People were unable to distinguish between artificial intelligence-generated faces and real ones. Not only that, but participants in the experiment even described "photos" of non-existent people as more real than existing people.
It is known that AI has learned to "fake" the faces of real people, creating their "avatars" in situations in which they have never been. Such technology, for example, can be used in the creation of pornographic content "featuring" celebrities. So as of today, any photo or even video can be questioned. And in 2019, researchers at the University of York found that people perceive photos of hyper-realistic masks as more believable than real human faces. This can be exploited by burglars and fraudsters.
A new study by scientists from the Universities of Lancaster (UK) and California (USA), the conclusions of which were presented in the journal PNAS, has shown that artificial intelligence is capable of creating faces which appear more real to humans than real ones. The three experiments involved 315, 219 and 223 participants respectively. In the first, the volunteers classified 128 AI-created faces as real or synthesised. The accuracy, therefore, was 48 per cent.
In the second experiment, the new participants, who were trained to recognize real and artificially created faces, identified the same 128 "photos," more accurately, but not by much. The "hit" percentage rose to only 59 percent. In the third stage, subjects rated the reliability of the 128 images on a scale from one (very unreliable) to seven (very reliable).
As a result, the average score of real faces was 4.48 points lower than that of synthesised and generated AI. That is, participants were 7.7 per cent more likely to trust the latter rather than the former. Researchers concluded that artificially created faces appear more real to people than natural faces. Interestingly, images of women in the participants elicited more trust than photos of men. The same was true for black faces compared to South Asian faces. Smiling faces were also given higher credibility ratings than those whose facial expression was neutral.