New Training Method Significantly Improves Ability to Detect AI-Generated Faces

Distinguishing between real human faces and those generated by artificial intelligence is becoming increasingly difficult, according to a new study from Australian National University (ANU). Researchers have found that while the average person is no better than chance at spotting AI-generated faces, a novel training method can significantly enhance this ability.

The study, published in the journal PNAS, highlights that AI-generated faces are rapidly improving, making them nearly indistinguishable from real ones. This advancement fuels a rise in AI-powered fraud, with projected losses in the United States alone reaching $40 billion by 2027. Traditional advice, such as looking for 'AI artifacts' like unusual numbers of fingers or misaligned teeth, is becoming outdated as AI can easily avoid or correct these errors.

Instead of focusing on specific flaws, the ANU researchers developed a training method that teaches individuals to rely on 'global impressions.' Participants were shown labeled images of both AI-generated and genuine human faces and asked to rate them based on six characteristics: facial distinctiveness, memorability, proportionality, symmetry, attractiveness, and expressiveness. The goal was not to teach explicit rules, but to cultivate an intuitive sense for what looks 'right.'

IMAGE4

Before the training, participants correctly identified AI imposters hidden among two real humans only 41% of the time. Their accuracy in identifying a single human face as real was 52%, and an AI-generated face was correctly labeled with 47% accuracy. However, after a brief online training session utilizing the new method, average accuracy in spotting AI fakes doubled. Some participants even achieved near-perfect results.

IMAGE3

These findings were replicated by a team at the University of Victoria in Canada, led by Professor Jim Tanaka and Dr. Eric Mah. Dr. Mah noted that the replication confirmed the effectiveness of the training, demonstrating significant improvement in a new group of participants in a different country. The low cost and ease of online implementation suggest the training program could be scaled effectively.

IMAGE2

The researchers explain that this method works because facial impressions are formed rapidly and intuitively, and are sensitive to the systematic biases inherent in AI algorithms. While algorithmic deepfake detection tools exist, they can be complex and opaque. The study argues for an urgent need to improve human AI-detection abilities to combat the growing threat of deepfake scams.