On Algorithmic Fairness and the EU Regulations

TL;DR

Frames algorithmic fairness through non-discrimination in EU contexts and situates analysis around AI Act and GDPR, focusing on recruitment systems. Employs an argumentative compliance approach that maps GDPR bases and AI Act requirements to bias-mitigation strategies across a set of scenarios. Demonstrates when bias correction is legally feasible under EU regulations and where non-compliance may arise, highlighting anonymization as a practical mitigation and the public-interest pathway as particularly impactful. Provides regulatory-oriented guidance for compliant design and governance of high-risk AI systems in Europe.

Abstract

The short paper discusses algorithmic fairness by focusing on non-discrimination and a few important laws in the European Union (EU). In addition to the EU laws addressing discrimination explicitly, the discussion is based on the EU's recently enacted regulation for artificial intelligence (AI) and the older General Data Protection Regulation (GDPR). Through a theoretical scenario analysis, on one hand, the paper demonstrates that correcting discriminatory biases in AI systems can be legally done under the EU regulations. On the other hand, the scenarios also illustrate some practical scenarios from which legal non-compliance may follow. With these scenarios and the accompanying discussion, the paper contributes to the algorithmic fairness research with a few legal insights, enlarging and strengthening also the growing research domain of compliance in AI engineering.