The “false faces” or synthetic faces are now more imperceptible for a human being. A study conducted by the University of Texas asked several hundred people to distinguish real people from faces generated by an algorithm.
The study process is interesting. A first group of 315 subjects were asked, on two portraits side by side, which one was fake. Same request for a second group, of 219 people who were briefly trained in the identification of false faces, in particular the defects left by artificial intelligence in certain places. A third and final group of 233 participants rated the reliability of the 128 images presented to the first two groups, on a scale of 1 to 7.
US authorities refuse to grant copyright to an AI
In their answers, the subjects of the first group obtain less than one correct answer out of two (48.2%). For the second group, better prepared, the percentage increases slightly to reach 59% of correct answers. Finally, the ratings of the last group assign on average a higher reliability score to fake faces (4.82) than to real people (4.48).
Real (R) and synthetic (S) faces best (top and top middle) and worst (bottom and bottom middle) accurately classified. Source: PNAS / Sophie Nightingale and Hany Farid.
” We’re not saying that every image generated is indistinguishable from a real face, but a significant number of them are.» deplore Sophie Nightingalen co-author of the study. She also admits to being concerned about the ease of access to the technologies that make it possible to create these synthetic portraits. She’s not wrong.
Almost two years ago, during the presidential election in the United States, a young American of 17 years old had created a fake profile of a fake candidate on Twitter. The mysterious Andrew Walz sported a profile picture taken from the site thispersondoesnotexist.com, which, as its name suggests, features portraits generated by an artificial intelligence. Twitter will eventually certify his profile before being alerted by media coverage of the deception.
In their conclusions, the two co-authors encourage “ those who develop these technologies wonder whether the associated risks outweigh their benefits. If so, then we discourage the development of a technology simply because it works.. »