讲 座 信 息

主题：Biases and Inconsistencies in GPT Models

时间：2024年7月19日（周五）13:30-15:30

地点：浙江大学紫金港西区成均苑8幢1017

概要：The rising popularity of ChatGPT and other large language models (LLMs) has led to increasing studies investigating their susceptibility to mistakes and biases. However, most of these studies focus on models exclusively trained on English texts. This research is one of the first studies that systematically investigate the cross-language political biases and inconsistencies in LLMs, specifically GPT models with multilingual capacity. Using two languages, English and Chinese, we asked GPT the same questions about the United States and China. We found that the Chinese model presented the least negative sentiment toward China, whereas the English model was the most critical against China. Both models presented similar sentiments toward the US. This disparity cannot be explained by the potentially lower level of robustness of GPT models trained on non-English corpora. Moreover, both the Chinese and English models tended to be less critical towards the issues of their “own country,” represented by the language used, than the issues of “the other country.” This suggests that GPT multilingual models could potentially develop an “in-group bias” based on their training language. This study sheds light on the impact of AI-powered tools on information transmission and communication in an increasingly divided world

主讲人： 张尹霰，纽约市立大学皇后学院,社会学系助理教授

Yinxian Zhang is a political sociologist who is broadly interested in public discussion, ideology, state-society relations, and regime legitimacy. Her current book project combines qualitative and computational methods to explore the nature of the Chinese public sphere and its changes under the influence of state control. Yinxian received her Ph.D. in sociology from the University of Chicago in 2019. Before joining the Department of Sociology at Queens College, Yinxian was a China Public Policy Postdoctoral Fellow at the Ash Center for Democratic Governance and Innovation at Harvard University.

评议&主持人：李林倬，浙江大学社会学系新百人计划研究员

芝加哥大学社会学博士，研究领域为知识社会学、计算社会科学、经济社会学。研究方向为知识系统和经济系统中的创新、颠覆、不确定性和层级相关的议题。研究发表在Poetics, Plos One, npj Urban Sustainability, 社会学研究等中英文期刊上。