大模型幻觉:人机传播中的认知风险与共治可能  

The Hallucinations of Large Language Models:Perceived Risk and the Possibility of Co-governance in Human-computer Communication

作　　者：张铮[1] 刘晨旭 ZHANG Zheng;LIU Chen-xu(School of Journalism and Communication,Tsinghua University,Beijing 100084,China)

机构地区：[1]清华大学新闻与传播学院,北京100084

出　　处：《苏州大学学报（哲学社会科学版）》2024年第5期171-180,共10页Journal of Soochow University(Philosophy & Social Science Edition)

基　　金：国家社会科学基金艺术学项目“新型数字文化消费对Z世代生活方式的影响研究”(项目编号:22BH156)的阶段性成果。

摘　　要：大模型在人机交互中可能产生“幻觉”,即生成的内容看似合理但与事实相悖。幻觉问题的产生并非偶然,由技术发展的局限性和用户输入的提示词与情境等共同所致。大模型幻觉对人的认知和信息传播构成风险,并对“数据范式”下的知识生产和隐性知识的价值显露带来挑战;同时,认知依赖下人机交互影响着人类的思维模式和判断力。为应对这些挑战,可通过提高训练数据的质量并明晰治理标准、增强模型透明性和可信性、实现有监督的机器动态自治等策略完善大模型未来发展方向。大模型幻觉无法完全避免,从另一个角度看,幻觉现象在创造性领域可能激发创新思维,为人类文化与艺术的发展注入新的活力与动力。Large language models can cause hallucination in human-computer interaction,generating content that seems reasonable but is contrary to the facts.The problem of hallucination is not accidental,but caused by a combination of technological limitations and user input prompts and contexts.The hallucination of large language models poses a risk to human cognition and information dissemination,and challenges the production of knowledge and the disclosure of the value of tacit knowledge under the data paradigm,while human-computer interaction under cognitive dependency affects human thought patterns and judgment.To meet these challenges,the large language models can improve the future development direction by improving the quality of training data and clarifying the governance standards,increasing the transparency and trustworthiness of the models,and realizing the supervised dynamic autonomy of machines.The hallucinations of large-scale language models cannot be completely avoided,and from another perspective,the phenomenon of hallucination may stimulate innovative thinking in the creative field,injecting new vitality and power into the development of human culture and art.

关 键 词：大模型幻觉 认知风险 知识生产 人机协同治理 人机传播 

分 类 号：G206[文化科学—传播学]