Empowering Affected Individuals to Shape AI Fairness Assessments: Processes, Criteria, and Tools

Affected individuals define fairness differently than experts: The study found that people directly impacted by AI credit rating decisions articulated diverse and nuanced fairness criteria that went beyond the predefined protected attributes and metrics typically used by AI experts and regulators, suggesting current fairness assessments may miss important concerns.

People ground fairness in concrete model features: Participants developed their fairness notions by connecting them to specific, understandable features of the AI system (such as which factors the algorithm considers), indicating that affected individuals can meaningfully engage with technical fairness concepts when given appropriate tools and context.

Both outcome and procedural fairness matter: The research identified that participants cared about two distinct types of fairness—whether decisions were fair in their outcomes and whether the decision-making process itself was fair and transparent—suggesting comprehensive fairness assessment requires attention to both dimensions.

Interactive tools can translate fairness values into measurable criteria: The study demonstrated that an interactive prototype enabled participants to convert their abstract fairness concerns into concrete, quantified criteria, pointing to the feasibility of designing tools that bridge the gap between what people care about and what can be technically measured.

Inclusive fairness assessment requires involving decision subjects: The findings provide empirical support for the emerging call to include affected individuals in AI fairness evaluation, as their participation can reveal fairness considerations that expert-led assessments alone would overlook.