Journal DOI 10.68956/journal.legal-engineering-technology

Creative Commons source reading · Legal language models

GDPR and Large Language Models: Technical and Legal Obstacles

Georgios Feretzakis; Evangelia Vagena; Konstantinos Kalodanis; Paraskevi Peristera; Dimitris Kalles; Athanasios Anastasiou

Originally published in Future Internet · 2025-03-28

From the article

Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.

Abstract

Large Language Models (LLMs) have revolutionized natural language processing but present significant technical and legal challenges when confronted with the General Data Protection Regulation (GDPR). This paper examines the complexities involved in reconciling the design and operation of LLMs with GDPR requirements. In particular, we analyze how key GDPR provisions—including the Right to Erasure, Right of Access, Right to Rectification, and restrictions on Automated Decision-Making—are challenged by the opaque and distributed nature of LLMs. We discuss issues such as the transformation of personal data into non-interpretable model parameters, difficulties in ensuring transparency and accountability, and the risks of bias and data over-collection. Moreover, the paper explores potential technical solutions such as machine unlearning, explainable AI (XAI), differential privacy, and federated learning, alongside strategies for embedding privacy-by-design principles and automated compliance tools into LLM development. The analysis is further enriched by considering the implications of emerging regulations like the EU’s Artificial Intelligence Act. In addition, we propose a four-layer governance framework that addresses data governance, technical privacy enhancements, continuous compliance monitoring, and explainability and oversight, thereby offering a practical roadmap for GDPR ali

Editorial reading guide

  • Which legal corpus, jurisdiction and evaluation task does the article use?
  • What evidence supports the reported performance, and where could the method fail in practice?