生成式人工智能传播中的偏向与规制——以ChatGPT为例  

The Bias and Regulation in the Dissemination of Generative Artificial Intelligence:A Case Study of ChatGPT

作　　者：周茂君[1,2] 郭斌 ZHOU Maojun;GUO Bin

机构地区：[1]武汉大学媒体发展研究中心 [2]武汉大学新闻与传播学院,湖北武汉430072

出　　处：《学习与实践》2024年第1期33-41,共9页Study and Practice

基　　金：国家社会科学基金项目“中国新媒体广告规制研究”(项目编号:17BXW094)。

摘　　要：ChatGPT的诞生标志着人类信息传播技术出现了重大突破。它凭借预训练和大模型获得了强大的“泛化能力”,可以基于上下文内容推断出最合适的答案。然而,作为基于概率推断的生成式人工智能,ChatGPT也存在着一定的缺陷。从内部推断逻辑来看,基于人类文本知识进行预训练的模型,会复刻人类社会的偏见与缺陷,在传播过程中会进一步强化偏向,忽视边缘群体。受制于外部环境的介入,ChatGPT也会受到资本力量和政治立场的影响。对于生成式人工智能的发展,要从训练数据和模型设计阶段防止偏向的发生,同时也要发挥行业力量、制度优势和用户技术素养,共同引导其健康发展。The emergence of ChatGPT represents a significant breakthrough in human information dissemination technology.With its powerful"generalization capabilities"achieved through pre-training and the use of large-scale models,ChatGPT is able to provide contextually appropriate responses based on the input it receives.However,as a generative artificial intelligence that relies on probabilistic inference,ChatGPT is not immune to inherent flaws.From an internal reasoning perspective,models pre-trained on human textual knowledge tend to replicate societal biases and shortcomings,which can be further exacerbated during dissemination and potentially marginalize underrepresented groups.In addition,external environmental factors,such as capital forces and political stances,can also influence ChatGPT.To ensure the healthy development of generative Al,it is essential to prevent bias from occurring at both the training data and model design stage.Furthermore,concerted efforts from industry stakeholders,leveraging institutional advantages and enhancing users'technical literacy,are crucial for guiding its progress in a positive direction.

关 键 词：ChatGPT 人工智能 传播偏向 规制 

分 类 号：G206[文化科学—传播学] TP18[自动化与计算机技术—控制理论与控制工程]