Abstract
This article explores the methodological and normative challenges of fine-tuning Large Language Models (LLMs) for international law. It focuses on three central architectural choices: (1) the selection of relevant domain-specific data, which raises questions related to legal sources doctrine, representativity and inclusion; (2) the integration of human expertise for reinforcement learning, and the requirement to address interpretative diversity; and (3) the capacity of the user interface to represent legal disagreement and to provide information about limitations and weaknesses of the model. Those architectural choices embed jurisprudential assumptions about the nature of international law and legal interpretation. As part of what has been labelled the “participative turn in AI design”, this work emphasises a user-centred approach that considers this main question: what is a legal LLM for, from the user’s perspective? This perspective reinforces the need for participative design processes that reflect the plurality of international legal voices. The article ultimately shows that the most transformative aspect of legal LLMs may not be their computational capabilities, but their demand for positioning on the nature and function of legal interpretation in international law. The article presents visual illustrations providing ideal-typical models of design choices depending on users related assumptions. Designing legal LLMs, far from bypassing legal theory expertise, re-engages foundational questions of jurisprudence and requires space for interdisciplinary co-design.
