摘要
算法解释旨在为算法治理提供认知基础。既有的算法解释路径大多通过对模型内在因果机制的还原来建构算法治理的认知。然而,该路径却因当下大模型所具有的复杂性而在技术上难以为继,无法满足算法治理的认知期待。算法解释规范的法理基础应当在回应算法治理需求的同时,顺应解释技术的能力边界。在此背景下,“目的论解释”作为对“因果论解释”的有效补强手段,无须解构晦涩的技术细节,即可重建治理者对大模型的理解。为此,算法治理的认知基础可以并应当从聚焦模型内在机制转向关注模型的外在目的。具体而言,在争议解决型、行政监管型两种治理场景下,沿技术主义路径识别模型输出被赋予的外在设计目的、部署目的以及模型自发涌现的内在输出目的,并据此评价与规制模型行为。这是大模型治理认知基础的转向,也是人工智能规制法理的转向。
Algorithm interpretation aims to provide a cognitive basis for algorithm governance.Existing algorithm interpretation paths mostly construct the cognition of algorithm governance by restoring the inherent causal mechanisms of the model.However,due to the complexity of current large models,this path is technically unsustainable and fails to meet the cognitive expectations of algorithm governance.The legal basis of algorithm interpretation norms should respond to the demands of algorithm governance while conforming to the capability boundaries of interpretation techniques.Against this backdrop,teleological explanation,as an effective reinforcement to causal explanation,can reconstruct the governance’s understanding of large models without deconstructing obscure technical details.Therefore,the cognitive basis of algorithmic governance should and can shift from focusing on the internal mechanisms of the model to paying attention to the external purposes of the model.Specifically,in the cases of dispute resolution and administrative supervision,external design and deployment purposes assigned to the model output along the technical path,as well as the internal output purposes that emerge spontaneously by the model can be identified,with efforts made to evaluate and regulate the model behavior based on this.This represents a shift in the cognitive basis of large model governance and also a shift in the legal theory of artificial intelligence regulation.
作者
魏义铭
孙笑侠
WEI Yi-ming;SUN Xiao-xia
出处
《南通大学学报(社会科学版)》
北大核心
2025年第6期110-121,156,共13页
Journal of Nantong University:Social Sciences Edition
基金
教育部高校人文社会科学重点研究基地重大项目“科技与人权的关系研究”(20JD820004)。
关键词
大模型治理
算法解释
因果论解释
目的论解释
认知转向
large-model governance
algorithmic explanation
causal explanation
teleological explanation
cognitive turn
