People may mistake AI-generated faces for real people, and the error is not evenly distributed across racial groups. A study of face recognition found that white synthetic faces were often judged more real than photographs, while the same pattern did not emerge for images of people of color.
When synthetic faces seem more human
The researchers call this effect “AI Hyperrealism”: the tendency to see AI-generated faces as more real than human faces. Their experiments tested whether participants could distinguish between photographs and synthetic faces, and how confident they felt about their judgments.
In one experiment, 124 white adults saw a mixture of 100 AI-generated and 100 real white faces. Participants identified 66 percent of the AI images as human, compared with 51 percent of the real images. For images of people of color, AI-generated and real faces were both judged human about 51 percent of the time, regardless of participants’ race.
The finding matters because a mistake in judging one face does not necessarily stay a harmless guessing error. If synthetic faces appear convincing in contexts where people make decisions based on images, that could affect how those images are received.
Why the images looked convincing
The study used synthetic faces from StyleGAN2, an image generator released in 2020. The technology has since advanced rapidly, but newer image-generation models were not part of this research, so the results describe how participants responded to the images used in these experiments.
A second experiment asked 610 adults to rate AI and human faces on different attributes without telling them that some images were generated. The researchers used “face space” theory to examine which perceived traits might explain why some faces seemed human.
Responses pointed to proportionality, familiarity and lower memorability as factors linked to mistaking AI faces for people. The researchers suggested that synthetic faces’ attractiveness and “averageness” could make them seem more real. By comparison, the wider variation in real faces may have made some photographs feel less typical to participants.
That explanation helps show why visual realism is not simply a matter of adding detail. A face that seems familiar and balanced may feel authentic, even when it was generated, while real human variation can make a photograph feel unusual.
Confidence did not guarantee accuracy
People who made more errors were also more confident in their judgments. The researchers describe this relationship as a manifestation of the Dunning-Kruger effect: in this experiment, people who were more sure of their answers were more often wrong.
The study also found that a machine-learning system could correctly identify whether a face was real or AI-generated 94 percent of the time. That result shows a difference between human judgment and the system’s performance on the images tested. It does not mean people can reliably spot synthetic faces by simply looking harder, nor does the reported result establish how the system would perform on images made by newer models.
Bias can shape where the risks fall
The racial difference in the results raises questions about how people interpret synthetic faces and whose images seem most convincing. The article links the disparity to AI models being trained predominantly on images of white individuals, a bias already known in machine-learning research.
That imbalance could matter in areas that use faces to represent or identify people. The study authors raised possible implications for locating missing children, while the difficulty of recognizing synthetic faces more broadly could create opportunities for fraud or identity theft. The article also points to possible effects in areas such as online therapy and robotics.
The results do not show that every person will mistake an AI face, or that every generated face will seem real. They do show why systems and decisions that depend on face images need to account for both human misjudgment and uneven performance across racial groups. As image generation improves, the study’s findings offer a reason to take those differences seriously.