The latest Australian research shows that it is difficult for ordinary people to reliably distinguish faces generated by artificial intelligence from real faces. Even for a small number of people known as "super recognizers", the advantage is not obvious. Researchers warn that such highly realistic synthetic faces could pose serious challenges to security clearance and identity verification, as people are often overconfident in their ability to identify people.
The research was jointly conducted by psychology teams from the University of New South Wales (UNSW Sydney) and the Australian National University (ANU), and the paper was published in the British Journal of Psychology. The research team pointed out that today’s “synthetic AI faces” may even be more “real” in subjective perception than real portrait photos, and relevant experimental data provides support for this judgment.
James Dunn, a researcher at UNSW's School of Psychology, said people have long believed they could recognize AI-generated fake faces at a glance. Early AI systems did often generate distorted, deformed, and imperfect images, but with the rapid development of generative models, it has become extremely easy to create an extremely realistic face.
In the specific experiment, the two universities recruited a total of 125 participants, including 36 subjects identified as "super recognizers" and 89 ordinary participants as a control group. Everyone completed the same test online: real faces and AI-generated faces were randomly presented on the screen, and the subjects were asked to judge whether each was "real" or "synthetic."
The results showed that the judgment performance of ordinary people was only slightly better than the random level of "blinded" people, while the "super recognizers" had higher overall scores, but their advantages were very limited. What’s even more ironic is that regardless of their recognition ability, all participants had equally high confidence in their “ability to see through AI fake faces”, indicating significant overconfidence.
Research points out that modern AI-generated faces are no longer as full of obvious flaws as they were in the early days. Typical "algorithm traces" such as blurred backgrounds, missing organs, and weird proportions of facial features appear less and less. Subsequently, the potential risks of these extremely confusing fake faces are rapidly accumulating in scenarios such as online fraud, identity impersonation, and social engineering attacks, and both individuals and institutions may be harmed as a result.
ANU Associate Professor Amy Dawel points out that today's AI fake faces "do not reveal themselves because something is terribly wrong, but because they are too good". Rather than saying that they are abnormal, it is better to say that they are "too normal": the face is highly symmetrical, the facial features are well-proportioned, and the overall characteristics are at the "average" in a statistical sense. They look both pleasing and "pleasing to the eyes." She emphasized that this kind of "abnormally perfect" and "too average" appearance characteristics can itself be regarded as a potential danger signal generated by AI, and people with super facial recognition abilities may be better able to capture these subtle patterns.
The research team's next plan is to further analyze which perceptual cues these "super AI fake face detectors" rely on, in order to promote the improvement of technical automatic detection tools and mass recognition strategies. Currently, UNSW has launched the “UNSW Face Test” online test on its official website, which includes a free demo version of AI fake face recognition capabilities. Ordinary users can also evaluate their own recognition level.
After personally participating in the test, the author of the article found that he correctly identified 12 AI fake faces out of 20 faces twice. According to official instructions, this result is expected to be classified as a "super recognizer", while the average score of ordinary participants is about 11/20. As AI-generated images continue to improve, such a slight lead is enough to highlight the growing gap between technology and human perception.