Rounds & Square Pegs

As deepfake technology makes it impossible to trust what you see and hear, the implications extend well past fraud, into mental health, emergency response, and the basic architecture of public health communication.

A familiar scam, an unfamiliar mechanism

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 reassuring, the request straightforward: authorise a 25 million dollar transfer. She complied. Every face on that call, every voice, had been fabricated. The money was gone.

The fraud itself is not new. What is new is the mechanism: artificial intelligence manufacturing a credible spokesperson to deliver the request. That shift, from spreading false information to manufacturing the messenger, is the part that should concern anyone working in health.

When the messenger is fabricated

Dr. François Marquis, an intensive care physician in Montreal, discovered his face and voice being used to sell fraudulent health products online. His first concern was not his own reputation. "My primary worry is for the individuals who trust me," he told reporters, noting the damage this does to how patients relate to physicians generally.

A colleague, Dr. Alain Vadeboncoeur, found himself digitally cloned across videos discussing conditions entirely outside his specialty. Neither case is isolated. Both follow the same logic: fabricate the credibility, then use it to move product or move opinion. The World Health Organization warned of an "infodemic" during COVID-19, too much information, true and false, making trustworthy guidance hard to find. Deepfakes are a different problem. They do not just add noise. They manufacture the authority that cuts through it, for whoever controls the fabrication.

The mental health cost is not hypothetical

Clinical research on victims of deepfake-based image abuse documents rates of depression, anxiety, and post-traumatic stress comparable to survivors of physical assault. The psychological injury is not diminished by the imagery being fabricated. If anything, the dissonance, knowing it is fake while it looks completely real, adds its own layer of harm.

Young people carry a disproportionate share of this. The American Academy of Pediatrics reports that adolescent victims of deepfake-based sexual content experience shame, withdrawal, and in severe cases self-harm, and that most never disclose what happened to anyone. Healthcare workers are not exempt either: a study of Romanian frontline clinicians during COVID-19 found that those affected by false news reported significantly higher stress and insomnia than colleagues who felt insulated from it, along with damaged trust with their own patients.

Why the equity question cannot wait

Vaccine misinformation during COVID-19 spread with real efficiency: one analysis found roughly 46.6 percent of vaccine-related Facebook content contained misinformation, while fact-checks made up under half of the conversation, and over a quarter of those fact-checks repeated the false claim they were correcting. Deepfakes do not fix this problem. They accelerate it, by lending manufactured authority to whichever side deploys them first.

The populations least equipped to detect a fabrication are the same ones with the least institutional trust to begin with, often for good historical reason. Research on deepfakes in resource-limited settings documents real gaps in detection tools and awareness. This is the pattern every wave of health technology repeats: the people most likely to be harmed are the least likely to be included in building the defence.

Building a defence that actually holds

Detection research borrows the same architecture that creates the problem. Generative adversarial networks pit a generator against a discriminator; the same adversarial logic can train detection systems to catch the artifacts a fabrication leaves behind. But detection tools documented as effective in 2024 were already outdated against 2025's techniques. This arms race will likely never resolve in detection's favour on a fixed timeline.

Policy is catching up unevenly. Some jurisdictions now require disclosure of AI-generated content and establish platform liability; many do not. Media literacy needs an update too: "seeing is believing" assumed a world where fabrication took real skill. That assumption no longer holds, and the verification habits that should replace it are mostly still unbuilt.

What this means for you