Overview
- UNSW and ANU published findings in the British Journal of Psychology showing that laypeople are only slightly better than chance at telling AI-generated faces from real ones.
- The study tested 125 participants, including 36 super-recognizers, using an online task that screened out images with obvious flaws to focus on highly realistic outputs.
- Super-recognizers outperformed other participants by a modest margin, yet their accuracy remained far below their typical performance with real faces and overlapped with non-experts.
- Advances in face-generation have removed many telltale artifacts, with the most convincing synthetic faces tending to look unusually average, symmetrical, and well-proportioned.
- The authors urge skepticism about unaided visual judgments for identity checks across social media, dating, and recruitment, and plan to study potential 'super-AI-face-detectors' to see if their strategies are teachable.