Two years ago, catching an AI-generated face was almost too easy. You looked for the telltale glitches: six fingers, an earring floating in mid-air, a smear where a background should be. By 2026, those obvious tells are mostly gone, and some synthetic portraits now fool even seasoned observers.
A report from Scientific American argues that we’ve been hunting for the wrong things. Instead of scanning for defects, researchers say you should judge a face by six perceptual qualities: distinctiveness, memorability, proportionality, symmetry, attractiveness, and expressiveness.
The counterintuitive part? The very traits we instinctively read as “human” are often the giveaways. AI faces tend to be more symmetrical, more proportional, and more attractive than real ones.
“Our training directs people’s attention to global qualities that differ between AI and human faces,” says Amy Dawel of the Australian National University. “AI faces tend to be more symmetrical, proportional, and attractive, but without training, we often think these are markers of being human.”
Dawel, who led the work at ANU’s Emotions and Faces Lab, says guiding people toward these broader patterns produced striking results, with some participants reaching 100% accuracy. “It was amazing to see the dramatic improvement in people’s ability to detect AI faces,” she says.
The reason these images skew toward statistical perfection comes down to how they’re built. Generative models are trained on enormous databases of photos, largely scraped from the web without authorization, then used to construct what amounts to a statistically average human face. The output looks polished but oddly hollow.
That’s where the trained eye comes in. An AI face tends to be:
- Unmemorable — forgettable rather than striking
- Less expressive — emotionally flat
- Too proportional — features arranged with suspicious balance
- Too symmetrical — real faces are rarely mirror-perfect
The study focused on StyleGAN faces, among the most convincing fakes currently in circulation. “We’ve shown our training is effective for some of the most convincing fakes available, StyleGAN faces,” Dawel notes. “Now we need to find out whether that training generalises to other AI-generated faces.”
The encouraging takeaway is how little effort it takes. “We found that even relatively short training sessions helped participants improve their accuracy in detecting AI-generated faces, highlighting the potential for practical education tools in this area,” says ANU Honours student Tanya Georg, who ran the training.
“AI image generation technology is improving extremely quickly, and many people underestimate how convincing these faces can be,” Georg adds. “Research like this can help people navigate increasingly complex online environments.”
In other words: the next time a profile picture looks flawless, that flawlessness might be exactly the problem.
