The Brief
Science & Discovery 5 min read

Fake Faces Rated More Real Than Real Ones

NAVION

Share

For years, spotting an AI-generated face was a game anyone could win. Mangled ears, an unsettling surplus of teeth, a gaze that never quite landed: the tells were obvious and a little funny. A new study from psychologists at the University of Florida, published in Social Cognitive and Affective Neuroscience, suggests that game is over. The faces now look so convincing that people rate them as more realistic than photographs of actual humans.

What the experiment found

The research team recruited more than 150 participants, split between younger adults (18 to 37) and older adults (60 to 89). Each person viewed photos of either real human faces or AI-generated ones and rated how realistic each looked on a scale from 1 to 10. All the while, an fMRI machine scanned their brains.

The result runs against intuition. Both age groups gave higher realism scores to the computer-made faces than to the real ones. Study author Natalie Ebner put the problem bluntly: “Fake faces are being perceived as more real than real faces.” She added that people are doing worse than random guessing, because they are “actively being drawn in the wrong direction.”

That last point deserves attention. If people were simply unable to tell the difference, accuracy would hover around chance. Instead, the bias pushes judgment toward the wrong answer. A viewer who trusts their gut about which face “looks most real” will, on average, pick the fake.

The brain region that stayed quiet

The scans add a second layer. When participants looked at AI faces, activity dropped in the anterior insula, a region tucked inside the Sylvian fissure, the deep crease along the side of the brain. This area is involved in processing cues that signal untrustworthiness.

Muted activity there suggests that the synthetic faces simply never triggered a neural alarm. Nothing felt off, so nothing got flagged. This is a different kind of failure from the classic uncanny valley, where a near-human face provokes discomfort precisely because something is slightly wrong. These faces appear to have crossed that valley entirely.

The researchers’ explanation is elegant. AI faces are, in effect, an average of many faces. Distinctive quirks get sanded away, leaving something prototypical, attractive, and familiar. And familiar faces tend to feel more trustworthy. Real human faces carry asymmetries, odd proportions, and individual character. Those very differences may read to the brain as slightly less “typical,” and therefore slightly less real.

Why older adults rate fakes even higher

One finding stands out. Brain activity did not differ significantly between the age groups, yet older adults still judged AI faces as more real. Both groups rated actual faces similarly, but older participants scored AI faces 1.11 points higher in realism, compared with just 0.25 points for younger participants.

Ebner was careful to frame this correctly: “All adults have hyperrealism bias, not just older adults. But older adults have it more.” The researchers are not certain why. One possibility is that older people lean on the familiar more generally. Another is that they are less aware that AI-generated deepfakes exist at all.

What this means beyond the lab

The practical consequence is straightforward: the old advice to “look closely for flaws” is no longer reliable. Visual inspection, the instinct most of us would reach for, may actively mislead. A face that feels trustworthy and natural is not evidence of a real person behind it.

This matters most for scams. The researchers point out that the greater bias among older adults makes them especially attractive targets. Ebner framed the finding as a starting point for defense: “Identifying that this potential deception scheme is the first step to an intervention.” Her suggested habit is simple, if hard to practice: if something looks too good to be true, think twice before going along with it.

The study also hints at a deeper shift in how we should think about trust online. Our perceptual systems evolved to read faces in a world where every face belonged to a person. Those instincts carry hidden assumptions, such as the idea that a familiar, pleasant face signals safety. Synthetic faces can exploit that machinery precisely because they are optimized toward the average, and the average is what the brain finds comfortable.

There is also a candid note of realism in the source about the road ahead. Raising awareness sounds easy, but it is hard in an information environment already crowded with scams. Knowing a bias exists does not automatically switch it off, in the same way that knowing an optical illusion is an illusion does not make it disappear. Effective responses will likely need to go beyond individual vigilance and pair human judgment with other signals, rather than asking people to out-stare a machine.

In short

A University of Florida study found that both younger and older adults rated AI-generated faces as more realistic than real ones, while brain scans showed reduced activity in the anterior insula, a region tied to detecting untrustworthy cues. The likely reason is that AI faces blend many faces into a familiar, prototypical average that the brain finds easy to trust. Older adults showed a stronger bias, which makes them especially exposed to scams, though every group was affected. The lesson is that a face feeling real is no longer a reliable sign that it is real, and that awareness, caution, and independent verification matter more than a careful look.

Based on reporting from Nautilus.

Written by

NAVION