In an experiment, a research team from the Association for Psychological Science (APS) aimed to determine the difficulty in distinguishing AI-generated faces from real human photos. The results are surprising: AI-generated faces appear indistinguishable from real faces.
The data even suggest that AI-generated images were more often perceived as human. In 65.9 percent of cases, they were mistaken for humans. In comparison, real human faces seemed less authentic, correctly identified in only 51.1 percent of cases—just slightly better than the 50 percent chance through random guessing. However, the experiment has certain limitations.
Initially, only faces that could theoretically be identified as belonging to individuals of white ethnicity were used. Consequently, only white individuals were selected as the test group to differentiate between AI-generated and real images. This was done to eliminate various variables that could skew the results. The assumption is that faces from different cultures might stand out disproportionately for different ethnicities.
Furthermore, the experiment is not universally applicable for several reasons. The final test group that contributed to the results consisted of 124 individuals—61 men, 62 women, and one person identifying as non-binary. All participants were from the United States and had not lived in another country for more than two years.
AI-generated using StyleGAN2
For generating the images, Nvidia’s neural network StyleGAN 2 was used. It generates images of realistically appearing faces. The real photos were sourced from the AI training dataset Flickr-Faces-HQ. Subsequently, the faces generated by AI were inspected for artifacts and cleaned—such as obvious doubly generated faces in the background. Thus, this small-scale study might not necessarily reflect the technological advancements of current image generators.
Participants were asked to view the images from a distance of 50 cm and at an angle of 12 degrees, assessing and categorizing them based on their observations. They primarily used criteria such as image quality—coloring, rendering artifacts, or obvious editing errors—as well as specific facial features like eyes, ears, neck, and teeth to distinguish between real and generated photos.
In a second experiment, a separate test group of 610 individuals (290 men, 312 women, 8 non-binary) was asked to rate the images on a scale of 0 to 100 based on specific criteria and to determine whether it was a human or artificially created face. The criteria included aspects like age, facial symmetry, attractiveness, and the expressiveness of a face. Using these values, the team developed a classifier that aimed to recognize the faces created via StyleGAN2. It performed reasonably well: the software could identify 95 percent of human faces and 93 percent of AI-generated faces accurately.
READ MORE: MacOS 14 Sonoma Release Plagued by Bugs: Users Struggle Amidst System Flaws
