With artificial intelligence, it is not easy to distinguish real from fake. Especially when this technology is used to generate ultra-realistic faces. Researchers at the University of Lancaster have found that it is extremely difficult for us to distinguish a real face from another face created from scratch by algorithms.
They carried different experiments to determine how realistic these synthetic faces are. The researchers first asked 315 participants to identify, among 800 portraits, the faces of real people and those generated by artificial intelligence. They achieved an accuracy of 48.2%, slightly lower than by chance.
The second group of participants was trained to recognize artificial faces before taking the same test. Their result? 59% accuracy.
These mediocre scores aren’t surprising given the sophistication of the artificial intelligence software used to create ultra-realistic faces with just a few clicks. They are so good that some people use them for entertainment or parody purposes.
One of them liked to have Elon Musk appear in Stanley Kubrick’s movie, 2001: A space odysseywhile another made one deep fake of Barack Obama calling Donald Trump a “dips***”.
This type of scenario gets a lot of attention and raises a lot of questions about using deep fakes to spread false information. Lancaster University researchers wanted to test whether synthetic faces elicited a sense of trust in viewers. They asked 223 participants to rate the reliability of real and synthetic faces by giving them a score from 1 to 7.
The faces generated by artificial intelligence are considered to be 7.7% more reliable than human faces on average. A disturbing phenomenon that the researchers explained by the fact that the synthetic faces looked rather “average”. This impression of being ordinary creates a sense of confidence in most people.
“Our evaluation of the photorealism of AI synthesized faces indicates that synthesis engines have moved through the uncanny valley [a theory by roboticist Masahiro Mori that affinity for robots increases with their resemblance to humans, ed. note] and are able to create faces that are indistinguishable – and more reliable – than real faces,” said the study, published recently in the scientific journal PNAS.
So caution is advised when you come across a photo or a video on the Internet. Especially the ones made with the Chinese application Zao. It allows users to insert their face into movie or video clips in seconds with a simple selfie. The result is sometimes impressively realistic. Proof that it’s more important than ever on the internet to be wary of appearances.