"I think this is fair": Uncovering the Complexities of Stakeholder Decision-Making in AI Fairness Assessment

Stakeholders prioritize different fairness concerns than AI experts: The study found that non-experts considering a credit rating scenario identified fairness features "far beyond legally protected features," suggesting that affected communities care about broader dimensions of fairness than the standard protected categories (like race or gender) that AI experts typically focus on.

One-size-fits-all fairness metrics don't match real-world needs: Rather than accepting standard fairness metrics, the 26 stakeholders in the study "tailored metrics for specific contexts," indicating that fairness assessment should be customized to particular situations rather than applying uniform technical standards across all AI applications.

Affected people want stricter fairness protections than experts set: Stakeholders "set diverse yet stricter fairness thresholds" compared to typical AI expert practices, suggesting that those impacted by AI decisions are willing to accept lower system performance if it means fairer outcomes—a trade-off that technical experts might not prioritize.

Stakeholders prefer participatory fairness design: The research reveals that affected stakeholders "even preferred designing customized fairness," pointing to a desire for meaningful involvement in how fairness is defined and measured rather than having fairness standards imposed by external experts.

Current AI fairness governance excludes crucial perspectives: The study demonstrates that stakeholder input produces "more complex" fairness assessments than expert-only approaches, suggesting that excluding affected communities from fairness decisions results in incomplete and potentially inadequate fairness protections.