Artificial intelligence (AI)-generated white faces now look more realistic than human faces, according to new research co-authored by a University of London, Los Angeles (UCL) academic. In the study, led by researchers at the Australian National University and published in Psychological Science, more people rated AI-generated faces as more human-like than real faces.
Co-author Dr Eva-Krumhuber (UCL Psychology & Language Sciences) said: "AI has reached astonishing levels of realism, and here we found that sometimes it even looks more real than reality - hyper-realistic - so we can easily be tricked into thinking the AI-generated faces are real."
In the study, the researchers showed 124 participants different images of white faces and asked them to judge whether the faces were real or generated by the StyleGAN2 algorithm. For the AI faces, participants judged them to be real two-thirds of the time - more often than real faces.
Racial bias and perception of AI faces
However, this pattern was not found in other studies using the same algorithm and including faces of different races. The reason for this disparity is that AI algorithms tend to be trained disproportionately on white faces.
Dr Amy Doyle (Australian National University), the study's lead author, said: "If white AI faces are consistently perceived as more realistic, then the technology could end up reinforcing racial bias online, with serious consequences for people of color." This problem is already evident in AI technology currently used to create professional-looking avatars. When used on people of color, the AI modifies their skin tone and eye color to that of white people."
Artificial Intelligence’s Hyperrealism and Its Consequences
Researchers have found that one of the problems with "hyper-realistic" artificial intelligence is that people often don't realize they are being fooled.
Elizabeth Miller, co-author of the study and a doctoral student at the Australian National University, said: "Worryingly, people who believe that the AI face is the most realistic tend to be the most confident that their judgment is correct. This means that people who mistake the AI impostor for a real person do not know that they have been deceived."
Researchers also discovered why AI faces can fool people.
Dr. Krumhuber said: "In the past, there were often inconsistencies between artificial and human-like faces, creating the uncanny valley effect. In particular, the eyes (i.e. refraction, shadows) gave away clues as to whether the face was real or artificial. It now appears that we have overcome the valley effect in static images."
Dr Doyle added: "It turns out that there are still physical differences between AI and human faces, but people tend to misunderstand these differences. For example, white AI faces tend to be more disproportionate and people mistake them for being a sign of humanity. However, we cannot rely on these physical cues for long. AI technology is developing rapidly, and the differences between AI and human faces may soon disappear."
Impact on misinformation and identity theft
Researchers believe this trend could have serious implications for the spread of misinformation and identity theft, necessitating action.
Dr Doyle said: "AI technology cannot be compartmentalized and only technology companies know what is going on behind the scenes. There needs to be greater transparency around AI so that researchers and civil society can detect problems before they become big problems."
Researchers believe that raising public awareness can also play an important role in reducing technology risks.
Dr. Klumhuber, who researches and teaches at UCL, explores the social cognitive and emotional processes behind emerging technologies and their impact on the human psyche: "Given that humans are no longer able to recognize the face of AI, society needs tools that can accurately identify AI impostors."
Dr Doyle concluded: "Educating people about the perceived authenticity of AI faces can help instill appropriate skepticism in the public about the images they see online."
Reference: AI Hyperrealism: Elizabeth J. Miller, Ben A. Steward, ZakWitkower, Clare A. M. Sutherland, Eva G. Krumhuber, and Amy Dawel, "Psychological Science", November 12, 2023.
doi:10.1177/09567976231207095
Compiled source: ScitechDaily