Modeling Epistemological Principles for Bias Mitigation in AI Systems: An Illustration in Hiring Decisions

Accuracy alone is insufficient for responsible AI: The paper demonstrates that traditional AI design focused only on maximizing accuracy can lead to "economically irrational and socially unacceptable scenarios of discrimination and unfairness," highlighting that hiring systems optimized purely for predictive accuracy may still discriminate against protected groups.

Biased training data is a hidden source of discrimination: The research identifies that AI systems trained on biased datasets can perpetuate unfairness in ways that remain undetected in "black-box models"—systems whose decision-making processes are opaque—making it critical to scrutinize the data used to train hiring algorithms before deployment.

A structured testing approach can reduce bias: Rather than accepting an AI model as-is, the paper proposes selecting and improving models through systematic attempts to identify and refute unfair outcomes, similar to how scientific theories are tested; this means hiring systems should be repeatedly challenged and refined to catch discriminatory patterns.

Fairness requires explicit design choices: The paper emphasizes that fairness in AI "must be explicitly addressed" rather than assumed to follow from technical optimization, suggesting that organizations using AI for hiring must actively build fairness considerations into their system design from the start.

Epistemological principles offer a rational framework: By drawing on philosophical logic about how we validate knowledge, the paper provides a principled method for choosing between AI models based on both accuracy and fairness, offering decision-makers a structured way to justify their AI choices on rational grounds.