Measuring Epistemic Unfairness for Algorithmic Decision-Making

Researchers Camilla Quaresmini, Lisa Piccinin, and Valentina Breschi have introduced a groundbreaking quantitative framework for measuring epistemic unfairness in algorithmic systems. Their work, published on arXiv (2604.22675), addresses a critical gap in AI auditing: current fairness metrics focus on predictive outcomes like error rates and group parity, but ignore epistemic harms—damage to how individuals are perceived, believed, and empowered to participate in knowledge creation. The framework models these harms as deficits across three key features: credibility (whether an algorithm's outputs are trusted), uptake (how information is absorbed), and epistemic agency (the ability to influence knowledge systems).

The researchers propose two evaluation stances: resource inequality, which applies distributive fairness indices directly to epistemic goods like information access, and capability/rights inequity, which measures how algorithmic outputs constrain users' epistemic opportunities. They translate canonical fairness indices into epistemic terms, enabling detection of issues like exclusionary tails (where certain groups are systematically ignored) and hierarchical concentration (where a few voices dominate). A simulation study of recommender-mediated opinion dynamics demonstrates how these indices capture evolving unfairness over time, even when standard fairness constraints are satisfied. This framework makes epistemic harms explicit for system designers and auditors, offering a practical tool for longitudinal evaluation under iterative deployment.