Variable Selection in the Context of AI Fairness

Variable selection matters for fairness: The paper demonstrates that how organizations choose which data variables to include in AI systems directly affects fairness outcomes. Removing variables that seem "sensitive" can actually hide bias rather than eliminate it, potentially disadvantaging certain groups.

Excluding data can backfire: Counterintuitively, the research suggests that keeping all relevant variables—rather than removing them—may better reduce implicit bias and allow for more detailed fairness assessments across different subgroups, challenging common assumptions about data minimization.

Ethics and math must work together: The paper emphasizes that fairness in AI cannot be solved by mathematics alone. Developers need to combine technical approaches with philosophical ethics and social awareness to create systems that are both mathematically sound and ethically responsible.

Regulatory compliance requires interdisciplinary teams: With regulations like the EU AI Act now mandating fairness considerations, the authors argue that organizations must bring together mathematicians, ethicists, social scientists, and domain experts to properly evaluate and ensure equitable AI systems.

Current approaches are insufficient: Traditional variable selection methods used in AI development don't adequately account for ethical considerations or social context, meaning many existing systems may inadvertently perpetuate inequities across different demographic groups.