开源新生态:人工智能蒸馏技术的风险与治理OACHSSCD

[Purpose]Traditional theoretical research focuses on the model training paradigm centered on data size and pays insufficient at-tention to the risks in the application of AI distillation technology.The article clarifies the dual risks of distillation technology applications and proposes corresponding governance strategies,with a view to providing behavioral guidelines for AI developers and contributing inno-vative solutions for AI regulators.[Method]The literature research method is used to sort out the development history and basic principles of distillation technology,and the descriptive research method and case study method are used to analyze the legitimacy risk and secondary risk of distillation technology application.[Result/Conclusion]The distillation technology takes data and model knowledge as the appli-cation object,which on the one hand exposes student model developers to copyright infringement,unfair competition and breach of con-tract,and on the other hand exacerbates the proliferation of infringement risks on the output side of student models,and creates problems of attribution of liability in the absence of norms.The risk management of distillation technology should be in line with the concept of"open source innovation",while promoting the open utilization of data and model knowledge,clarifying the causality of the infringement risk on the output side and improving the rules of infringement liability determination.