[论文解读] Influence of Normative Theories of Ethics on the European Union Artificial Intelligence Act: A Transformer-Based Analysis Using Semantic Textual Similarity

Mehmet Murat ALBAYRAKOĞLU, Mehmet Nafiz Aydin|arXiv (Cornell University)|Jan 19, 2026

Ethics and Social Impacts of AI被引用 0

一句话总结

论文通过基于变换器的 STS 集成，量化美德伦理、义务伦理与结果主义在语义上与欧盟 AI 法案的对齐程度，结果显示义务伦理的对齐度最高。

ABSTRACT

Despite being regarded as a significant step toward regulating Artificial Intelligence (AI) systems and its emphasis on fundamental rights, the European Union Artificial Intelligence (EU AI) Act is not immune to moral criticism. This research aims to investigate the impact of three major normative theories of ethics (virtue ethics, deontological ethics, and consequentialism) on the EU AI Act. We introduce the concept of influence, confirmed by philosophical and chronological analysis, to examine the underlying relationship between the theories and the Act. As a proxy measure of this influence, we propose using Semantic Textual Similarity (STS) to quantify the degree of alignment between the theories (influencers) and the Act (influencee). To capture intentional and operational ethical consistency, the Act was divided into two parts: the preamble and the statutory provisions. The textual descriptions of the theories were manually preprocessed to reduce semantic overlap and ensure a distinct representation of each theory. A heterogeneous embedding-level ensemble approach was employed, utilizing five modified Bidirectional Encoder Representations from Transformers (BERT) models, built on the Transformer architecture, to compute STS scores. These scores represent the semantic alignment between various theories of ethics and each of the two components of the EU AI Act. The theories were evaluated by using voting and averaging, with findings indicating that deontological ethics has the most significant overall influence.

研究动机与目标

• 通过比较文本与经典规范性伦理理论（美德伦理、义务伦理、结果主义）来评估欧盟 AI 法案的道德基础。
• 将对齐建模为伦理理论与监管语言之间的关系影响。
• 应用语义文本相似度量化理论描述与法案组成部分（序言与条文）的对齐程度。
• 对理论描述进行预处理，以减少语义重叠并使每一理论的特征描述更具区分性。

提出的方法

• 用五个轻量化 Transformer 编码器（SBERT、ALBERT、DistilBERT、RoBERTa、TinyBERT）表示理论描述和法案文本。
• 计算理论描述与法案组成部分之间的语义文本相似度（STS）分数。
• 通过投票与平均来聚合分数，生成文档层面的对齐估计。
• 分别分析法案序言与法定条文以捕捉意向性与操作性伦理基础。

实验结果

研究问题

• RQ1 美德伦理、义务伦理与结果主义与欧盟 AI 法案的语义对齐程度如何？
• RQ2 哪一规范理论在法案及其组成部分（序言 vs 条文）上表现出最强的整体对齐？

主要发现

• 义务伦理在与欧盟 AI 法案的两个组成部分上显示出最高的整体语义对齐度。
• STS 分数来自五个编码模型的异质集成，并通过投票/平均进行聚合。
• 对法案进行了分别分析：序言（意向性基础）与法定条文（操作性基础）。
• 对理论描述进行预处理以减少语义重叠并强调每一理论的独特特征。