What happens when AI detectors fail? Researchers say we need to be trained to recognize fake AI faces
Artificial intelligence has become amazingly good at creating fake human faces. It’s good, in fact, that the old tricks people used to rely on – counting fingers, spotting twisted earrings, or looking for a distorted background – are quickly becoming obsolete. According to new research highlighted by the BBC, the next line of defense may not be the best AI detector at all. It may be someone who is better trained.
Researchers at the University of Aberdeen, working closely with the Australian National University, have found that people can significantly improve their ability to distinguish AI-generated faces from real ones after a relatively short period of systematic training. Instead of hunting for obvious glitches, participants are taught to see subtle patterns that modern image producers still struggle to replicate consistently.
The AI race forces humans to evolve as well
For many years, identifying AI-generated images felt trivial. Early models often produced six fingers, mismatched earrings or impossible shadows. But today’s generators, powered by systems like StyleGAN3 and new distribution models, have gone far beyond those chattering errors. As a result, researchers argue that relying on visual defects is no longer an effective strategy.
Instead, participants were trained to judge six cognitive attributes that AI commonly shares. These include unusually perfect facial proportions, over-proportioned features, above-average attractiveness, normal-looking facial structures, limited emotional expression, and faces that are surprisingly hard to remember after you look away.
The results were impressive. Before training, participants correctly identified AI-generated faces only about 40 percent of the time. After about an hour of guided learning and repeated exposure to both real and artificial faces, accuracy increased to about 80 percent. A number of participants even achieved perfect detection scores. More importantly, their self-esteem was better aligned with their actual performance, something previous research suggested was often lacking.
Why seeing AI matters more than ever
This is no longer just an academic exercise. Deepfake technology is already being used in financial fraud, political influence campaigns and internet identity scams. The BBC points to Deloitte estimates that suggest losses from deep AI-enabled fraud in the United States could rise to £40 billion next year, up from around £12 billion in 2023. It also refers to a widely reported case in Hong Kong where fraudsters allegedly used an immersive video call to convince an employee to transfer £25 million. Meanwhile, a previous investigation by the Associated Press found an AI-generated LinkedIn profile that successfully infiltrated US policy circles.
The study also highlights another important problem: AI systems remain less reliable in producing older faces, younger faces and people from underrepresented ethnic groups due to biases in their training data. Those imperfections may provide useful clues to human observers.
Perhaps the most interesting takeaway is that the human brain seems to learn as much as AI itself. By repeatedly seeing examples of real and fake faces, people gradually develop a more accurate sense of authenticity rather than relying on a single rendering. Researchers believe that instinct may become one of our most powerful tools as productive AI continues to improve.
The irony is hard to ignore. As artificial intelligence gets better at impersonating people, humans may have to start training themselves the way machines do – with data, repetition, and pattern recognition. AI detectors may continue to improve, but research suggests they should not be the only defense. Human judgment still has a role to play; it just needs improvement.
