Aberdeen researchers demonstrate training doubles AI face detection accuracy

A multinational team of researchers—including scholars from the University of Aberdeen, the Australian National University, the University of Victoria and the University of New South Wales—found that a brief, structured training session can dramatically improve people's ability to identify AI‑generated faces. In a pre‑post study of 45 participants, accuracy in spotting synthetic faces rose from roughly 40 % before training to nearly 80 % after exposure to examples and instruction.

The training focused on six perceptual cues that remain harder for current generators to mimic consistently: facial symmetry, proportionality, above‑average attractiveness, generic or non‑distinctive structure, limited emotional expressiveness, and low memorability. Participants learned to evaluate these qualities rather than relying on obvious visual glitches such as extra fingers or distorted accessories, which modern models like StyleGAN3 have largely eliminated.

Researchers note that improving human judgment is crucial as deepfake technology is increasingly employed in financial fraud, political influence operations and identity scams. The study underscores that while AI detectors continue to evolve, human observers equipped with the right training can provide a valuable line of defence against increasingly realistic synthetic media.