生成式人工智能科普内容生产的话语策略偏向研究——基于算法审计的方法

Research on the Discoursal Strategy Bias of Generative ArtificialIntelligence Produce Science Popularization Content：Based on theMethod of Algorithm Auditing

中文关键词: 生成式人工智能 算法审计 内容生产 话语策略 身份表征

英文关键词: generative artificial intelligence  algorithm auditing  content production  discursivestrategies  identity representation

中文摘要:

      生成式人工智能技术的出现与快速发展极大改变了当前的科学传播生态。该技术目前不仅有效赋 能了科普创作者，使得科普内容生产变得更为快捷便利，更以非人类行动者身份深度融入社会化协同传播体 系。想要更好地实现生成式人工智能对于当前科普主体生态格局的良性介入，关键在于系统解析其作为科学 传播者在生成科普内容时的策略偏向，以及其与人类科学传播者之间使用策略可能存在的差异。本文基于算 法审计的方法，通过对国内外代表性生成式人工智能工具生成的科普内容进行编码分析发现，生成式人工智 能在进行科普内容生产时切实采用了一系列多元化的话语策略，且其所采用的话语策略与人类科学传播者相 比存在一定差异，整体上呈现出一种偏向于微观化且身份模糊的特点。其内部基于不同语料库大模型的话语 策略也存在着相当的差异性，可能需要进一步区别对待。这一结果或可为后续进一步完善对于生成式人工智 能科普生产的规制与引导提供经验借鉴。

英文摘要:

      The emergence and rapid development of generative artificial intelligence（GenAI）have greatly changed the current science popularization ecology. It not only effectively empowers current science popularzation creators，making the production of science popularization content more convenient and faster，but also directly intervenes in the current socialized collaborative science popularization pattern as a non-human actor. To better realize the benign intervention of GenAI in the current ecological pattern of science popularization，it is first necessary to systematically analyze its strategy bias as a science communicator when generating science popularization content and the possible differences between its science popularization strategy use and that of human science popularization creators. Based on the method of algorithm auditing，this paper finds，through coding and analyzing the science popularization content that representative GenAI tools generated at home and abroad，that GenAI has indeed adopted a series of diversified discourse strategies when producing science communicators，and there are certain differences between the discourse strategies adopted and those of human science communicators. Overall，it presents a characteristic of being inclined to be microcosmic and having an ambiguous identity. There are also considerable differences in the discourse strategies of large models based on different corpora within them，and they may need to be treated differently. This result may provide experience and reference for further improving the regulation and guidance of science popularization by GenAI in the future.