Study Finds Training Improves Ability to Detect AI-Generated Faces

Distinguishing between human and artificial intelligence-generated faces is becoming increasingly difficult, according to a study from the Australian National University (ANU). The research suggests that while the average person struggles to tell the difference, training can significantly improve detection abilities.

AI-generated faces are becoming so sophisticated that traditional methods of spotting them, such as looking for 'AI artifacts' like misaligned teeth or extra fingers, are no longer effective. These errors are easily corrected by AI algorithms, rendering advice based on them outdated. The study, published in the journal PNAS, highlights that this advancement in AI capabilities is contributing to a rise in AI-powered fraud, with projected losses in the United States alone reaching $40 billion by 2027.

Researchers found that a new training method, focusing on 'global impressions' rather than specific features, effectively sharpens natural intuition. Participants were exposed to both real and AI-generated faces and asked to rate them on six criteria: facial distinctiveness, memorability, proportionality, symmetry, attractiveness, and expressiveness. This process was not about teaching explicit rules, but rather about developing an intuitive sense through repeated exposure and practice.

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

This training method's effectiveness was further validated by a replication study at the University of Victoria, Canada. Researchers there observed similar improvements in participants, suggesting the findings are robust and the training program can be implemented at scale with minimal cost. The ANU team argues that improving human AI-detection abilities is urgently needed to counter deepfake scams, as current algorithmic detection methods can be opaque and have potential hidden flaws.

The principle behind the training is that humans form rapid, intuitive facial impressions that are sensitive to the systematic biases inherent in AI algorithms. By directing attention to broader, global characteristics, individuals can hone their innate ability to discern whether a face 'looks right,' a skill that often goes unleveraged without specific training. This approach aims to equip people with a more reliable defense against increasingly convincing AI-generated imagery.