CHAPTER 7. ARTIFICIAL INTELLIGENCE IN INTERNATIONAL LAW SCHOLARSHIP: THE NARRATIVE OF CHATGPT

This record contains the full transcripts of three research conversations conducted with successive versions and functionalities of OpenAI’s GPT‑4 model, used as empirical material for a chapter in the edited volume Global Order in the Age of AI (Oxford University Press, forthcoming). The conversations were conducted by [authors’ names, incl. Emilia Justyna Powell] and focus on the theme “Legitimacy of International Law” (Example 1 in the chapter). Each session explores the same substantive topic but uses a different version or functionality of GPT‑4: Session 1 (5 April 2023) – ChatGPT‑4 (released 14 March 2023) Session 2 (23 October 2024) – ChatGPT‑4o (released 13 May 2024) Session 3 (28 July 2025) – GPT‑4 with Deep Research functionality All three conversations were conducted using the paid, professional versions of the respective models. The dataset allows readers to trace, in a granular way, how the model’s responses evolved over time in terms of legal reasoning, structure, and depth. In the chapter, we use these transcripts to examine: how different GPT‑4 variants reason about the legitimacy of international law, the stability and variability of AI‑generated legal arguments across model versions, the models’ abilities to manage complex normative debates, and the broader implications of AI‑assisted reasoning for international law scholarship and practice. Session 1 (ChatGPT‑4) documents the earliest conversation, in which GPT‑4 already displays sophisticated natural‑language abilities and a reasonably coherent treatment of doctrinal and theoretical debates, but with more limited contextual depth and fewer explicit meta‑reflections on its own reasoning process. Session 2 (ChatGPT‑4o) illustrates notable changes in speed, contextual understanding, and multimodal‑oriented design. In this session, GPT‑4o provides more nuanced, better‑structured, and often more responsive answers, with improved coherence over longer exchanges and a more refined ability to track the conversation’s normative and empirical dimensions. Session 3 (GPT‑4 – Deep Research) showcases the use of GPT‑4’s Deep Research functionality. Here the model engages in multi‑step reasoning, synthesizes information across multiple documents or sources, and produces systematically structured and explicitly reasoned answers. This session is used in the chapter to explore the potential and limitations of iterative, research‑style interactions with AI in legal scholarship. Together, the three transcripts serve as a comparative dataset documenting the evolution of GPT‑4‑based systems as tools for AI‑assisted legal reasoning. They are intended as: supplementary material to the book chapter (for transparency and reproducibility), a teaching resource for courses on international law, AI and law, or methodology, and a reference point for future empirical work on large language models in legal contexts. Users are encouraged to cite this dataset alongside the book chapter when using or discussing the transcripts in academic or policy work.本数据集收录了针对OpenAI的GPT‑4模型不同迭代版本与功能开展的三项研究对话的完整转录文本，作为即将由牛津大学出版社（Oxford University Press）出版的编著文集《人工智能时代的全球秩序》（Global Order in the Age of AI）某一章节的实证研究材料。对话由[作者名单，含Emilia Justyna Powell]主持，主题为“国际法的合法性”（为本章节示例1）。 每项会话围绕同一实质主题展开，但采用GPT‑4的不同版本或功能： 会话1（2023年4月5日）——ChatGPT‑4（2023年3月14日发布） 会话2（2024年10月23日）——ChatGPT‑4o（2024年5月13日发布） 会话3（2025年7月28日）——搭载深度研究（Deep Research）功能的GPT‑4 三项对话均通过对应模型的付费专业版本完成。 本数据集可让读者以精细化视角追踪模型应答在法律推理、文本结构与论证深度层面的时序演进脉络。在本章节中，我们借助这些转录文本展开四项研究： 1. 不同GPT‑4变体如何推演国际法的合法性命题； 2. AI生成的法律论证在不同模型版本间的稳定性与变异性； 3. 模型应对复杂规范性辩论的能力； 4. AI辅助推理对国际法学界研究与实务的更广泛影响。 会话1（ChatGPT‑4）为最早开展的对话，此时GPT‑4已展现出成熟的自然语言处理能力，可对教义学与理论辩论展开较为连贯的阐释，但上下文深度有限，且对自身推理过程的显性元反思较少。 会话2（ChatGPT‑4o）体现出显著的性能迭代：在响应速度、上下文理解能力与多模态导向设计层面均有提升。本次会话中，ChatGPT‑4o的应答更具精细化层次、结构更清晰，且响应性更强；长对话中的连贯性更优，同时在追踪对话的规范性与经验维度上能力更为精进。 会话3（搭载Deep Research功能的GPT‑4）展示了GPT‑4深度研究功能的应用场景。在此会话中，模型开展多阶段推理，跨多文档或来源整合信息，并生成结构严谨、推理过程明确的应答。本章节借助本次会话，探讨了法学研究中与AI开展迭代式、研究型交互的潜力与局限。 三项转录文本共同构成了一份比较数据集，记录了基于GPT‑4的系统作为AI辅助法律推理工具的演进历程。本数据集的定位如下： 1. 作为本书章节的补充材料，以保障研究的透明度与可重复性； 2. 作为国际法、人工智能与法律或法学方法论相关课程的教学资源； 3. 作为未来开展法律语境下大语言模型（Large Language Model, LLM）实证研究的参考基准。 若在学术或政策工作中使用或讨论本转录文本，敬请将本数据集与本书章节一并引用。