Abstract: Generative AI has triggered a transformation in the paradigms of knowledge production and dissemination across various fields. Meanwhile, biases such as "favoring males over females" and "discrimination against minorities" have been laid bare. Previous research on the ethics of technology has largely been speculative, with few empirical studies following the social constructivist approach and focusing on the role of humans, while insufficient attention has been paid to non-human actors. In light of this, the study employs actor-network theory to examine the labor process of generative AI data annotators, exploring how human and non-human actors intertwine to inscribe biases into intelligence during the practical construction of these biases. The research identifies three causes of bias production: "individual behavior," "constraints of objects," and "capital orientation." Finally, the study offers suggestions at the end for managing biases in generative AI through "identification discussions," "process optimization," and "ethical frameworks."

摘要： 生成式AI在各个领域引发了知识生产与传播范式的变革，与此同时，“爱男轻女”、“爱白欺黑”等偏见现象暴露无遗。既往研究对技术伦理的探讨多为思辨，少数实证研究遵循社会建构的理路，聚焦于人在其中发挥的作用，但对物的关注不足。有鉴于此，研究基于行动者网络理论，跟随生成式AI数据标注师的劳动过程来考察在偏见的实践建构中，人类与非人行动者是如何在缠绕交汇下将偏见“铭刻”于智能之中的。研究梳理出“个体行为”、“物的掣肘”与“资本导向”三个偏见生产的原因，并在文末提出“识别磋商”、“流程优化”、“伦理框架”的建议以管理生成式AI的偏见问题。