Summary: If you believe you can reliably spot an AI-generated face by sight alone, recent research suggests you are likely overconfident — and potentially at risk. A new study shows that humans, including those with exceptional face-recognition ability, are frequently deceived by the most advanced face-generation systems.
Even so-called “super-recognisers” — people with a rare, natural talent for identifying faces — performed only slightly better than chance when asked to distinguish real photos from AI-generated portraits. Paradoxically, the clearest clue that an image is synthetic isn’t a distortion or glitch, but an absence of human imperfection: AI faces often appear unusually symmetrical, well-proportioned and statistically average, making them paradoxically “too perfect” to be real.
Key Facts
- The Overconfidence Gap: Many people rely on outdated visual cues (crooked teeth, messy backgrounds, obvious artefacts) that modern AI generation has largely eliminated.
- Super-Recognisers Fooled: People who usually excel at recognizing human faces struggled almost as much as the general population when judging synthetic faces.
- The “Too Average” Rule: AI-generated faces tend to be more symmetrical and average in their proportions than real faces — traits our brains associate with attractiveness and familiarity, but which can signal artificiality.
- Security Risks: Misplaced confidence in visual judgment increases vulnerability to scams, fake social or professional profiles, and identity manipulation in recruitment or dating contexts.
- Need for Scepticism: Researchers urge greater scepticism of digital images and recommend relying on verification beyond visual inspection alone.
Source: UNSW
Most people think they can spot AI-generated faces, but that belief is now out of date, according to research conducted by UNSW Sydney and the Australian National University (ANU).
As face-generation models have advanced, the familiar signs of synthesis have faded. The new study warns that continued confidence in visual inspection alone could expose individuals and organisations to fraud, fabricated identities and social-engineering tactics.
“Until recently, people were confident they could identify a fake face,” says Dr James Dunn of the UNSW School of Psychology. “But faces produced by the most advanced generation systems are no longer easily detectable by the visual rules people learned from older tools.”
In experiments reported in the British Journal of Psychology, researchers recruited 125 participants, including 36 super-recognisers and 89 control participants. Each participant completed an online test showing a sequence of faces, and judged whether each image was a photograph of a real person or created by AI. Images with obvious artefacts had been removed before testing to focus on the most convincing examples.
“Participants with average face-recognition ability performed barely above chance,” Dr Dunn reports. “Super-recognisers did better, but only by a small margin. Across groups, people remained confident in their judgments even when performance didn’t justify that confidence.”
The end of artefacts
Confidence often stems from cues that worked in the past. Early AI images were betrayed by clear visual artefacts: malformed ears, teeth that didn’t align, glasses merging into faces, or backgrounds bleeding into hair. But as generation models improved, those errors diminished. The most realistic outputs now lack the obvious flaws that people typically look for, so quick visual checks no longer provide reliable protection.
“Many people base their judgment on examples from popular consumer tools,” says ANU psychologist Dr Amy Dawel. “Those familiar interfaces don’t reflect the sophistication of the leading face-generation systems, and relying on them can create a false sense of security.”
The study also found overlap between groups: some ordinary participants outperformed certain super-recognisers, indicating this challenge is not simply experts versus everyone else.
Too good to be true
If obvious flaws are disappearing, what cues remain? Ironically, the most advanced AI faces are flagged by what they lack: small, natural imperfections. They tend to be unusually average — highly symmetrical and ideally proportioned. Traits that normally signal attractiveness and familiarity therefore become indicators of synthetic origin.
“They’re almost too perfect,” Dr Dawel explains. “Those subtle human asymmetries that make a face unique are often missing, and that very uniformity can be the giveaway.”
What to do about it
The modest advantage super-recognisers showed in the study stemmed from greater sensitivity to hyper-average features. Even so, their limited success suggests that spotting AI faces is not a skill that can be reliably taught with simple tricks. Rather than training people to find artefacts that may no longer exist, the broader lesson is to update assumptions and verification practices.
Practically, this means treating visual evidence with caution: verify identities through independent channels, check metadata where available, use platform verification features, and be wary of requests for sensitive information from profiles that were judged solely by appearance. In professional contexts — hiring, networking, online marketplaces and dating — policies and tools that go beyond visual inspection will reduce risk.
“A healthy level of scepticism is needed,” Dr Dunn says. “For a long time a photograph implied a real person. That trust is being undermined by rapidly improving AI.”
Dr Dawel adds: “As face-generation technology continues to improve, the gap between what looks plausible and what is real may widen. Recognising the limits of visual judgment will become increasingly important.”
Looking ahead
The researchers note an intriguing possibility: some people may be naturally better at spotting AI faces than others, hinting at a potential category of “super-AI-face-detectors.” The team plans to study what cues these individuals use and whether any of those strategies can be effectively taught or embedded in detection tools.
- Good with faces? Visit the UNSW Face Test page where you can test your face recognition skills and see how well you can spot AI-faces.
Key Questions Answered:
A: About 2% of the population has an extraordinary ability to remember and identify faces, even after brief encounters. These individuals are often relied upon in police and security settings, yet even they are challenged by highly realistic AI-generated faces.
A: Paradoxically, look for “perfection.” If a face appears mathematically average, perfectly symmetrical and lacks minor human imbalances, it may be more likely to be AI-generated. Real faces almost always show small asymmetries.
A: Humans are predisposed to trust photographic evidence. If we assume we are interacting with a real person based on a convincing image, we are more likely to share personal information or fall for financial and social-engineering scams.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context added by staff.
About this AI and facial recognition research news
Author: Lachlan Gilbert
Source: UNSW
Contact: Lachlan Gilbert – UNSW
Image: The image is credited to Neuroscience News
Original Research: Open access. “Too Good to be True: Synthetic AI Faces are More Average than Real Faces and Super-recognisers Know It.” British Journal of Psychology. DOI: 10.1111/bjop.70063
Abstract
Too Good to be True: Synthetic AI Faces are More Average than Real Faces and Super-recognisers Know It
The rise of powerful generative models has produced synthetic faces that at times appear more human than photographs of real people. This study tested whether individual differences in face-recognition ability predict how well people discriminate AI-generated faces from real ones. Super-recognisers (N = 36) outperformed a typical sample by 15% and a higher-performing motivated control group by 7% (Cohen’s d = 0.55; N = 89). Individual differences showed a positive association between human face-recognition skill and AI-face discrimination. Those better at detecting AI faces were also more sensitive to the “hyper-average” appearance of these images.
Deep neural networks optimized for face identity processing confirmed that AI faces occupy a more central position in face-space. Centrality correlated with a higher probability that super-recognisers would judge a face as AI-created, a pattern not observed for control participants. The finding that super-recognisers interpret hyper-averageness as a cue to artificiality links evolved expertise in face processing to AI-face detection and challenges assumptions about the structure of human face perception.