作　者：冯玉军 沈鸿艺
关键词：大模型幻觉；公权力机关；人工智能应用；风险分级；幻觉规制；自由裁量权
摘　要：

随着中国人工智能技术的不断发展，法律场景下的大模型幻觉现象愈发普遍。通过访谈调研发现，在公权力机关应用大模型的现实场景中，已经出现了包括法条幻觉、事实幻觉与涵摄幻觉在内的多类典型幻觉。在对公安机关、检察院和法院的醉酒危险驾驶类案件决策场景的专项测试中，发现幻觉集中性呈现在一至两个决策点上。考虑到错误大小、决策模式以及问责模式的转变，法律场景下产生的幻觉有必要进行类型化规制。规制思路应当从场景风险分级的基本原则切入，通过权利影响的处分性和自由裁量权两个维度的组合，形成“高风险、中风险、低风险”的划分格局。依据风险等级与来源针对性地设计“前端预防—中端监督—后端问责”的三阶段规制路径，实现分类分级的精细化治理。

Categorized Regulation of Large-scale Model Hallucinations In Legal Scenarios
——Focusing on the Risks of Large-scale Model Application in Public Authorities
Feng Yujun, Shen Hongyi
（Renmin University of China）
Abstract With the continuous development of artificial intelligence in China, large-scale AI model hallucinations have become increasingly prevalent in legal scenarios. Field interviews and investigations reveal that the practical deployment of large-scale models by public authorities has given rise to various typical hallucination categories, namely, statutory provision hallucination, factual hallucination, and subsumption hallucination. Targeted tests on case adjudication involving drunk driving across public security authorities, procuratorates and people's courts demonstrate that such hallucinations tend to converge at merely one or two pivotal decision-making nodes. Given the changes in severity of errors, decision-making patterns, and accountability models, categorical regulation of legal-domain large-scale model hallucinations is urgently warranted, The regulatory approach, starting from the basic principle of scenario-based risk classification, should establish a "high, medium, and low-risk" classification pattern by combining the two dimensions of dispositive impact on legal rights and discretion scope. Targeted at respective risk levels and risk sources, a three-stage regulatory framework of "front-end prevention — in-process supervision —post-hoc accountability" can be formulated to deliver refined, tiered governance by category.
Key words hallucinations in large-scale AI models; public authorities; artificial intelligence applications; risk grading; regulating hallucinations; discretion

 

■ 作者简介 冯玉军，中国人民大学法学院教授，北京 100872；沈鸿艺，中国人民大学法学院博士研究生。

[PDF]（下载数：）