文化主体性与生成式人工智能的价值导向干预  

Cultural Subjectivity and Value-Oriented Intervention of Generative AI

作　　者：马文 陈云松[1] Ma Wen;Chen Yunsong

机构地区：[1]南京大学社会学院

出　　处：《江苏社会科学》2024年第1期66-74,I0003,共10页Jiangsu Social Sciences

基　　金：国家社会科学基金重大项目“大数据驱动的网络社会心态发展规律与引导策略研究”(19ZDA149)的阶段性成果。

摘　　要：生成式人工智能通过对海量互联网数据的深度学习来生产信息,不可避免地带有数据提供主体所在时空的文化观念,因此,其生成的内容具有时空上的局限性和动态演化的特征。以大型语言模型ChatGPT和文心一言为比较研究对象,通过对真实的种族争议案例进行内容生成测试实验,从文化主体性角度探讨基于不同时空语料的生成式人工智能平台的价值观差异。结果表明,基于中文训练的文心一言语言模型并未展示出与ChatGPT明显区分的当代中国公众所持有的文化主体意识。该现象与自21世纪初开始积累的中文互联网语料的文化主体性时间演化特征相关。据此提出,通过对语言模型语料库进行时间轴加权或复制训练以调整“时空折叠”,能够对生成式人工智能平台的价值观导向进行干预。Generative AI produces information through deep learning of massive Internet data,inevitably carrying the cultural concepts of the time and space where the data providers are located.Therefore,the content it generates has limitations of time and space and features of dynamic evolution.Taking large language models ChatGPT and Wenxin Yiyan as comparative research objects,this paper conducts content generation test experiments on real cases of racial disputes,and discusses the value differences between generative AI platforms trained on different corpora from the perspective of cultural subjectivity.The results show that the Wenxin Yiyan language model trained on Chinese corpus does not exhibit contemporary Chinese public cultural subjectivity consciousness that is clearly distinguished from ChatGPT.This phenomenon is related to the temporal evolution characteristics of cultural subjectivity of the Chinese corpus on the Internet accumulated since the beginning of the 21st century.Accordingly,it is proposed that by weighting the time axis of the language model corpus or duplicate training to adjust"spacetime folding",the value orientation of generative AI platforms should be intervened.

关 键 词：生成式人工智能 ChatGPT 文心一言 文化主体性 时空折叠 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] G05[自动化与计算机技术—控制科学与工程]