西安交通大学学报（社会科学版）2026，Vol.46Issue(3)：139-149,11.DOI:10.15896/j.xjtuskxb.202603013
大模型代理的风险治理
Risk Governance of Large Language Model Agents
摘要
Abstract
Large Language Model agents(LLM agents)stand as a new evolution of artificial intelligence(AI)technology,embodying the governance philosophy of artificial intelligence paternalism(AI paternalism).LLM agents are primarily engaged in technical information screening and instruction conversion.The ultimate attribution of rights and obligations,however,remains rooted in the relationships among users,AI operators,and third-party service providers.These relationships include the basic legal bond between users and third-party service providers,the technical service relationship between users and AI operators,and the technical interaction between AI operators and third-party service providers. This paper first employs a normative research approach.Building on a comparison between legal paternalism and AI paternalism,it highlights that the latter differs from the former in terms of intervention subjects,methods,and mechanisms-while sharing similarities in intervention objectives,doubts about intervention,and degrees of intervention.In conducting a substantive analysis of the concept of AI paternalism,the cost-benefit analysis method is introduced.It argues that when the law examines AI paternalism from a paternalistic perspective,the decision of whether to allow AI to act as a paternalistic agent in certain contexts should be based on a social cost-benefit analysis of LLM.On the benefit side,LLM can enhance decision-makers' information capacity,counteract the increase of entropy in the decision-making process,and rectify human biases in decision outcomes.On the cost side,compared with traditional human decision-making,LLM can reduce decision-making costs;yet when measured against autonomous decision-making,they may give rise to risks such as technical failure and the risk of restricting free will.Consequently,on the premise that LLM agents contribute to improved social welfare,the role of AI paternalism should be better leveraged by innovating the decision confirmation mechanism,behavior supervision mechanism,and liability allocation mechanism for large model agents. This paper makes two key extensions.First,it focuses on LLM agents as an emerging form of AI technology,offering an in-depth analysis of this new technology from the perspective of legal relations while exploring the theory of AI paternalism that underpins it.Second,unlike most articles on AI law that adopt a doctrinal analysis approach,this paper introduces the cost-benefit analysis method to examine the consequences of applying LLM agents in decision-making. This study uncovers the internal logic of AI's participation in decision-making,helping to better understand the relationship between humans and machines in the era of human-machine symbiosis and providing intellectual support for the standardized application of large model agent technology.关键词
大模型代理/法律关系/人工智能家长主义/人工智能体/技术价值评估/风险治理/成本收益分析Key words
large language model agent(LLM agent)/legal relationship/artificial intelligence paternalism(AI paternalism)/artificial intelligence agent(AI agent)/technology value evaluation/risk governance/cost-benefit analysis分类
社会科学引用本文复制引用
张钦昱,王淑瑶..大模型代理的风险治理[J].西安交通大学学报（社会科学版）,2026,46(3):139-149,11.基金项目
北京市习近平新时代中国特色社会主义思想研究中心重大项目(25LLFXA134). （25LLFXA134）
