During the study, researchers analyzed how people perceive generated images and why most participants cannot accurately distinguish them from real ones. The focus was on a group of "super-recognizers" — individuals with exceptional abilities to differentiate and remember faces.
The study involved 36 such experts and 89 volunteers with high scores on perception tests. They were shown 200 images, half of which were created using a neural network, while the other half consisted of real photographs. All images were selected to be similar in gender, emotional expression, and other basic characteristics.
The results were impressive: ordinary participants could not distinguish the "fakes" from the original images, and their ability to recognize the difference was close to random guesses. Super-recognizers performed better, but even they achieved an accuracy of only about 57%.
This data indicates that the task remains challenging even for experts.
Researchers also identified an important pattern: the better a person recognizes real faces, the more effectively they differentiate artificial ones. There is a consistent relationship between these skills, suggesting that the perception of AI-generated portraits is based not on searching for technical defects but on fundamental mechanisms of face perception.
A unique effect was observed when the group interacted. When eight super-recognizers combined their opinions, their accuracy significantly increased. In the control group, the "wisdom of the crowd" did not manifest, indicating a high level of self-reliance and accuracy in the experts' self-assessment.
To understand the reasons for the identified differences, scientists analyzed the images using neural networks trained for face recognition. This allowed them to create a "face space" map — a multidimensional model in which each face is represented as a set of specific characteristics.
It turned out that real faces are unevenly distributed in this space and differ in many unique details. At the same time, generated images are concentrated closer to the center — in the area of the "averaged" face.
In other words, AI tends to create the most typical, statistically averaged portraits. Researchers termed this effect "hyper-averageness," which arises from the peculiarities of generative models: algorithms suppress rare and unstable features, focusing on the most common ones. As a result, what emerges is not a specific person but an idealized image with minimal deviations from the norm.
Paradoxically, this is what makes AI-generated faces more convincing. Most people possess unique combinations of features that rarely occur together. Such faces are statistically "uneven," while the neural network creates more harmonious and "correct" images.
Super-recognizers, as researchers found out, intuitively recognize this feature. They focus not on the attractiveness, youth, or emotional expression of faces, but on their "similarity to the averaged image." This criterion helps them recognize generated images.
However, the experts themselves cannot clearly explain their methods. Their approach is intuitive and formed based on unconscious experience.
The authors of the study emphasize that even the most experienced observers encounter the limits of their abilities. With the development of generative technologies, the task will become increasingly complex.
The practical implications of the research could impact various fields. Scientists warn that using AI-generated faces in psychological experiments, training, or legal processes may distort perception and influence people's decisions. Such images are not neutral — they are systematically biased towards the "ideal norm."
In the future, researchers suggest developing hybrid detection systems, where algorithms will be combined with human expertise. Computers will analyze statistical patterns, while specialists will interpret complex situations. The ability to notice subtle deviations from the norm may become an important skill in the digital age. In conclusion, the study highlights that identifying "fakes" is not only a technological challenge but also a question of adapting human perception to a changing reality.
