Humans have been successfully trained to spot AI-generated faces in a study involving researchers at the University of Aberdeen.
AI-generated deepfake faces have become so realistic that it is difficult for people to tell them apart from photos of real humans, contributing to increases in AI-related fraud.
“Training on visual artifacts, like looking for a sixth finger or odd earrings, has had limited success, partly because the AI is getting too good, and fraudsters may avoid using pictures with obvious flaws anyway,’’ said lead researcher Associate Professor Amy Dawel, Director of the Australian National University (ANU) Emotions and Faces Lab.
“Our training directs people’s attention to global qualities that differ between AI and human faces. AI faces tend to be more symmetrical, proportional and attractive, but without training we often think these are markers of being human.”
Now, the team, based in Australia, Canada and the UK, have shown that people can learn to recognise AI faces, given the right exposure, representing an important breakthrough for combating AI-misinformation.
The researchers trained people to spot AI-generated faces by drawing their attention to six perceptual qualities: distinctiveness, memorability, proportionality, symmetry, attractiveness and expressiveness.
The ability of all participants to spot AI faces improved, with “high performers” achieving near perfection.
The UK lead for the project, Dr Clare Sutherland from the University of Aberdeen, said: “This project is particularly exciting because it’s not always been straightforward to train people on face recognition skills.”
“However, over time, as the team worked with the face images, we started to get a feel for which faces ‘looked like’ AI ones. We have also found that 'super-recognisers’, or people who are inherently good with faces, do a slightly better job at discriminating AI faces.”
“Our new work shows that people can indeed learn to recognise AI faces without needing much instruction - instead, we just alert people to some of the key differences between the faces; for example, AI faces tend to look less emotional than real human faces.”
The study, Training Humans to Detect AI-generated Faces, is published in the scientific journal PNAS.
Associate Professor Dawel said: “It was amazing to see the dramatic improvement in people’s ability to detect AI faces. We've shown our training is effective for some of the most convincing fakes available, StyleGAN faces. Now we need to find out whether that training generalises to other AI-generated faces.”
“We are also working on how to optimise the training – making it shorter and ensuring the benefits last over time.’’
Dr Sutherland added:“I find it particularly striking that people’s first impressions of the faces, such as their emotional expressiveness, can give important clues helpful for telling the faces apart, even when it’s not very explicitly clear to people which faces are real or not.”
“We know from our other work at Aberdeen how rich and important first impressions are in many different contexts. However, only some impressions are actually helpful - some mislead!”
ENDS