Applications of Risk Science to AI Fairness Evaluation: Principles, Challenges, and Best Practices

Most AI fairness studies fail to adequately communicate uncertainty: The paper analyzed 22 common fairness metrics and found that while most studies characterize how severe bias problems are, they typically don't explain the uncertainty around whether these problems will actually occur or how confident we should be in severity estimates—a critical gap that risk science has long established as essential for decision-making.

Current AI evaluation practices don't follow established risk assessment principles: By comparing AI fairness scholarship to decades of best practices in risk science (which systematically studies how to understand, assess, and communicate risks), the authors found that AI evaluation work largely ignores these proven methodologies, potentially leading to incomplete or misleading conclusions about real-world harms.

Hiring and employment AI systems are particularly under-evaluated: The paper's detailed examination of fairness studies for AI-mediated hiring and resume screening revealed significant gaps in how bias risks are characterized, suggesting that workers may be affected by algorithmic decisions without adequate understanding of the actual risks involved.

Better communication tools are needed for stakeholders: The authors propose an "AI Risk Report Card" designed to help policymakers, companies, and other decision-makers understand and act on fairness risks—suggesting that current reporting methods don't adequately translate technical findings into actionable information for those responsible for managing AI systems.

Risk science offers proven solutions: The paper demonstrates through a case study that incorporating established risk science principles into AI fairness evaluation is feasible and produces more complete assessments of potential harms.