[Paper Review] Fairness and Explainability in Automatic Decision-Making Systems. A challenge for computer science and law

Thierry Kirat, Olivia Tambou+2 more|arXiv (Cornell University)|May 14, 2022

Digitalization, Law, and Regulation4 citations

TL;DR

This paper bridges computer science and law by analyzing fairness and explainability in automated decision-making systems, emphasizing that technical choices in supervised learning carry social and legal implications. It compares U.S. and European legal frameworks on fairness, highlights the polysemy of legal explainability concepts, and calls for interdisciplinary collaboration to address normative, technical, and legal trade-offs in algorithmic decision systems.

ABSTRACT

The paper offers a contribution to the interdisciplinary constructs of analyzing fairness issues in automatic algorithmic decisions. Section 1 shows that technical choices in supervised learning have social implications that need to be considered. Section 2 proposes a contextual approach to the issue of unintended group discrimination, i.e. decision rules that are facially neutral but generate disproportionate impacts across social groups (e.g., gender, race or ethnicity). The contextualization will focus on the legal systems of the United States on the one hand and Europe on the other. In particular, legislation and case law tend to promote different standards of fairness on both sides of the Atlantic. Section 3 is devoted to the explainability of algorithmic decisions; it will confront and attempt to cross-reference legal concepts (in European and French law) with technical concepts and will highlight the plurality, even polysemy, of European and French legal texts relating to the explicability of algorithmic decisions. The conclusion proposes directions for further research.

Motivation & Objective

• To examine how technical choices in supervised machine learning inherently carry social and legal implications beyond algorithmic design.
• To compare fairness standards in U.S. and European legal systems, particularly regarding disparate impact and facially neutral decision rules.
• To analyze the legal and technical dimensions of explainability in algorithmic decisions, especially the ambiguity and plurality in European and French legal texts.
• To identify gaps in interdisciplinary research on national legal systems and their impact on algorithmic fairness and explainability.
• To propose future research directions integrating legal norms, social values, and technical feasibility in automated decision-making systems.

Proposed method

• Uses a contextual legal analysis to compare fairness standards in U.S. anti-discrimination law and European non-discrimination law.
• Applies empirical risk minimization (ERM) as the core technical framework for supervised learning, with focus on loss functions, hypothesis classes, and evaluation benchmarks.
• Classifies biases into data bias and algorithmic bias, tracing their origins across data collection, model selection, and evaluation processes.
• Cross-references legal concepts of explainability (e.g., in GDPR and French law) with technical interpretability methods, highlighting conceptual polysemy.
• Proposes axiomatic frameworks to model decision-makers’ values, such as 'what you see is what you get' versus 'we’re all equal', to assess fairness implications.
• Engages in normative analysis to explore trade-offs between fairness, public safety, and other social objectives, using examples like the COMPAS recidivism tool.

Experimental results

Research questions

• RQ1 How do technical choices in supervised learning—such as data collection, model selection, and loss functions—produce unintended social consequences?
• RQ2 What are the key differences in fairness standards between U.S. and European legal systems regarding disparate impact in automated decisions?
• RQ3 How do legal concepts of explainability in European and French law differ from or align with technical interpretability methods in machine learning?
• RQ4 To what extent can legal and technical constraints be reconciled in the design of fair and explainable algorithmic systems?
• RQ5 How can social values and normative preferences be formally modeled in algorithmic decision-making to ensure alignment with legal and ethical standards?

Key findings

• Technical choices in supervised learning—such as data collection, model class, and loss function—directly influence fairness outcomes and carry inherent social implications.
• U.S. and European legal systems promote different fairness standards: the U.S. emphasizes disparate impact and individual fairness, while Europe emphasizes non-discrimination and procedural rights.
• Legal texts on explainability in Europe and France exhibit polysemy, with multiple interpretations of 'explainability' that complicate legal compliance and technical implementation.
• The study identifies a critical gap in research on national legal systems outside the U.S. and EU, especially in Asia, Oceania, and Latin America.
• There is a fundamental tension between maximizing public safety and reducing racial disparities in algorithmic decisions, as shown in the COMPAS recidivism case.
• Fairness cannot be reduced to a purely technical problem; it requires explicit social and legal normative choices, such as between equality of opportunity and demographic parity.