人机互动中不确定性表达的双重效应： 基于争议性科学话题的在线实验研究

Dual Effects of Uncertainty Expression in Human-Computer Interaction：An Online Experimental Study Based on a Controversial Scientific Issue

中文关键词: 人工智能 不确定性 大语言模型 科学传播 人机交互

英文关键词: artificial intelligence  uncertainty  large language models  science communication  human computer interaction

中文摘要:

      随着人工智能在科学传播中的深入应用，理解其不确定性表达对受众信息处理及人机交互行为的 影响，已成为亟待解决的理论与实践议题。本研究以转基因技术这一高争议性话题为案例，采用2（不确 定性：高vs. 低）× 4（角色模拟：无角色vs. 记者vs. 科学家vs. 普通网民） 的组间实验设计，依托自研的 大语言模型对话平台，收集并分析了547名参与者的自报告与在线交互数据，系统地探讨了不确定性沟通 在人机交互中的作用机制。研究发现，高不确定性陈述对受众产生双重效应。在认知评价层面，高不确定 性显著地降低了对信息内容及其来源的信任，并强化了风险感知、弱化了收益评估，呈现出明显的负面偏 向。然而，在行为层面，高不确定性条件下的对话长度与轮次均显著增加，反映出受众更强烈的信息寻求 动机与探索倾向。这表明，不确定性在人机交互情境中同时激活了两条相对独立的心理路径：认知评价路 径（负面评价）与行为调节路径（积极参与）。此外，大语言模型的角色模拟对信任与风险/收益感知无显 著影响，但显著地改变了交互行为，提示受众对AI 内容的评价更依赖内容特征而非角色标签。本研究拓展 了不确定性沟通理论在人工智能时代的适用范围，并为优化科学传播策略提供了理论依据与实践启示。

英文摘要:

      With the deep penetration and widespread application of artificial intelligence technology in the field of science communication，understanding the mechanisms by which its expression of uncertainty influences audience information processing and human-computer interaction behavior patterns has become a pressing theoretical and practical issue in science communication research. Based on this，this study selected genetically modified technology，a highly controversial and socially concerned scientific and technological issue，as the research object. We adopted a 2（uncertainty level： high vs. low）×4（simulated role: no role simulation vs. journalist vs. scientist vs. ordinary netizen） between-subjects experimental design. Through an independently developed AI dialogue platform based on large language models，we collected and analyzed self-reported and online interaction data from 547 Chinese participants to explore the complex mechanisms of uncertainty communication in human computer interaction contexts. The research results reveal a phenomenon of significant theoretical importance：the high uncertainty expression driven by artificial intelligence has produced a significant dual effect on the audience. Specifically，at the cognitive evaluation level，highly uncertain statements significantly reduced the audience's trust in information content and information sources，and at the same time showed a clear negative bias in the risk-benefit perception dimension，that is，significantly enhanced the audience’s risk perception and correspondingly weakened their cognitive evaluation of the benefits of related technologies. However，at the human-computer interaction behavior level，the dialogue length and number of dialogue turns between the audience and the large language model under high uncertainty conditions both showed a significant increasing trend，reflecting a stronger information seeking motivation and more active exploratory behavior tendencies compared to low uncertainty conditions. This finding indicates that uncertainty in human-computer interaction communication contexts simultaneously activates two relatively independent psychological processing processes: the cognitive evaluation path（leading to negative evaluation）and the behavioral regulation path（promoting active participation）. In addition，the role simulation of large language models had no significant effect on trust and risk/benefit perception，but had a significant effect on interaction behavior，indicating that the audience's evaluation of AI-generated content is more based on content characteristics than on role labels. The findings of this study not only enrich the applicability of uncertainty communication theory in the era of artificial intelligence，but also provide important theoretical guidance and practical insights for optimizing science communication strategies.