D3.6. Report on paradigms, policies and metrics for algorithmic fairness

This deliverable reviews key concepts, methods, and policy frameworks for achieving algorithmic fairness, especially in the context of artificial intelligence and automated decision-making. It outlines three types of bias—pre-existing, technical, and emergent—and explains how these can lead to various forms of discrimination, including direct, indirect, intersectional, and emergent. The report presents fairness metrics such as demographic parity, equal opportunity, and counterfactual fairness, and discusses their strengths and trade-offs. It describes strategies for bias detection and mitigation across the machine learning pipeline, including pre-processing (e.g. reweighing and data repair), in-processing (e.g. adversarial debiasing), and post-processing (e.g. outcome adjustment). Tools like Fairlearn, AI Fairness 360, and Aequitas are compared based on functionality, licensing, and usability. The report also addresses the importance of high-quality, representative data, offering practical recommendations on dataset documentation and transparency (e.g. datasheets, nutrition labels). Finally, it examines regulatory and ethical frameworks, including GDPR, the EU AI Act, ISO standards, and voluntary guidelines, to support fair and responsible AI development and deployment.