
Artificial intelligence has become so proficient at generating human faces that telling a real photograph from a synthetic one is increasingly difficult. But according to a new study from the Emotions and Faces Lab at the Australian National University (ANU), that skill can be taught.
The research found that people can be trained to identify AI-generated faces, significantly improving their accuracy even when the images appear highly realistic. The findings could provide a valuable new tool in the fight against the growing wave of AI-enabled fraud fueled by deepfakes.
According to the study’s lead author, Professor Amy Dawel, traditional detection methods have shown clear limitations. Looking for visual glitches—such as an extra finger or mismatched accessories—is becoming less effective as AI models grow increasingly sophisticated and obvious mistakes become rarer.
Instead, the researchers developed a different approach that teaches participants to focus on the overall characteristics of a face rather than isolated flaws. AI-generated faces tend to be unusually symmetrical, well-proportioned, and conventionally attractive. Ironically, these are precisely the features that people often instinctively associate with real human faces.
The training program focuses on six perceptual qualities: distinctiveness, memorability, proportionality, symmetry, attractiveness, and expressiveness. After completing the training, every participant improved their ability to identify AI-generated faces, while the top performers achieved near-perfect accuracy.
The findings were independently replicated by a research team at the University of Victoria in Canada, led by Professor Jim Tanaka and Dr. Eric Mah. Reproducing the results with a new group of participants in a different country strengthens the study’s conclusions. Because the training can be delivered entirely online, the researchers say it could be deployed at scale at relatively low cost, offering a practical tool to help combat digital fraud.

