What Is the Own-Race Effect?
The own-race effect, also called the cross-race effect or other-race effect, is one of the most replicated findings in face recognition research. People are more accurate at recognising and distinguishing faces from their own racial group than faces from other racial groups. This effect occurs in memory tests (recognising a face seen previously), in matching tests (deciding whether two photos show the same person), and in eyewitness identification tasks.
The magnitude of the effect is substantial. Across many studies, within-race recognition accuracy averages about 10–15 percentage points higher than cross-race accuracy under matched conditions. This is a large and practically significant difference, with direct implications for eyewitness reliability and for the design of face recognition systems.
Why Does It Occur?
The dominant explanation is the perceptual learning hypothesis: people develop expertise in recognising the face variations most relevant to distinguishing between faces they regularly encounter. People raised in monoracial environments develop fine-grained discrimination ability for the within-group face differences that matter in their social context, at the expense of discrimination ability for face differences characteristic of groups they encounter less.
The effect is not fixed. Extended cross-race contact improves cross-race recognition accuracy. People who have lived in diverse environments, worked extensively with members of another racial group, or completed laboratory training on other-race faces show reduced cross-race effects compared to those without such experience. This plasticity is evidence for the perceptual learning account.
Implications for AI Systems
The own-race effect in humans has a direct analogue in AI face recognition: training data bias. A model trained predominantly on faces from one demographic group will have better-calibrated representations for that group, because the training objective sees more face pairs within the well-represented group and receives more specific gradient signal for their within-group differences.
Ollie's network was trained on MS1MV2, about 5.8 million photos of 85,742 people from many regions, ages and backgrounds. Achieving good performance across all demographic groups requires not just diversity in the training data, but balance, ensuring that the loss function receives comparable numbers of within-group comparison examples from all groups represented in the database.