Rounds & Square Pegs

In January 2024, a finance worker in Hong Kong joined what looked like a routine video call with her company's chief financial officer. The face was familiar. The voice was familiar. The ask was ordinary: authorise a transfer. She did. Every person on that call had been fabricated. Roughly 25 million dollars left the building.

Fraud is not new. Manufacturing a credible messenger on demand is. That is the part that matters for health informatics, because clinical and public health systems still treat video and audio as stronger evidence than text. Patients trust a face they recognise. Staff forward a clip that "looks real." A chief medical officer's voice on a voicemail can move behaviour faster than a PDF policy ever will.

When the messenger is the payload

In Montreal, intensive care physician Dr. François Marquis found his face and voice used to sell fraudulent health products online. His first concern, he told reporters, was not his reputation. It was the patients who thought they were hearing him. A colleague, Dr. Alain Vadeboncoeur, was cloned into videos about conditions outside his specialty. Same pattern: steal the credibility, then sell the product or the opinion.

COVID-19 already showed how hard it is to keep trustworthy guidance visible when the information environment is flooded. The World Health Organization called that an "infodemic": too much information, true and false, competing for attention. Deepfakes do not only add volume. They forge the authority that used to cut through volume. Whoever can fabricate a trusted clinician, official, or journalist gets to borrow that trust without earning it.

If you work in health systems, this is not an abstract media story. It is an attack on the same trust channels your programmes depend on: clinician identity, institutional voice, emergency alerts, vaccine communication, and the patient who believes the face on the screen.

Harm that shows up in charts, not only in headlines

The fraud cases get the press. The clinical harm is quieter and often private.

Research on victims of deepfake-based image abuse documents depression, anxiety, and post-traumatic stress at rates comparable to survivors of physical assault. Knowing the imagery is fabricated does not cancel the injury. The dissonance is part of the injury: it looks completely real, and the person still has to live with it circulating.

Adolescents carry a disproportionate share. The American Academy of Pediatrics has described shame, withdrawal, and in severe cases self-harm among young people targeted by deepfake sexual content, with most never disclosing what happened. That lands in paediatric, primary care, and school health settings whether or not the clinician has language for "deepfake."

Clinicians are not insulated as bystanders either. Studies of pandemic-era misinformation among frontline workers, including work from Romania on COVID-related false news, found higher stress and insomnia among those who felt targeted or flooded by false claims, along with damaged trust in their own patient relationships. Fabricated clinician identities intensify that pressure: you can be both the supposed source and the person cleaning up the damage.

Equity is not a later chapter

Vaccine misinformation on social platforms during COVID-19 moved with ugly efficiency. Analyses of Facebook vaccine content found large shares of posts carrying false claims, with fact-checks making up a smaller share of the conversation, and a meaningful fraction of those fact-checks repeating the false claim in order to refute it. Deepfakes do not solve that architecture. They make it cheaper to attach a trusted face to whichever claim travels first.

The populations least equipped to detect a fabrication often already have the least institutional trust, frequently for good historical reason. Work on deepfakes in resource-limited settings documents gaps in detection tools, literacy programmes, and platform response. This is the same pattern health technology keeps repeating: the people most exposed to harm are least represented in building the defence.

If your programme serves communities that already treat official messaging with caution, a fabricated "doctor video" does more damage there than in a clinic where patients can walk down the hall and ask a real human.

Detection will lag. Design for that.

Detection research uses the same adversarial setup that generates the forgeries: train a discriminator to catch artifacts, then watch the generator improve until yesterday's detector fails. Tools that looked strong in 2024 were already weak against 2025 techniques. Treating "we will detect it" as the primary control is a bet against an arms race you do not control.

A useful Canadian framing of that problem sits in The Evolution of Disinformation: A Deepfake Future, the October 2023 report from an unclassified CSIS Academic Outreach workshop (World Watch: Expert Notes, Cat. No. PS74-19/2023E-PDF). It is not a formal CSIS position; it collects independent expert papers under Chatham House rules. The through-line that matters for health systems is blunt: deepfakes erode trust in visual and audio evidence itself, not only in particular claims, and mitigation that relies on detection alone will keep falling behind generation. The report also stresses human rights-centred responses, media literacy, and organisational protocols alongside any technical detector. That mix matches what clinics and public health shops can actually do.

Policy is uneven. Some jurisdictions require disclosure of AI-generated content and put liability on platforms; many do not. Media literacy still leans on habits built for a world where fabricating a convincing video took real skill and money. That world is gone. The replacement habits, verify the channel, distrust unexpected urgency, confirm identity out of band, are only partly built into health organisations.

Where this lands in informatics work specifically

Most health informatics students will not be asked to build a deepfake detector. Plenty will be asked to design patient portals, staff alert channels, telehealth consent flows, or public-facing campaign sites. Those surfaces are exactly where fabricated identity becomes an operational risk.

If your system sends appointment reminders by SMS with a clinic name in the from-line, ask how a patient would know a fake reminder is fake. If your organisation posts clinician videos for patient education, ask who owns the takedown path when a clone appears. If you are evaluating a vendor that wants to use synthetic patient voices in training data, ask what consent and disclosure look like when the synthetic voice could be mistaken for a real person.

The discipline habit is the same as with CDS: do not stop at "cool capability." Ask failure mode, who is exposed, and what the recovery path is when trust breaks.

What to do if this is your job

You do not need to become a deepfake forensic analyst. You do need a short operational posture.

Treat unexpected video or audio "from" a clinician, executive, or public health official as unverified until confirmed through a known channel. Phone the person. Use the EHR message. Use the official listserv. Do not reply inside the suspicious medium.
Assume emergency and vaccine communications will be impersonated. Build authentication into the workflow before the crisis, not during it: signed messages, known URLs, staff scripts for "how we actually send alerts."
When a patient's or clinician's likeness is abused, treat it as a clinical and safety issue, not only a PR issue. Document, escalate, and connect people to support. Do not minimise because "it was fake."
In any AI or communications project, ask who is most exposed if a fabricated message lands in their feed, and whether those communities were in the room when you designed the response.

Deepfakes will keep getting cheaper. Trust will not. The informatics work is to stop treating "I saw it" and "I heard it" as enough, and to build verification into systems that still assume a face is a credential.

Deepfakes are a health systems problem, not just a fraud problem