Imagine a patient who arrives at her doctor’s clinic furious. She shows her doctor a video of him — white coat, plausible exam room, familiar cadence — endorsing an over-the-counter hormone supplement for menopausal symptoms, dismissing standard therapies as “pharma scams,” and offering a discount code.
But the physician never recorded that message. Someone built a deepfake from online recordings, including interviews, webinars, and patient-facing videos, and used the synthetic likeness to sell an unregulated product. This scenario is no longer hypothetical. Investigations have documented AI-generated videos impersonating specific clinicians whom they name to promote supplements and other dubious treatments on major platforms.
Medical misinformation is often treated as a content problem: debunk falsehoods, reduce amplification, pressure platforms to remove harmful posts. Deepfakes push this problem into new terrain. They undermine the credibility that makes digital care, ranging from telehealth visits to patient portals and even social media, possible. And they raise a basic question: What, and whom, can patients trust?
Call this underlying requirement ”epistemic security”: the degree to which clinicians and patients can believe that what they see, hear, and document is authentic enough to act on safely. In digital medicine, that trust rests on three layers: identity (who is speaking), record (what belongs in the chart), and evidence (what we accept as real findings). Deepfakes threaten all three.
First, identity. Deepfake “doctors” exploit the cues patients are taught to trust: a familiar face, a clinic logo, the tone of professional certainty. For patients, it becomes harder to know whether a video featuring their doctor is genuine advice or a persuasive imitation. For clinicians, a second hazard emerges: the “liar’s dividend.” When forgery is plausible, people who did say something reckless can deny authentic recordings as synthetic. The result is not only deception, but degraded accountability.
Second, the clinical record and workflow. Consider a voice-cloned “attending” calling overnight to change an opioid dose; a resident, hearing a trusted voice amid urgency and fatigue, complies. Or consider a telehealth encounter for a controlled substance: the patient presents with a synthetic face and a borrowed identity, and the documentation looks routine until a discrepancy triggers review. Deepfake methods also threaten diagnostic media. Researchers have demonstrated that deep learning can tamper with medical images while preserving a clinically plausible appearance. If images, recordings, or screenshots can be manipulated without leaving a trace, the electronic health record risks becoming less a faithful chronicle and more another contested digital artifact, undermining continuity, collaboration, and trust in documentation.
Third, the evidence base. Generative models can create synthetic datasets that, when responsibly governed, may protect privacy or support method development. But the same tools reduce the cost of fabrication. Paper mills already manufacture manuscripts and results at scale, and an investigation published in Nature has described how industrialized fraud and untrustworthy clinical trials already exist in the literature. A future in which bad actors generate a “trial,” complete with convincing tables, figures, and patient trajectories, should not be dismissed as fanciful. Even if fully synthetic randomized trials are not yet documented as being passed off as real-patient evidence, the incentives are obvious: publication, promotion, and profit.
The scale of this problem in routine clinical practice is not yet well measured. But in safety-critical systems, uncertain prevalence is not a reason for complacency. These attacks are cheap, scalable, and asymmetric: They can be created in minutes and may take hours to unwind, with consequences that spill across patients and institutions.
The response should not be framed solely as an arms race to detect every forgery. Detection matters, but adversarial systems evolve. What health care needs is a practical trust infrastructure, or norms and controls that make verification routine. The National Institute of Standards and Technology has emphasized layered approaches — provenance, watermarking, detection, and auditing — because no single method is sufficient.
- Establish trusted channels. Clinical instructions and results must be delivered through the patient portal or a verified phone tree, not screenshots, forwarded videos, or social-media clips.
- Require verification for high-risk requests. Medication changes, controlled substances, urgent orders, and major care-plan changes should trigger callback verification to a known number or a two-factor check, and the chart should record that it occurred.
- Treat provenance as patient safety. For high-stakes clinical media such as key diagnostic photographs, select imaging exports, clinician-recorded audio/video, retain a tamper-evident chain of custody showing when the file was created and how it was edited. Emerging standards such as the Coalition for Content Provenance and Authenticity (C2PA) “Content Credentials” point toward how provenance can travel with media.
- Create a clear escalation pathway. “Suspected synthetic media” should be a one-click report to information security and risk management, so clinicians are not forced to improvise in real time.
- Train for the conversation. Clinicians need scripts for patients who bring fabricated “doctor videos” into the exam room. Patients need to know where a practice does (and does not) publish advice, and they should be explicitly invited to ask, “Did you actually say this?”
None of this is frictionless. Callback verification, provenance-aware media handling, and new escalation pathways add steps to already burdened clinical workflows, and many health systems will not have the IT infrastructure to implement them seamlessly. But the alternative is to leave frontline clinicians improvising when authenticity is in doubt. The goal should not be perfect verification of every artifact on day one; it should be a pragmatic, risk-based standard for the highest-stakes communications and media.
Regulators and platforms have parallel responsibilities. Using a clinician’s name, image, or credentials in a synthetic endorsement without consent should be treated as a deceptive practice, with liability directed toward those who commission or distribute the content. The Federal Trade Commission’s authority over deceptive advertising offers a ready lever when deepfakes are used as marketing. Journals and regulators should also tighten disclosure expectations when data are synthetic or heavily augmented, and require clear descriptions of how generated data were validated against observed data before clinical claims are made.
Deepfakes shift the burden of trust from recognition to verification. Digital medicine runs on an infrastructure of authenticity that is easy to take for granted — until it fails. We should rebuild that infrastructure now, before patients arrive not only uncertain about what is true, but uncertain about whether their clinician is real.
DEEPFAKE RADIOLOGY
Deepfake images can mislead viewers, upend elections, and instigate violence. They can also, researchers say, disrupt medical care.
In a study published Tuesday in Radiology, an international team of researchers tested whether 17 radiologists could tell the difference between real X-rays and those generated by ChatGPT. Only 41% noticed that anything was awry when they were initially asked to diagnose patients based on the synthetic images. Even once they knew to look out for deepfake X-rays, they only differentiated them accurately 75% of the time.
