Holistic vs Analytic Processing
Face processing research has documented consistent cultural differences in processing style. Western observers typically process faces more analytically, attending relatively more to individual features. East Asian observers process faces more holistically, attending to the overall gestalt with less focused attention on individual features like the eye region.
These differences are visible in eye tracking data: observers from different cultural backgrounds have characteristically different scan paths when viewing faces. They also manifest in performance differences: conditions that disrupt holistic processing (the face inversion effect, the composite face effect) tend to be somewhat smaller in observers with more analytic processing styles.
What This Means for AI
AI face recognition is not culturally variable, it applies the same computation to every face. This is both its strength (consistent) and a potential limitation (may not capture culturally variable aspects of identity perception). The features weighted most heavily by the network, stable mid-face geometry, are relatively culture-independent in how they predict identity.
However, when human observers evaluate celebrity matches and disagree with the AI's ranking, cultural processing style may contribute to the disagreement. An analytic processor may focus on a specific feature that differs; a holistic processor may respond to gestalt impression. The AI's embedding captures a different combination than either.