Semantic Alignment Between Normative Theories of Ethics and the European Union Artificial Intelligence Act: A Transformer-Based Semantic Textual Similarity Analysis

TL;DR

This study uses Transformer-based semantic textual similarity to analyze how well the EU AI Act aligns with three major normative ethical theories, revealing deontological ethics as most aligned.

Contribution

It introduces a novel method employing ensemble Transformer models to quantify semantic alignment between ethical theories and legal texts.

Findings

Deontological ethics shows the highest semantic alignment with the EU AI Act.

Transformer-based STS effectively measures moral grounding in legal regulations.

Separate analysis of preamble and provisions captures different ethical dimensions.

Abstract

The European Union Artificial Intelligence (EU AI) Act, which explicitly references fundamental rights and ethical principles, is a comprehensive regulatory framework for governing Artificial Intelligence (AI) systems. This study examines the moral grounding of the EU AI Act by analyzing the semantic alignment between three canonically distinct normative ethical theories (virtue ethics, deontological ethics, and consequentialism) and the Act's regulatory language. Building on philosophical and chronological considerations, the concept of influence is treated as a relational construct between the theories of ethics and the regulatory text. As a proxy for this relationship, Semantic Textual Similarity (STS) is employed to quantify the degree of alignment between the theory descriptions and the Act. The Act's preamble and statutory provisions are analyzed separately to capture its intentional and operational ethical groundings. To describe each theory distinctively and to reduce semantic overlap among theories, theory descriptions are manually preprocessed. To compute similarity scores, a heterogeneous embedding-level ensemble approach, comprising five lightweight Transformer-based encoders (SBERT, ALBERT, DistilBERT, RoBERTa, and TinyBERT), is used. To represent document-level alignment estimates, voting and averaging are used to aggregate STS scores. The findings indicate that deontological ethics exhibits the highest overall semantic alignment with both components of the EU AI Act.