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.