Formalising Anti-Discrimination Law in Automated Decision Systems

Legal standards matter more than technical metrics alone: The paper reveals that existing "fairness metrics" in AI often miss what actually constitutes discrimination under law. Simply measuring prediction differences between groups isn't enough—UK anti-discrimination law requires examining whether decisions are justified by legitimate business reasons, a distinction most AI fairness research overlooks.

We don't know what we don't know about our data: The authors highlight a critical blind spot: most fairness approaches assume we fully understand how data was generated and what causes outcomes, but this is rarely true in practice. The paper's "conditional estimation parity" metric accounts for this uncertainty, making it more realistic for real-world deployment than current methods.

US-centric AI regulation creates global problems: The paper documents how algorithmic fairness research has been dominated by US legal frameworks, leading to misconceptions about what's lawful elsewhere. Since UK and European anti-discrimination law differs significantly, companies and AI developers following US-based guidance may unknowingly violate laws in other jurisdictions.

Technical and legal expertise must work together: The authors demonstrate through a real case study that detecting unlawful discrimination requires both machine learning specialists and legal experts collaborating from the start, not as an afterthought. Current practice often separates these disciplines, allowing discrimination to slip through.

Actionable frameworks now exist for practitioners: Rather than abstract theory, this work provides concrete guidance that ML teams, policymakers, and legal professionals can actually use to build and audit systems for compliance with anti-discrimination law.