On the Need and Applicability of Causality for Fairness: A Unified Framework for AI Auditing and Legal Analysis

TL;DR

The paper tackles the problem of evaluating fairness in AI by establishing a causal framework that integrates legal and technical perspectives. It leverages structural causal models and the potential outcomes framework, using $TE$, $NDE$, $NIE$, and $PSE$ to decompose discrimination along explicit causal pathways. It analyzes legal evidentiary challenges in the EU through regulatory instruments like the AI Act and the AI Liability Directive, alongside landmark cases, to illustrate how causal analysis can inform accountability despite AI opacity. The authors advocate for a pragmatic, framework-backed approach that combines causal discovery, counterfactual reasoning, and path-specific analysis, while acknowledging limitations in graph specification, transportability, and data availability. The work aims to equip policymakers, practitioners, and AI developers with principled tools for transparent, accountable, and fair algorithmic decision-making, and outlines concrete directions for future research.

Abstract

As Artificial Intelligence (AI) increasingly influences decisions in critical societal sectors, understanding and establishing causality becomes essential for evaluating the fairness of automated systems. This article explores the significance of causal reasoning in addressing algorithmic discrimination, emphasizing both legal and societal perspectives. By reviewing landmark cases and regulatory frameworks, particularly within the European Union, we illustrate the challenges inherent in proving causal claims when confronted with opaque AI decision-making processes. The discussion outlines practical obstacles and methodological limitations in applying causal inference to real-world fairness scenarios, proposing actionable solutions to enhance transparency, accountability, and fairness in algorithm-driven decisions.

Figures (5)

• Figure 1: Causality is crucial for auditing algorithmic fairness because it accurately identifies whether automated systems produce genuinely discriminatory outcomes or merely reflect correlations; helps to disentangle the paths relating sensitive attribute and the outcome and justifiable dependency from proxy discrimination; causal analysis aligns with causal evidence for establishing liability in court-practice.
• Figure 2: Confounder structure.
• Figure 3: Mediator structure.
• Figure 4: Collider structure.
• Figure 5: Causal graph with two mediated paths.