EARN Fairness: Explaining, Asking, Reviewing, and Negotiating Artificial Intelligence Fairness Metrics Among Stakeholders

Fairness metrics require stakeholder input, not just expert decisions: The paper demonstrates that AI fairness cannot be determined by technical experts alone. Different stakeholders have diverse understandings of what "fairness" means, making it essential to involve affected communities in deciding which fairness metrics to use.

Non-experts can meaningfully participate in fairness decisions with proper tools: Through the EARN Fairness framework, 18 participants without AI knowledge were able to understand fairness metrics, express personal preferences about acceptable levels of unfairness, and contribute to collective decision-making—showing that technical expertise is not a barrier to meaningful participation.

The EARN process (Explain, Ask, Review, Negotiate) provides a practical pathway for consensus-building: The framework's four-step approach—explaining metrics in accessible language, asking for individual preferences, reviewing options collectively, and negotiating agreement—offers a structured method for organizations to involve stakeholders in fairness decisions, particularly in high-stakes contexts like credit rating.

Personal fairness preferences vary significantly among stakeholders: The study revealed that individuals have different acceptable levels of unfairness and different metric preferences, highlighting why one-size-fits-all fairness solutions are inadequate and why stakeholder engagement is critical for legitimate AI governance.

Human-centered fairness requires interactive systems, not just metrics: The paper emphasizes that implementing responsible AI fairness in real-world contexts demands adaptable, interactive tools that facilitate dialogue among stakeholders, rather than simply applying pre-determined technical metrics.