Researchers test whether people can learn to spot AI faces

Psychologist Dr Clare Sutherland at the University of Aberdeen and Prof Amy Dawel at the Australian National University have been studying whether people can be trained to identify AI-generated human faces. Older clues such as extra fingers or mismatched accessories are becoming less useful as image generators improve, pushing researchers toward more subtle signals.

The team created a pool of thousands of AI-generated faces using StyleGAN3 and trained participants to assess six perceptual qualities: symmetry, proportionality, attractiveness, distinctiveness, expressiveness and memorability. AI faces were described as often too symmetrical, more pleasant-looking, more generic, less emotionally expressive and harder to remember. The technology was also said to be less proficient at recreating non-white, older or younger faces because more of its training involves young white people.

After exposure to real and AI faces with feedback, participants typically increased their accuracy score from about 40% to 80%, with a few individuals achieving close to 100% accuracy. The researchers also found that confidence improved after training. The stakes include fraud and political espionage: Deloitte has predicted US losses from AI deepfake scams could rise to £40bn next year, up from £12bn in 2023, and a Hong Kong-based firm employee transferred £25m after a deepfake video call.