If it's too good to be true, it probably isn't.
In laboratories and living rooms alike, humanity is confronting a quiet crisis of perception: the faces we see may no longer belong to anyone who exists. Researchers at the University of Aberdeen have discovered that ordinary people can be trained to distinguish AI-generated faces from real ones with 80% accuracy — not by hunting for obvious flaws, but by cultivating a subtler intuition about symmetry, distinctiveness, and the ineffable quality of being human. The stakes are not merely philosophical; deepfake fraud has already cost billions and compromised national security, and the gap between what machines can fabricate and what eyes can detect is narrowing with each passing model. This is the oldest of human struggles rendered in pixels: the effort to know what is real before reality itself becomes negotiable.
- AI-generated faces have grown so convincing that the familiar red flags — distorted ears, extra fingers, visual glitches — have effectively vanished, leaving people without reliable instincts for detection.
- The consequences are already severe: a £25 million fraud executed via deepfake video call, a fictitious AI-generated persona used by Russian intelligence to infiltrate American policymaking circles, and projected fraud losses of £40 billion in the US alone.
- University of Aberdeen researchers found that training people to notice six subtle perceptual qualities — symmetry, proportionality, attractiveness, distinctiveness, expressiveness, and memorability — can lift detection accuracy from 40% to 80% within a single hour.
- A troubling pre-training pattern emerged: the most confident people were often the most wrong, suggesting that overconfidence in our own perception may be as dangerous as the fakes themselves.
- The arms race is self-defeating by design — AI systems are already absorbing the academic research that describes how to spot them, meaning the detection window that training opens may be closing even as researchers publish their findings.
In a University of Aberdeen laboratory, psychologist Dr Clare Sutherland places two photographs side by side — one a real person, one a machine's invention. To the untrained eye, they are nearly identical.
The old tells have disappeared. AI learned from its early mistakes, and so did the fraudsters who exploit it. Deepfake scams cost the United States roughly £12 billion in 2023, with projections reaching £40 billion by next year. In one documented case, an employee transferred £25 million to criminals after a video call with a deepfake of their own boss. Russian intelligence, meanwhile, built a fictitious LinkedIn persona complete with an AI-generated face to infiltrate circles of American national security officials.
Sutherland and colleague Prof Amy Dawel set out to discover whether ordinary people could be taught to see what machines were doing. Working with thousands of AI-generated faces, they identified six perceptual qualities that AI handles differently than nature does: symmetry, proportionality, attractiveness, distinctiveness, expressiveness, and memorability. Real faces carry small asymmetries and imperfections. AI faces tend toward aesthetic polish and generic averageness — pleasant, balanced, and oddly forgettable.
There is no single magic tell. The training works by building intuition. Participants exposed to both real and AI-generated faces, then told which was which, improved from roughly 40% to 80% accuracy within an hour. The human brain, it turns out, learns through pattern recognition in ways not entirely unlike the generative AI it is trying to unmask.
One unexpected finding: before training, the most confident participants were often the most wrong. After training, confidence rose — but this time it tracked actual performance. The research also surfaced a structural bias: AI performs worse on non-white, older, and younger faces, reflecting the skewed demographics of its training data.
The window for human detection may be narrowing. AI models have likely already absorbed the academic papers describing how to spot them. For now, reality and fabrication remain distinguishable — but the machines are reading the instruction manual, and they are learning fast.
In a laboratory at the University of Aberdeen, psychologist Dr Clare Sutherland holds up two photographs side by side. 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, to the untrained eye, is nearly invisible.
Artificial intelligence has become so skilled at generating human faces that the old tells—the extra finger, the warped ear, the obvious glitch—have largely disappeared. The machines learned from their mistakes. Fraudsters learned too. This poses a problem that extends far beyond academic curiosity. In 2023, AI deepfake scams cost the United States roughly £12 billion. Deloitte projects that figure will climb to £40 billion by next year. In one documented case, an employee at a Hong Kong firm transferred £25 million to criminals after a video call with a deepfake version of their boss. The technology has also been weaponized for political espionage: Russian intelligence, according to an Associated Press investigation, created a fictitious LinkedIn profile of a woman named Katie Jones, complete with AI-generated photograph, and used it to infiltrate circles of prominent American policymakers and national security officials.
Sutherland and her colleague Prof Amy Dawel, director of the Australian National University Emotions and Faces Lab, began asking whether ordinary people could be taught to see what machines were doing. The answer, they found through experiments involving thousands of AI-generated faces created with StyleGAN3, is yes—but the method requires abandoning the search for obvious flaws. Instead, the researchers identified six perceptual qualities that AI tends to handle differently than nature does. Symmetry: real faces carry asymmetries, a drooping eyelid or a lopsided smile, the small imperfections that make us human. AI tends toward balance. Proportionality: very large noses or protruding ears appear less frequently in machine-generated faces. Attractiveness: AI faces, Sutherland explains, tend to be more pleasant to look at—aesthetically polished in ways that real faces rarely are. Distinctiveness: AI faces cluster toward the average, looking generic rather than memorable. Expressiveness: they show less emotional range. Memorability: they fade from memory more easily than real faces do.
These qualities sound fuzzy because they are. There is no single definitive tell, no magic detail that unmasks a fake. Instead, the training works by building intuition. Researchers exposed participants to both real and AI-generated faces, then told them which was which. Within an hour, accuracy improved dramatically—from roughly 40 percent to 80 percent. Some participants achieved near-perfect scores. The human brain, it turns out, learns in a way not entirely unlike how generative AI itself learns: given enough data, pattern recognition improves, even when the person doing the recognizing cannot fully articulate what they are seeing.
One unexpected finding emerged from the confidence data. Before training, people were overconfident in their ability to spot fakes, and the most confident were often the most wrong. After training, confidence increased—but now it was calibrated to actual performance. Sutherland notes this matters. If you do not know when you are right, you cannot act on the information. The research also revealed a limitation: AI performs worse on non-white faces, older faces, and younger faces, because its training data skewed heavily toward young white people.
The implications ripple outward. In Australia, a politician is proposing that AI-generated political content be watermarked and disclosed. Sutherland acknowledges that the technology has legitimate uses—showing how a missing child might age, enabling creative work when people know AI was involved and consent to it. But the arms race is accelerating. AI models have likely already absorbed the academic papers describing how to spot them. They are learning. The window for human detection may be narrowing. For now, though, the good news is simple: we are not yet living in a world where reality and fabrication are indistinguishable. The bad news is that we are moving toward one, and the machines are reading the instruction manual.
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 or odd earrings, has had limited success, partly because the AI is getting too good.— Prof Amy Dawel, Australian National University
La Conversación del Hearth Otra perspectiva de la historia
So people can actually learn to spot these things? It seems like it would be impossible.
It's not impossible, but it requires a different kind of attention. You're not looking for a smoking gun—a sixth finger or something obviously wrong. You're learning to feel the difference.
Feel the difference? That sounds vague.
It is vague, intentionally. The researchers identified six qualities—symmetry, attractiveness, distinctiveness, and others—that AI handles differently than human faces do. But none of them alone is definitive. It's the accumulation.
And people can actually improve at this in an hour?
Yes. Accuracy jumps from about 40 percent to 80 percent just through exposure and feedback. Some people get close to perfect. It's the same way the AI itself learns—give it enough data and it gets better, even if nobody fully understands the mechanism.
But if AI is reading these research papers, doesn't that mean the machines will just get better at fooling us?
Almost certainly. That's the real problem. We're in an arms race, and the other side gets to read our playbook.