Search for: All records

We applied techniques from psychology --- typically used to visualize human bias --- to facial analysis systems, providing novel approaches for diagnosing and communicating algorithmic bias. First, we aggregated a diverse corpus of human facial images (N=1492) with self-identified gender and race. We tested four automated gender recognition (AGR) systems and found that some exhibited intersectional gender-by-race biases. Employing a technique developed by psychologists --- face averaging --- we created composite images to visualize these systems' outputs. For example, we visualized what an average woman looks like, according to a system's output. Second, we conducted two online experiments wherein participants judged the bias of hypothetical AGR systems. The first experiment involved participants (N=228) from a convenience sample. When depicting the same results in different formats, facial visualizations communicated bias to the same magnitude as statistics. In the second experiment with only Black participants (N=223), facial visualizations communicated bias significantly more than statistics, suggesting that face averages are meaningful for communicating algorithmic bias.