Cultural Subjectivity and Value-Oriented Intervention of Generative AI
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.
万方数据
维普
