not a reason to refuse imaging — it is a reason to ask questions and hold providers accountable.

Sources

  • RSNA study on deepfake X‑rays and their detectability (March 2026)
  • RSNA warning on privacy risks from AI‑generated medical images
  • Existing literature on de‑identification weaknesses in medical imaging (various, 2020–2025)
  • HIPAA compliance guidance on synthetic data (current status: incomplete)

Last updated: May 2026