A recent study published in the British Journal of Psychology suggests that people with exceptional face recognition skills are slightly better at telling artificial intelligence-generated faces from real ones. The research provides evidence that computer-generated faces tend to look mathematically hyper-average, a subtle clue that these top-tier face recognizers subconsciously detect. Overall, the findings indicate that visual intuition alone is no longer enough to spot modern synthetic faces, highlighting a growing vulnerability to digital deception.
Historically, the human face processing system evolved to extract emotion and social meaning from real people. Now, artificial intelligence programs can generate synthetic faces that are nearly indistinguishable from real humans.
These synthetic faces pose a significant threat in the real world. Bad actors routinely use artificial faces for illegal activities, such as creating fake profiles for corporate cyberespionage, running online dating scams, and spreading propaganda through automated accounts. Because earlier artificial intelligence programs made obvious visual errors, such as distorted teeth or strange backgrounds, people could easily spot the fakes.
As technology has advanced, those obvious glitches have largely disappeared. Scientists wanted to find out if certain people possess higher-level perceptual abilities that allow them to spot more subtle, structural differences between real and generated faces.
“AI-generated faces are now so realistic that most people can’t reliably tell them apart from real faces,” said study author James Dunn, a lecturer at UNSW Sydney and principal investigator in the Face & Forensic Psychology Research Lab. “That creates real-world risks—from scams and fake job applicants to misinformation campaigns using synthetic identities. At the same time, we know that some people are exceptionally good at recognising faces (‘super-recognizers’), but no one had tested whether that expertise helps with AI detection. We wanted to understand not just who is better at spotting AI faces, but also why.
For their study, the scientists recruited a total of 125 participants. The sample included 36 super-recognizers, who had previously scored in the top tier of standardized face recognition tests, and 89 highly motivated control participants with above-average but not exceptional skills. The participants completed an online task where they viewed a series of 200 face images. (You can take the test here.)
Half of these images were photographs of real white men and women. The other half were artificial intelligence-generated faces designed to match the real faces in gender, posture, and expression. For each image, participants had to decide whether the face was real or computer-generated, and then they rated their confidence in their decision on a scale from zero to 100.
To understand the hidden structure of the artificial faces, the researchers also analyzed the images using artificial neural networks. These are advanced computer programs designed to mimic the way the human brain processes information, specifically trained here to recognize face identities. The scientists used these computer models to map out a mathematical landscape called face-space.
A face located at the center of this mathematical space is considered highly average, meaning it lacks distinct or unusual physical proportions. The computer analysis provided evidence that artificial intelligence-generated faces are overwhelmingly located closer to this center than real faces. In other words, the artificial faces were exceptionally typical and symmetrical, lacking the natural quirks of real human faces.
When looking at the human participants, the researchers found that typical individuals performed no better than a coin flip. The control group correctly identified the faces only 50.7 percent of the time. The super-recognizers performed slightly better, achieving an average accuracy of 57.3 percent.
While the super-recognizers only had a modest advantage, they demonstrated a deeper awareness of their own performance. When super-recognizers felt highly confident about a guess, they were more likely to be correct. The control participants showed no such relationship between their confidence and their actual accuracy.
The researchers also found that super-recognizers and regular participants relied on entirely different visual clues. Super-recognizers subconsciously used face-space centrality as a warning sign. When a face looked too perfectly average and symmetrical, the super-recognizers tended to classify it as artificial.
“One striking finding was that AI faces occupy the very centre of ‘face-space,’ a kind of mental map of face, more than real human faces do, and that clue is what gives super-recognizers their advantage,” Dunn told PsyPost. “Traditional theories might assume the most human-looking faces should sit at the center, but we found the opposite, and this comes back to the ‘hyper-averageness’ of AI faces.”
Regular participants completely missed this structural clue. Instead, the control group relied heavily on perceived youthfulness. Regular participants frequently made the mistake of assuming that older-looking faces were real human beings, which led to incorrect guesses.
To see if group collaboration could boost accuracy, the scientists ran a statistical simulation called the wisdom of crowds. They combined the answers of multiple participants to see if a consensus choice was more accurate. Aggregating the responses of eight super-recognizers boosted the group’s detection accuracy significantly, but applying the same mathematical technique to the control group produced no improvement at all.
One potential misinterpretation of this study is the assumption that human experts can consistently protect us from artificial intelligence deception. Even the highly skilled super-recognizers only achieved an accuracy rate of 57 percent, which is far lower than their typical performance on standard face identification tasks.
“The performance advantage we observed was meaningful but not dramatic,” Dunn explained. “Super-recognizers performed about 7% better than motivated controls, and their accuracy improved further when we combined responses across small groups. However, even they were far from perfect. So while face expertise helps, it’s not a complete solution to real-world AI deception—at least not yet. We hope to use the insights from this paper to develop training that can deliver greater benefits to everyone in detecting AI-faces.”
In addition, “some of the people who were best at spotting AI-faces were not super-recognizers,” Dunn said. “This gives us hope, it means that everyone, even people with prosopagnosia or face blindness, could be good at spotting AI-faces from real ones.”
The researchers note that visual judgment is simply no longer reliable for high-stakes security situations. The subtle structural differences that super-recognizers detect are too slight to depend on for fraud prevention or verifying identities online. Because artificial faces are designed to mimic real statistical properties, their features often overlap with those of exceptionally attractive or symmetrical real people.
Future research will need to explore new ways to improve detection accuracy. Scientists hope to investigate whether hybrid systems that combine human judgment with algorithmic tools can offer better protection against synthetic media. They also plan to look for individuals who might possess a specific, natural talent for detecting artificial faces, distinct from traditional face recognition skills.
“We’re interested in whether people can be trained to better detect the statistical cues that distinguish AI faces from real ones,” Dunn explained. “We’re also exploring hybrid approaches—combining human judgments with algorithmic tools. More broadly, we want to understand how increasing exposure to AI-generated faces might reshape human face perception over time.”
“One broader implication is that AI-generated faces may not be neutral stand-ins for real people. Because they are systematically more “average,” they may influence memory, trust judgments, and even how children develop their mental representation of faces. As synthetic identities become more common online, understanding these subtle perceptual shifts will be increasingly important.
The study, “Too good to be true: Synthetic AI faces are more average than real faces and super-recognizers know it,” was authored by James D. Dunn, David White, Clare A. M. Sutherland, Elizabeth J. Miller, Ben A. Steward, and Amy Dawel.
