Title: Position: Moving Past Sensitive Attributes, ML Fairness Must Measure Structural Injustice Through Social Determinants

Abstract:

Current algorithmic fairness research predominantly defines unfairness as discrimination based on sensitive attributes. Yet, this narrow framing obscures the reality of structural injustice, which emerges through social determinants—contextual factors that influence both attributes and outcomes without being tied to specific individuals. This position paper contends that the field must expand its scope to quantify structural injustice via these social determinants, rather than relying exclusively on sensitive attributes. Leveraging interdisciplinary perspectives, we argue that existing technical frameworks are insufficient for capturing structural injustice because they often treat contextual variables as noise to be smoothed out, rather than as critical signals requiring audit.

To illustrate the practical necessity of this paradigm shift, we present a theoretical model of college admissions, a demographic analysis utilizing U.S. Census data, and a case study on breast cancer screening within an integrated U.S. healthcare system. Our findings reveal that mitigation efforts focused solely on sensitive attributes may inadvertently generate new forms of structural injustice. We assert that auditing for structural injustice through social determinants must occur before any mitigation takes place. Furthermore, we urge the development of novel technical approaches that transcend the traditional concept of fairness as mere non-discrimination based on sensitive attributes.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC