If it's too good to be true, it probably isn't.
At the University of Aberdeen, researchers have discovered that ordinary people can be taught to perceive what machines have learned to conceal — not by cataloguing flaws, but by cultivating a felt sense of what makes a human face genuinely human. As AI-generated faces shed their obvious imperfections, the frontier of detection has moved inward, toward subtler perceptual instincts that training can awaken. The stakes are no longer merely academic: billions of dollars and the integrity of human identity itself hang in the balance of this quiet, escalating contest between creation and recognition.
- AI-generated faces have become so convincing that the old detection shortcuts — extra fingers, misaligned features — have been quietly engineered out of existence, leaving most people unable to tell the real from the invented.
- The human cost is already measurable: a single deepfake video call convinced a Hong Kong employee to wire £25 million to fraudsters, and fabricated AI personas have infiltrated the highest levels of political discourse.
- Researchers found that training people to notice six subtle perceptual qualities — symmetry, proportion, attractiveness, distinctiveness, expressiveness, and memorability — doubled detection accuracy, from 40% to 80%.
- Crucially, training also recalibrated participants' confidence, closing the dangerous gap between how certain people felt and how correct they actually were.
- The deepest irony looms over every advance: the published research that teaches humans to spot deepfakes becomes training data for the AI systems generating them, ensuring the arms race never truly ends.
Dr Clare Sutherland stands before two photographs at the University of Aberdeen. One shows a real person. The other is a machine's invention. The difference exists — but most people cannot find it.
Artificial intelligence has grown so fluent at generating human faces that the old tells have been engineered away. A sixth finger, an impossible shadow — these giveaways no longer reliably appear. What remains is subtler. Sutherland and colleagues across Australia, Canada, and the UK asked a practical question: can ordinary people be trained to detect what the eye no longer naturally catches?
The answer is yes — but only through a shift in approach. Rather than hunting for defects, participants learned to attune themselves to six perceptual qualities that AI faces tend to share. Real faces carry asymmetries, proportional variety, emotional expressiveness, and a kind of memorable distinctiveness that comes from lived experience. AI faces lean toward symmetry, toward the averaged and attractive, toward the generic. The training was less about rules than about developing a feel. Using thousands of synthetic faces generated by StyleGAN3, researchers found that accuracy climbed from roughly 40% to 80% — and participants' confidence became properly calibrated to their actual ability, correcting a previously dangerous overconfidence.
The urgency behind this work is concrete. Projected losses from deepfake fraud could reach £40 billion annually. A Hong Kong employee transferred £25 million after a video call from what appeared to be their boss — it was not. Russian intelligence once ran a fictitious LinkedIn persona that successfully infiltrated conversations with senior US officials.
Sutherland acknowledges legitimate uses for the technology — helping families imagine a missing child's appearance, supporting creative work in good faith. But the darker applications are multiplying faster than the defenses. And the final irony is inescapable: every paper published on how to catch deepfakes becomes data the machines can learn from. The research that arms humans also, in time, arms the systems they are trying to see through.
Dr Clare Sutherland holds up two photographs at the University of Aberdeen. One is a real person—an Australian academic. The other is not a person at all, but a machine's invention of one. The difference, she knows, is there to be found. Most people cannot find it.
Artificial intelligence has become so fluent at generating human faces that the old tells have stopped working. A sixth finger, a misaligned ear, an impossible shadow—these were once the giveaways. But AI learns from its mistakes, and fraudsters learn from AI. The obvious flaws have been engineered away. What remains is subtler, harder to name, and yet detectable if you know what to look for.
Sutherland and her colleagues—including Prof Amy Dawel, director of the Australian National University Emotions and Faces Lab, and researchers across Canada and the UK—set out to answer a practical question: Can ordinary people be trained to tell the difference? The answer, they found, is yes. But the training required a shift in approach. Instead of hunting for obvious defects, participants learned to attune themselves to six perceptual qualities that AI-generated faces tend to share. Real human faces carry asymmetries—a slightly drooping eyelid, a lopsided smile—that make us recognizably human. AI tends toward symmetry, toward the mathematically balanced. Real faces vary wildly in proportion; AI faces cluster toward the average. Real faces age, sag, show character. AI faces tend to be more attractive, more generic, less emotionally expressive, less memorable. The training was not about learning rules. It was about developing a feel.
The researchers created a pool of thousands of synthetic faces using StyleGAN3, one of the most sophisticated face-generation tools available. They tested participants before and after training, exposing them to both real and artificial images, then revealing which was which. The results were striking. Accuracy improved from roughly 40 percent to 80 percent—sometimes higher. A few participants achieved near-perfect scores. What happened in the human brain during this learning process mirrored, ironically, what happens inside the AI itself: exposure to data, repetition, gradual refinement of pattern recognition, improved performance without necessarily understanding the mechanism.
Confidence mattered too. Earlier research had shown that people were often overconfident in their ability to spot deepfakes, and the most confident were often the most wrong. After training, participants grew more confident—but now their confidence was calibrated to their actual ability. They knew when they were likely correct. That knowledge, Sutherland noted, is what allows people to act on what they've learned.
The stakes for this research are not abstract. Deloitte has projected that losses from AI deepfake fraud in the United States alone could reach £40 billion in the coming year, up from £12 billion in 2023. A concrete example: an employee at a Hong Kong firm received a video call from what appeared to be their boss. It was not. The deepfake was convincing enough that the employee transferred £25 million to the fraudsters. The technology has also been weaponized for political espionage. In 2019, an Associated Press investigation uncovered a LinkedIn profile belonging to a woman named Katie Jones, presented as a Russia and Eurasia specialist with connections to prominent Washington think tanks. The profile was fictitious—a deepfake created by Russian intelligence that had successfully infiltrated conversations with top US political officials.
There are legitimate uses for the technology. Sutherland acknowledges that AI can help families imagine how a missing child might look at different ages, or assist in creative work when used transparently and in good faith. But the darker applications are proliferating faster than the defenses against them. The final irony is this: the academic papers describing these training methods, once published, become data that AI systems can learn from. The arms race continues. The machines are reading the research on how to catch them.
Citas Notables
If you don't know when you're correct or not, you can't really do anything with that information.— Dr Clare Sutherland, University of Aberdeen
Training on visual artifacts like looking for a sixth finger has had limited success, partly because the AI is getting too good.— Prof Amy Dawel, Australian National University Emotions and Faces Lab
La Conversación del Hearth Otra perspectiva de la historia
Why does training work so well? What are people actually learning when they look at these images?
They're not learning a checklist. They're developing an intuition. It's like how you can recognize a friend's voice without being able to describe what makes it distinctive. The brain is picking up on patterns—the way AI tends to smooth things out, make them more average, more pleasant. After you've seen enough examples, you just feel it.
But couldn't the AI just learn those same patterns and fix them?
Absolutely. That's the problem. The moment this research is published, the AI systems read it and start adjusting. They can make faces less symmetrical, less attractive, more emotionally expressive. It's a genuine arms race.
So the training is temporary? It only works until the AI catches up?
For now, yes. But there's something else happening too. The human brain is flexible in a way machines aren't—at least not yet. We can adapt to new tricks faster than we might think. The real value is that people stop being passive. They start looking closely. They develop skepticism.
The Hong Kong case—£25 million transferred because of a video call. How does training prevent that?
It doesn't, not entirely. But it shifts the burden. If more people are trained to be suspicious, if more organizations require additional verification for large transfers, the fraud becomes harder to pull off. It's not a perfect defense. It's a friction point.
You mentioned AI is worse at non-white faces, younger faces, older faces. Why?
Because the training data is skewed. Most of the images AI learns from are young white people. So it's better at generating those faces convincingly. It's a bias baked into the technology itself.
What happens next? Does this research change anything?
It changes what's possible for individuals and organizations that take it seriously. But the real question is whether institutions will act on it—whether banks will require extra verification, whether platforms will invest in detection tools, whether people will actually pause and look closely before they trust what they see.