Recognizing Bias in Facial Recognition

This paper examines the dangerous intersection of racial bias and facial recognition technology (FRT) in law enforcement. Through the case study of Robert Williams—a Black man wrongfully arrested in Detroit due to a false algorithmic match—the author illustrates how FRT systems, which are trained on databases like driver’s license photos, frequently misidentify people of color. Despite procedural safeguards, such as human oversight, these systems often fail because human checkers may harbor their own biases or succumb to 'technochauvinism,' the blind belief that computer-generated results are infallible.

Research by Dr. Joy Buolamwini, specifically the 'Gender Shades' project, provides empirical evidence that FRT is significantly less accurate for individuals with darker skin tones compared to those with lighter skin. The author argues that these disparities stem from insufficiently diverse training datasets. While some argue that diversifying data can 'fix' the technology, the author contends that this approach is insufficient. Improving accuracy merely makes surveillance more efficient at targeting marginalized communities, thereby exacerbating existing social inequalities.

The widespread adoption of FRT by law enforcement agencies, often justified by large taxpayer investments in surveillance infrastructure, has profound consequences for civil rights. The author argues that because these systems have historically failed to provide reliable results and are inherently prone to racial bias, the only truly antiracist approach is to ban the use of facial recognition in policing entirely. The paper calls for a shift from a 'nonracist' stance—which claims neutrality while maintaining the status quo—to an 'antiracist' stance that actively challenges the policies and tools that perpetuate systemic harm.