Attractiveness shapes beliefs about whether faces are real or AI-generated, study finds

A new study published in Acta Psychologica sheds light on how people decide whether a face is real or artificially generated—and the answer may have more to do with personality and perception than objective clues. Despite viewing only authentic photographs of real people, most participants in the study confidently labeled many faces as fake. Their judgments were strongly influenced by subjective impressions of attractiveness and trustworthiness, as well as personality traits like narcissism and honesty-humility.

As artificial intelligence becomes more capable of producing realistic images, especially of faces, the boundaries between what is real and what is synthetic are becoming harder to identify. In this rapidly changing landscape, understanding how people determine the authenticity of images is increasingly important. The researchers sought to explore which psychological factors drive these judgments, especially in the absence of clear visual markers.

“Witnessing the incredible acceleration of the technological ability to create artificial content (such as pictures of people) that are – or will soon become – indistinguishable from their ‘genuine’ counterparts, led us to two observations,” said study author Dominique Makowski, a lecturer in psychology at the University of Sussex and head of the Reality Bending Lab.

“The cyberspace (the digital environment created by social media, the internet, and virtually mediated communication) we navigate will soon become filled with content of ambiguous nature, for which the certainty of knowing if it is real and authentic is not guaranteed. At the same time, knowing that something is real or fake is epistemologically important and carries consequences for adaptive psychological processes such as the formation of trust, beliefs, memorization, and decision making.”

The research team recruited 150 participants to complete a multi-part study. First, participants filled out a series of personality questionnaires measuring traits such as narcissism, honesty-humility, paranoid ideation, and attitudes toward artificial intelligence. Then they were shown 109 images of real faces drawn from the American Multiracial Face Database. The images were carefully selected to feature individuals with neutral expressions, representing a diverse range of racial backgrounds.

Participants rated each face on four characteristics: attractiveness, beauty, trustworthiness, and familiarity. Later, they were told that about half of the images they had seen were AI-generated, though this was not true—all the images were real. The same faces were shown again in a randomized order, and this time, participants were asked to judge whether each face was real or fake, and how confident they were in their judgment.

Surprisingly, across all participants, about 44% of the images were judged to be fake—even though every face was a real person. The judgments varied not only between individuals but also within the same person across different faces. This variability suggests that judgments about authenticity are influenced by subjective impressions rather than objective features of the images.

One of the most important takeaways from the study is how easily people can be led to believe that real faces are fake. Simply telling participants that some of the images might be AI-generated was enough to drastically alter their beliefs. Despite having seen the faces already and rated them on other dimensions, participants changed their judgments when primed to expect fakes.

“People’s beliefs about what is real are surprisingly flexible,” Makowski told PsyPost. “People were easily led to the belief that ‘real’ images were AI-generated, and the decision as to which pictures were fake was highly variable from individual to individual, suggesting the role of subjective (idiosyncratic) processes.”

One of the most consistent findings was the role of facial attractiveness. For men, faces rated as more attractive or beautiful were more likely to be judged as real. For women, the relationship was more complex. Both highly attractive and highly unattractive faces were more likely to be judged as real, suggesting a U-shaped relationship. Confidence in these judgments also followed a similar pattern for women, with the highest confidence seen in responses to the most and least attractive faces.

“Judgments of facial attractiveness, which are made automatically and unconsciously, influence whether we believe that an image was AI-generated,” Makowski said. “This influence might be different for men and women.”

Trustworthiness also influenced judgments, particularly for female participants. Women were more confident in judging faces as real or fake when the faces were rated as highly trustworthy or untrustworthy. However, there was no strong evidence that trustworthiness directly affected whether a face was judged as real or fake for men.

Familiarity had weaker effects overall. However, among men, faces that seemed familiar were more likely to be judged as real, and participants felt more confident in these decisions. This finding supports past research showing that familiarity can enhance the feeling that something is genuine, although the effect here was modest.

Beyond face-related judgments, the researchers also examined how personality traits influenced reality beliefs. Narcissistic traits, particularly those related to seeking recognition and manipulating others, were associated with a higher likelihood of judging faces as real. These individuals were also more confident in their judgments. This fits with past research showing that narcissistic individuals often display greater self-assurance and are more likely to overestimate their abilities, including their skill in detecting authenticity.

In contrast, people scoring high in honesty-humility—a personality trait associated with sincerity and modesty—were less confident in their judgments. These individuals may have been more cautious in their responses, reflecting a general reluctance to make bold claims without certainty.

There was also a tentative link between paranoid ideation and a greater tendency to judge faces as real. The researchers speculated that heightened sensitivity to social cues, common in individuals with paranoid thinking, might increase the perceived emotional salience of the faces, leading to stronger impressions of authenticity.

Interestingly, general attitudes toward artificial intelligence had only limited effects. The only factor that mattered was participants’ enthusiasm about AI, which was linked to greater confidence in their judgments. However, beliefs about how realistic AI-generated content can be did not significantly influence decisions about whether faces were real or fake.

The researchers also found that the order in which images were shown had little effect on belief judgments, although participants were less confident in their decisions toward the end of the session. The delay between first viewing the face and making a real-fake judgment had no effect on the likelihood of calling a face real or fake, but it did affect confidence—participants were less confident when there was a longer delay.

There were some limitations to the research. All of the faces were shown twice, which could have introduced familiarity effects. The researchers also gave participants a false instruction that half of the images were fake, which may have created a bias toward identifying some faces as AI-generated. Although the instruction did not specify how many faces should be labeled real or fake, it may have led participants to second-guess their initial perceptions. Future research could use different designs to minimize these influences, such as showing new faces in the second phase or using a more implicit measure of authenticity judgment.

Another limitation was that the effects of facial ratings and personality traits, while statistically significant, were relatively small. This suggests that other factors—perhaps less conscious or more situational—also play a role in how people judge the authenticity of faces. The researchers recommend future studies explore additional cues, such as specific facial features or expressions, and examine how people consciously or unconsciously use these features to make reality judgments.

Looking ahead, Makowski said he plans to “confirm and expand these findings for other types of materials, such as art.” He added, “There will be more coming soon! Stay tuned by checking the Reality Bending Lab’s website.”

The study, “Too beautiful to be fake: Attractive faces are less likely to be judged as artificially generated,” was authored by Dominique Makowski, An Shu Te, Ana Neves, Stephanie Kirk, Ngoi Zi Liang, Panagiotis Mavros, and S.H. Annabel Chen.