法律大语言模型的司法应用及其规范  

Judicial Applications and Regulation of Large Legal Language Models

作　　者：魏斌 Wei Bin

机构地区：[1]浙江大学光华法学院、数字法治实验室

出　　处：《东方法学》2024年第5期57-73,共17页Oriental Law

基　　金：国家重点研发计划重点专项“基层社会网格治理数字化关键技术研究与应用示范”(项目批准号:2021YFC3300300);国家社科基金青年项目“智慧司法中法律推理与法律解析的融合路径研究”(项目批准号:21CFX006);浙江大学学科交叉预研专项“法律人工智能的可解释性问题研究”(项目批准号:S20230011)的阶段性研究成果。

摘　　要：法律大语言模型因其超级学习和“涌现”能力,在法律任务中表现出卓越的性能,引发了法律科技领域的新一轮革命。法律大语言模型的司法应用主要体现在法律语言理解、法律知识问答、法律预测和法律文本生成等领域。然而,随着其在审判等司法实践中的应用,暴露出可解释性弱和“幻觉”等问题。法律大语言模型仍然难以胜任法律人的核心工作,包括法律推理、司法证明、法律解释、法律论证和疑难案件中的道德判断等。法律大语言模型在情感、道德、逻辑推理、决策机制和经验学习等方面仍然与法律人有本质的差异,因此,法律大语言模型的司法应用需要在严格的规范之下合理使用。Large legal language models have triggered a new revolution in the field of legal technology due to their superior learning and"emergent"capabilities,demonstrating remarkable performance in various legal tasks.The judicial applications of large legal language models primarily include legal language comprehension,legal knowledge Q&A,legal prediction,and legal text generation.However,as these models are increasingly applied in judicial practices such as court trials,they reveal issues such as weak interpretability and"hallucinations."Large legal language models still struggle to perform the core tasks of legal professionals,including legal reasoning,judicial proof,legal interpretation,legal argumentation,and moral judgment in complex cases.Significant differences remain between large legal language models and legal professionals in areas such as emotion,morality,logical reasoning,decision-making mechanisms,and experiential learning.Therefore,the judicial application of large legal language models must be conducted under strict regulation to ensure their appropriate use.

关 键 词：生成式人工智能 法律大语言模型 法律大数据 司法应用 数据安全 法律科技 

分 类 号：DF0[政治法律—法学理论]