Regulating hallucination risks in large language models across their lifecycle: lessons for China from the European Union

Xu, Linyi (2026) Regulating hallucination risks in large language models across their lifecycle: lessons for China from the European Union. LL.M(R) thesis, University of Glasgow.

Abstract

The rapid development of large language models (LLMs) has raised scholarly and regulatory concerns about the phenomenon of hallucinations, which means LLMs produce factually inaccurate, misleading, or fabricated content in a confident way. This thesis argues that hallucinations constitute a distinct legal and regulatory challenge that cannot be adequately addressed through general artificial intelligence (AI) governance alone. undertakes a comparative legal analysis of how the European Union (EU) and China understand, regulate, and seek to mitigate the harms caused by hallucinations in LLMs, combining doctrinal method with tech and law method to assess both the theoretical coherence and the practical enforceability of hallucination-related regulations. To better address the uncertainty that technological change creates for legal regulation, this thesis develops a lifecycle-based analytical framework that divides the lifecycle of LLMs into five interrelated stages: model design, pretraining, fine-tuning and alignment, deployment and interaction, and monitoring and iteration. On this basis, it examines how hallucination risks arise across the lifecycle, how they generate harm for stakeholders at both the micro and macro levels, and how regulation responds to those harms in practice. It further compares the legal approaches adopted in the EU and China, identifying both convergences and divergences, as well as the normative gaps and enforcement dilemmas that remain in each jurisdiction. Through this comparison, the thesis draws lessons that may inform the future regulation of LLMs in China.

This thesis advances two main arguments encompassing both theoretical and practical dimensions. Theoretically, although the EU and China recognise hallucinations as part of broader Artificial Intelligence (AI) risks and incorporate them into their AI or generative AI governance frameworks, the specific harms they pose raise distinct technical and legal challenges that are insufficiently addressed by existing general regulations. Neither jurisdiction has yet developed a targeted legal response. However, key differences remain. The EU employs a horizontally integrated, legally codified framework anchored in the AI Act, with provisions spanning the entire lifecycle. In contrast, China adopts a vertically layered model and focuses more on the design, monitoring, and iteration stages. Practically, both jurisdictions implement measures to mitigate hallucination-related harms at both macro and micro levels. The EU’s approach tends to be broader in scope, while China’s is more operationally focused and implementable.

Overall, the thesis contributes to the broader discourse on AI regulation by proposing that hallucination risks demand not only lifecycle-based legal frameworks but also cross-jurisdictional learning to ensure both innovation and accountability in the governance of LLMs.

• Item Type: Thesis (LL.M(R))
• Qualification Level: Doctoral
• Additional Information: Supported by funding from the China Scholarship Council Scholarship (CSC).
• Keywords: Artificial Intelligence, hallucination, large language model, Artificial Intelligence Act.
• Subjects: K Law > K Law (General)
• Colleges/Schools: College of Social Sciences > School of Law
• Funder's Name: China Scholarship Council Scholarship (CSC)
• Supervisor's Name: Erickson, Professor Kristofer
• Date of Award: 2026
• Depositing User: Theses Team
• Unique ID: glathesis:2026-85971
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 26 May 2026 15:06
• Last Modified: 26 May 2026 15:09
• Thesis DOI: 10.5525/gla.thesis.85971
• URI: https://theses.gla.ac.uk/id/eprint/85971