Machine learning (ML) systems automating high-stakes decisions in hiring, finance, healthcare and other high-stakes domains and are facing intensifying legal scrutiny under GDPR, EU AI Act, and equality frameworks. This interdisciplinary study reveals critical misalignments between technical fairness strategies and anti-discrimination law and data protection law, drawing primarily on GDPR and anti-discrimination law in EU, UK and US frameworks, through a systematic review of 51 peer-reviewed studies (2020–2025). We identify three key gaps: (1) technically “fair” models frequently disregard individualised harm and proportionality principles; (2) dominant mitigation techniques (pre-/in-/post-processing) erase legally essential evidence of structural discrimination; (3) inconsistent use of bias, discrimination, and unfairness across ML literature creates ambiguity when these systems are assessed against evolving legal standards. To bridge these gaps, we propose a legally robust audit framework anchored in five principles: jurisdictional baseline specification, discrimination-fairness distinction, sector-specific duty integration, design of legally meaningful metrics, and preservation of procedural safeguards. This framework enables developers and regulators to advance ML systems that satisfy both statistical parity and substantive legal-compliance requirements across jurisdictions.
Liu et al. (Sat,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: