Fairness Explainability using Optimal Transport with Applications in Image Classification

Understanding why AI systems are biased, not just fixing outcomes: The paper addresses a critical gap in AI fairness research by focusing on explaining the root causes of discrimination rather than just correcting biased predictions. This distinction matters because understanding why a model discriminates helps prevent similar problems in future systems.

Identifying which image features drive unfair decisions: Using a mathematical technique called Wasserstein barycenters, the researchers can pinpoint specific regions or features in images that the AI system relies on when making biased decisions. This visual explainability is particularly valuable for image classification systems used in high-stakes domains like hiring, lending, or criminal justice.

Connecting fairness enforcement with explainability: The paper proposes an integrated approach that simultaneously enforces fair predictions while measuring how individual features contribute to bias. This dual approach creates a feedback loop where fairness improvements directly reveal which aspects of the training data or model are problematic.

Implications for trustworthy AI in critical decisions: The authors emphasize applications across "diverse domains" and "critical decision-making scenarios," suggesting this method could enhance accountability in sensitive areas where AI bias has documented harms—such as facial recognition systems, medical imaging, or employment screening tools.

Addressing human bias in training data as a root cause: The paper explicitly acknowledges that human biases embedded in training data are a "substantial" problem that current fairness methods often overlook, positioning this work as a step toward more transparent identification of where discriminatory patterns originate.