Algorithmic Unfairness through the Lens of EU Non-Discrimination Law: Or Why the Law is not a Decision Tree

Legal and technical definitions of fairness don't automatically align: The paper demonstrates that computer science notions of "algorithmic fairness" and EU legal concepts of "discrimination" operate under different frameworks and assumptions, meaning a system deemed "fair" by technical metrics may still violate non-discrimination law, and vice versa.

Fairness metrics alone cannot ensure legal compliance: The authors argue that technical fairness interventions have inherent limitations in addressing discrimination as defined by law. Legal compliance requires understanding the broader context and intent behind discrimination prohibitions, which cannot be reduced to mathematical formulas or decision trees.

Both fields make implicit normative assumptions that need to be explicit: The paper reveals that both computer scientists developing fairness metrics and EU courts applying non-discrimination law operate based on unstated value judgments about what equality means. Making these assumptions explicit is essential for practitioners and regulators to understand what they're actually measuring or enforcing.

Real-world discrimination cases reveal gaps in current AI practice: By analyzing actual EU case law examples through an algorithmic lens, the paper shows that current technical approaches to fairness often miss important dimensions of discrimination that the law recognizes, suggesting AI systems need to be evaluated against legal standards, not just technical benchmarks.

Regulators need interdisciplinary guidance: The paper's conclusion implies that effective AI governance requires collaboration between legal experts and technologists to bridge the conceptual gap, rather than allowing either discipline to independently define what constitutes fair and non-discriminatory AI systems.