A Unifying Human-Centered AI Fairness Framework

There is no single "fair" AI system: The paper identifies eight distinct fairness metrics that can conflict with each other, meaning organizations must deliberately choose which fairness priorities matter most for their specific context rather than assuming one technical solution works universally.

Fairness frameworks need to be understandable to non-experts: The authors emphasize using "consistent and easy-to-understand formulation for all metrics to reduce the learning curve for non-experts," recognizing that fairness decisions should involve diverse stakeholders—not just data scientists—in choosing how to balance competing values.

Multiple perspectives on fairness must be considered simultaneously: The framework combines individual fairness (treating similar people similarly), group fairness (equal outcomes across demographic groups), and intersectional fairness (considering overlapping identities), acknowledging that fairness is multidimensional and requires examining AI systems from different angles.

Real-world applications reveal practical trade-offs: Testing on actual datasets used in high-stakes decisions (criminal justice, credit, healthcare, employment) demonstrates that achieving fairness in one dimension often requires accepting trade-offs in others, making transparent stakeholder input essential for deployment decisions.

Stakeholders should explicitly assign priorities rather than hide behind technical defaults: The framework enables organizations to assign weights across multiple fairness objectives, making value judgments visible and enabling "multi-stakeholder compromises" rather than letting technical choices implicitly determine whose interests are protected.