多模态大语言模型侵权的韧性治理

打开文本图片集

中图分类号：D923 文献标识码：A D0I:10.7535/j.issn.1671-1653.2025.05.00

Resilient Governance of Infringement by Multimodal Large Language Models

MAChenlong，ZHANGJunfa (LawSchool，JingdezhenCeramicUniversity，Jingdezhen，China)

Abstract: The application of multimodal large language models has given rise to multimodal risks，posing chalenges to the determination of infringement liability. Based onthe doctrine of reliance interest，users of multimodal large language models are generally not suitable to be identified as responsible entity. Instead，the operators of multimodal large language models should be the duty bearers.Given that the existing legislation has not yet clearly defined the liability rules for operators, this paper took resilience theory as the analytical framework to construct a liability mechanism covering three stages of infringement occurrence. Before the infringement occurs，by distinguishing the stages and scenarios of model generation，the operators' duty of care in data preprocessing, differentiated review，and other links is typologically defined；When the infringement occurs，the operators' obligation to conduct hierarchical interception of infringement instructions and to make timely corrections after receiving infringement notifications is clarified，so as to regulate the model's generation behavior；After the infringement occurrs，a liability reduction clause based on the probability of failure is set up，and a compulsory liability insurance mechanism is introduced to balance the dual goals of technological innovation and legal regulation， and promote the sustainable development of the multimodal large language model industry.

Keywords: multimodal large language model； operator of large language model； resilience theory；duty of care;infringement liability

以深度求索（DeepSeek)为代表的多模态大语言模型标志着通用人工智能的开发又迈进了一大步。（剩余13001字）