According to researchers, people can no longer reliably distinguish between a real human face and an AI-generated face image. Sophie Nightingale of the Department of Psychology at Lancaster University in England and Hany Farid of the Department of Electrical Engineering and Computer Science at Berkley, California, studied human evaluations of both real photographs and images synthesized by artificial intelligence, and now no one can reliably tell the difference. came to the conclusion.
In one part of the study published in the Proceedings of the US National Academy of Sciences, people were able to detect fake images in only 48.2 percent of trials.
In another part of the study, participants were given some training and feedback to help them detect fakes. While the team that received the feedback detected real people in 59 percent of the trials, the results remained stable at this point.
In the third part of the study, participants rated faces as “reliable” on a scale of one to seven. Fake faces were rated as more reliable than real faces.
The researchers concluded, “A smiling face is more likely to be rated as trustworthy, but 65.5 percent of our real faces and 58.8 percent of synthetic faces are smiling, hence why facial expression alone is why synthetic faces are more reliable. It doesn’t explain how it’s rated,” he wrote.
The fake images were created using generative adversarial networks (GANs), a class of machine learning frameworks where two neural networks play some kind of competition with each other until the network trains itself to create better content.
This method starts with a random array of pixels and iteratively learns to create a face. Meanwhile, a splitter learns to detect the synthesized face after each iteration. Finally, the process is complete when the discriminator cannot tell the difference between real and synthesized faces. And apparently a human can’t tell the difference either.
The final images used in the study included a diverse array of 400 real and 400 synthesized faces representing Black, South Asian, East Asian and White faces. In contrast to previous studies that used primarily white male faces, male and female faces were included.
White faces were the least accurately classified, and male white faces were detected even less accurately than female white faces.
“We hypothesize that white faces are more difficult to classify as they are overrepresented in the StyleGAN2 training dataset and are therefore more realistic,” the researchers write.
Scientists acknowledge that while creating realistic faces is a success, it poses potential problems such as deep fakes. They say such activities have “serious implications for individuals, societies and democracies”.
4380
