summary: If you think you can “immediately tell” that a face is generated by AI, you may be overconfident and putting yourself at risk. A new study by reveals that humans are now constantly fooled by state-of-the-art facial generation systems.
Even “superrecognizers,” people with a rare natural talent for identifying faces, performed only slightly better than chance. Ironically, the study found that the most important “feature” of an AI’s face is not a glitch or a distorted ear, but perfection itself. AI faces tend to be “too good to be true”, unusually symmetrical, well-proportioned, and statistically average.
important facts
- Overconfidence gap: Most people rely on outdated cues (crooked teeth, messy backgrounds) that modern AI has already corrected.
- Super Recognizer was fooled: People who are typically good at recognizing human faces had about as much trouble identifying synthetic faces as the general population.
- “Too average” rule: AI-generated faces are often more symmetrical and “average” in proportion than real human faces, which our brains mistakenly interpret as more attractive and relatable.
- Security risks: This misplaced confidence makes individuals and organizations even more vulnerable to fraud, fraudulent LinkedIn profiles, and deepfake IDs in recruitment and dating.
- The need for skepticism: Researchers argue that we need to move toward a healthy level of skepticism regarding all digital images because visual judgment can no longer be trusted.
sauce: New South Wales
Most people believe they can identify faces generated by AI, but a study from UNSW Sydney and the Australian National University (ANU) proves that confidence is outdated.
Faces generated by AI are becoming almost impossible to distinguish from real faces, and researchers warn that this misplaced confidence could leave individuals and organizations even more vulnerable to scammers, fraudsters and bad actors.
“Until now, people have been confident in their ability to spot fake faces,” says researcher Dr James Dunn from the UNSW School of Psychology. “But faces created by state-of-the-art facial generation systems are no longer that easy to detect.”
In a research paper published in British Journal of Psychology, Researchers at UNSW and ANU recruited 125 participants, including 36 people with exceptional facial recognition skills known as super-recognizers, and 89 control participants, to take part in an online test in which they were shown a series of faces and asked to decide whether each image was real or generated by AI. Obvious visual defects were eliminated in advance.
“What we saw was that people with average facial recognition ability performed only slightly better than chance,” says Dr. Dunn.
“And while the super-recognizers performed better than the other participants, it was only by a small margin. What was consistent was people’s confidence in their ability to recognize AI-generated faces, even if that confidence did not match their actual performance.”
The end of crafts
Much of that confidence comes from cues that have worked before. Early AI-generated faces were often lost in obvious visual artifacts, such as crooked teeth, glasses that blended into the face, ears that weren’t attached properly, and strange backgrounds that bleed through hair and skin.
However, as face generation systems have improved, these types of errors have become much less common. The most realistic output no longer shows obvious flaws, leaving faces that look convincing at first glance, but are much harder to judge using the cues people are used to.
“Many people think they can still tell the difference because they have used popular AI tools such as ChatGPT or DALL·E,” says ANU psychologist Dr Amy Dowell. “However, these examples do not reflect how realistic state-of-the-art facial generation systems have become, and relying on them can give people a false sense of confidence.”
What interested the researchers was that even super-recognizers could be easily fooled. Although this group performed better on average, its advantage was modest and its accuracy remained well below that typically achieved when recognizing real human faces.
There was also considerable overlap between groups, with some groups other than super-recognizers outperforming super-recognizers. This shows that this is not simply a matter of experts versus everyone.
too good to be true
But if AI faces are so convincing, what should we be looking for?
“Ironically, the face of cutting-edge AI is lost not by its problems, but by being too right,” says Dr. Dowell. “Rather than an obvious defect, it has an unusually average trend, is very symmetrical, well-proportioned, and statistically typical.”
Qualities such as symmetry and average proportions usually indicate attractiveness and friendliness. However, current research raises red flags that they are artifacts.
“It’s as if their faces are too good to be true,” says Dr. Dowell.
what to do about it
The super recognition feature was as unremarkable as usual in tests with real human faces, showing only a small benefit. What differentiated them was the same quality identified in the study: a higher sensitivity to plausible, unusually average, and highly symmetrical faces. Still, limited success suggests that finding faces in AI is not a skill that can be easily trained or learned.
This finding also has practical implications, as relying solely on visual judgment is no longer reliable. This is important in a variety of situations, from social media and online dating to professional networking and recruiting. There, people often assume they can “tell right away” if a profile photo looks fake. False confidence can make individuals and organizations more vulnerable to fraud, fake profiles, and fabricated identities.
“There needs to be a healthy level of skepticism,” Dr. Dunn says. “For a long time, we were able to look at a photo and assume we were seeing a real person. That assumption is now being challenged.”
Rather than teach you tricks to finding synthetic faces, the broader lesson is to update your assumptions. The visual rules that many of us rely on were shaped by earlier, less sophisticated systems.
“As face generation technology continues to improve, the gap between what appears plausible and what is real may widen, and it will become increasingly important to recognize the limits of our own judgment,” says Dr. Dowell.
Looking to the future
Interestingly, Dr. Dunn thinks the research team may have discovered a new type of facial recognition device.
“Our research reveals that some people are already AI face-detecting detectives, suggesting that a ‘super AI face detector’ may exist.
“We want to learn more about how these people are able to spot fake faces, what cues they use, and see if we can teach these strategies to other people.”
- Are you good with your face? Please visit. New South Wales face test On this page, you can test your facial recognition skills and see how well you can recognize AI faces.
Answers to key questions:
answer: Approximately 2% of the population has an uncanny ability to remember and identify faces they saw only briefly a few years ago. Usually these are the golden rules of police work, but now even they are being overtaken by the realism of AI.
answer: Paradoxically, strive for “perfection.” If a face appears mathematically average, perfectly symmetrical, and without the slightest “human” imbalance, it is actually more There is a high possibility that it is AI. Real human faces almost always have slight asymmetries.
answer: We are hardwired to trust what we see. If you believe you’re talking to a real person on a professional network or dating app, you’re much more likely to share sensitive information or fall victim to financial fraud.
Editorial note:
- This article was edited by the editors of Neuroscience News.
- Journal articles were reviewed in full text.
- Additional context added by staff.
About this AI and facial recognition research news
author: lachlan gilbert
sauce: New South Wales
contact: Lachlan Gilbert – UNSW
image: Image credited to Neuroscience News
Original research: Open access.
“Too Good to Be True: Synthetic AI faces are more average than real faces, and super-recognizers know it.” (Author) British Journal of Psychology
DOI:10.1111/bjop.70063
abstract
Too Lie: Artificial AI faces are more average than real ones, and super-recognizers know it
The AI revolution is creating synthetic faces that look more human-like than photos of real people.
We tested whether individual differences in human face recognition ability could explain variation in AI and real face identification. Super Recognizers – People with an excellent ability to recognize human faces (N= 36) – outperformed a typical sample by 15% and 7% compared to a group of high-performing, motivated control participants (Cohen’s d= 0.55; N= 89).
Individual difference analysis revealed that this pattern was positively associated with human face recognition and AI face identification ability. The AI’s discriminatory ability was also related to an individual’s susceptibility to the “super-average” appearance of the AI’s face.
Deep neural networks optimized for face identification processing have confirmed a more concentrated distribution of AI faces in face space. Moreover, centrality was associated with a higher probability that the superrecognizer judged a face to be AI, whereas this pattern was not observed in controls.
The correct interpretation of hyperaverageness as a cue to artificiality by hyperrecognizers constitutes the first mechanistic link between evolved expertise in face processing and AI face detection, and addresses common misconceptions about the structure of human face space.
