What are the major problems with the 'hyperrealism' phenomenon, in which AI creates faces that look more realistic than actual human faces?

AI can generate photos of non-existent people's faces, and the accuracy of this is increasing at an incredible rate. There have been reports of suspected cases

of AI-made spies and non-existent engineers being used, and some research has shown that AI-generated faces are not only indistinguishable from real faces, but are also ' more reliable than real faces .' Amy Dowell, a senior lecturer at the Australian National University, has shown that 'AI can create faces that look more real than real human faces,' and explains the AI's ability to 'hyperrealism.'

AI Hyperrealism: Why AI Faces Are Perceived as More Real Than Human Ones - Elizabeth J. Miller, Ben A. Steward, Zak Witkower, Clare A. M. Sutherland, Eva G. Krumhuber, Amy Dawel, 2023

AI faces are 'more real' than human faces — but only if they're white | Live Science

It is known that when a human likeness reaches a certain level of quality, the ' uncanny valley' phenomenon occurs, in which disgust rises sharply. On the other hand, some research has shown that AI can overcome the 'uncanny valley' and generate faces that are so good that people feel that 'AI-generated faces are more trustworthy.'

Sophie Nightingale, a psychologist at Lancaster University, and Hany Farid, an information scientist at the University of California, Berkeley, used GANs , a type of artificial intelligence algorithm, to create realistic fictional faces that are indistinguishable from real faces. As a result of conducting tests in which the generated photos were mixed with real face photos and had them identified, it seems that depending on the group, the AI-generated photos were often judged to be 'real faces'.

AI-generated faces are indistinguishable from real faces and are more reliable than real faces - GIGAZINE

Based on this research, Dowell named this phenomenon, in which AI-generated faces seem more real, 'hyperrealism.' As a step further, he investigated whether racial factors influence the hyperrealism phenomenon.

In a study published in November 2023, participants were shown a total of 100 photos of faces, consisting of 'fictional humans generated by AI' and 'real humans.' After determining whether the photo was AI or human, participants rated their own opinion of their choice on a scale of 0 to 100.

In the images below, the top row is the 'image most judged to be' human ', and the bottom row is the 'image most judged to be' AI '. Of these, the top three from the left, which were judged to be the most human, were images generated by AI.

The results of the study showed that AI can only achieve hyperrealism when generating white people. AI-generated faces of people of color still fell into the uncanny valley, making it easy to distinguish between real and fictitious faces. Dowell explained the reason for this: 'The reason is simple: the learning data for AI algorithms is unbalanced. The faces that are trained are overwhelmingly white, and this bias means that only white faces can be made realistic.'

Dowell also found another social problem with AI from his research. According to the study, people who repeatedly misjudged AI-generated faces as 'real human faces' were also more likely to be confident in their choices. 'In other words, the people who are most fooled by AI are the least aware that they are being fooled. Generative AI is likely to be able to fool such people, so this shows that it should be developed with increased transparency and monitored by an independent organization,' Gwell argued.

'Racial bias in AI systems reinforces racial bias in the media people consume,' says AI expert Frank Buitendijk. 'This could have serious implications on a societal level, such as making teenagers feel that AI-generated looks are stereotypical and therefore ideal.'

in Software, Posted by log1e_dh