| 291 | 0 | 83 |
| 下载次数 | 被引频次 | 阅读次数 |
大语言模型的幻觉问题已嵌入内容生产与传播的每个环节，成为影响个体认知、社会分化、人机信任等的重要课题。然而不透明性、概率性与自主性作为人工智能系统的本质特征，大模型幻觉难以从技术层面完全消弭。从大模型技术底层逻辑出发，可分析大模型幻觉的定义内涵、表现特征与产生机制，并辩证看待大模型幻觉的消极影响与潜在价值。作为一种难以消解的复杂现象，大模型幻觉对于我们的挑战在于如何利用其潜在好处，同时尽量减少其消极影响。对幻觉率的容忍度应随场景而流动，温度参数的调整可在信息准确性与创意性之间进行平衡与取舍。此外，用户应提高“提示”素养，减少幻觉可能带来的负面效应，并在人机协同中最大程度地激发其潜在价值。
Abstract：Hallucinations within large language models have become deeply entrenched in every phase of content production and dissemination processes, thereby emerging as a crucial problem that exerts a significant impact on individual cognitive processes, social stratification, and human-machine trust relationships.Nevertheless, considering that opacity, probabilistic nature, and autonomy constitute the intrinsic characteristics of artificial intelligence systems, it is a difficult task to entirely eradicate hallucinations in large language models from a technical standpoint.This paper commences from the fundamental logic underpinning large language model technologies.It conducts an in-depth analysis of the definitional implications, expressive features, and generative mechanisms of hallucinations within large language models.Moreover, it adopts a dialectical approach to examine both the negative ramifications and potential values of these hallucinations.As a complex and intractable phenomenon, the primary challenge presented by large language model hallucinations for us lies in the optimization of leveraging their potential advantages while concurrently minimizing their adverse effects to the greatest extent possible.Consequently, the tolerance levels and adjustment strategies regarding the hallucination rate should be adaptable to diverse scenarios.Users are enabled to customize temperature parameters, thereby facilitating the achievement of a balance and enabling trade-offs between information accuracy and creativity.Additionally, users are required to enhance their proficiency in “prompting” to mitigate the potential negative consequences induced by hallucinations and to maximize the potential value of large language model hallucinations within the context of human-machine collaboration.
[1] BOMMASANI R,HUDSON D A,ADELI E,et al.On the opportunities and risks of foundation models.(2021-10-06)[2024-09-10].http://arxiv.org/abs/2108.07258.
[2] 闫坤如.人工智能价值对齐的价值表征及伦理路径.伦理学研究，2024(4):94-100.
[3] 张静，陈方迪.2023年度词都和AI相关：剑桥词典选“幻觉”,韦氏词典选“真实”.澎湃新闻.(2023-11-29)[2024-09-29].https://www.thepaper.cn/newsDetail_forward_25455617.
[4] HUANG L,YU W,Ma W,et al.A Survey on Hallucination in Large Language Models:Principles,Taxonomy,Challenges,and Open Questions.(2023-11-19)[2024-09-29].https://doi.org/10.1145/3703155.
[5] KALAI A T,VEMPAL S S.Calibrated language models must hallucinate.(2024-03-20)[2024-10-29].https://doi.org/10.48550/arXiv.2311.14648.
[6] DAHL M,MAGESH V,SUZGUN M,et al.Large legal fictions:profiling legal hallucinations in large language models.Journal of Legal Analysis,2024,16(1):64-93.[2024-06-21].https://doi.org/10.1093/jla/laae003.
[7] AGRAWAL G,KUMARAGE T,ALGHAMDI Z,et al.Can knowledge graphs reduce hallucinations in LLMs:A Survey.(2024-03-16)[2024-10-29].http://arxiv.org/abs/2311.07914
[8] 张铮，刘晨旭.大模型幻觉：人机传播中的认知风险与共治可能.苏州大学学报(哲学社会科学版),2024(5):171-180.
[9] 彭兰.从ChatGPT透视智能传播与人机关系的全景及前景.新闻大学，2023(4):1-16+119.
[10] HOHWY J.The predictive mind.Oxford:Oxford university press,2013.
[11] KAMBHAMPATI S.Can large language models reason and plan?Annals of the New York academy of sciences,2024,1534(1):15-18.[2024-03-26].https://doi.org/10.1111/nyas.15125.
[12] JI Z,LEE N,FRIESKE R,et al.Survey of hallucination in natural language generation.ACM computing surveys,2023,55(12):1-13.[2024-07-14].https://doi.org/10.1145/3571730.
[13] RAWTE V,CHAKRABORTY S,PATHAK A,et al.The troubling emergence of hallucination in large language models-an extensive definition,quantification,and prescriptive remediations.Proceedings of the 2023 conference on empirical methods in natural language processing,2023.[2023-10-04].https://doi.org/10.48550/arXiv.2310.04988.
[14] 梅夏英.复杂系统与智能涌现：未来数字法研究的范式图景.法学家，2024(5):45-59+191-192.
[15] 江积海，阮文强.新零售企业商业模式场景化创新能创造价值倍增吗?科学学研究，2020,38(2):346-356.
[16] 严昊，刘禹良，金连文，等.类ChatGPT大模型发展、应用和前景.中国图象图形学报，2023,28(9):2749-2762.
[17] POHL R F.Cognitive illusions:Intriguing phenomena in thinking,judgment,and memory.Oxfordshire:Taylor & Francis Group,2016.
[18] 杜文静.法律人工智能概率推理的困境与破解.学术研究，2022,45(4):29-34.
[19] 方婕妤.人类中心主义与媒介物质性之间的抉择——论唯物现象学的内在矛盾.新闻界，2025(1):87-96.
[20] BROWN T,MANN B,RYDER N,et al.Language models are few-shot learners.Advances in neural information processing systems,2020,33(4):1877-1901.[2020-07-22].https://doi.org/10.48550/arxiv.2005.14165.
[21] 朱迪亚，王白尔，麦肯齐.为什么：关于因果关系的新科学.江生，于华，译.北京：中信出版集团股份有限公司，2019.
[22] 许志伟，李海龙，李博，等.AIGC大模型测评综述：使能技术、安全隐患和应对.计算机科学与探索，2024,18(9):2293-2325.
[23] MUSHESHE M,CHRISPO M.A conceptual study of cognitive atrophy in homo sapiens through a darwinian analysis of overreliance on artificial intelligence.East African journal of interdisciplinary studies,2025,8(2):185-207.https://doi.org/10.37284/eajis.8.2.3463.
[24] 王浩伟，汪璠，王秉琰.主题视角下生成式人工智能生成内容与用户生成内容的比较.情报理论与实践，2023,46(10):200-207+199.
[25] 彭兰.智能生成内容如何影响人的认知与创造?编辑之友，2023(11):21-28.
[26] ZUCKER L G.Production of trust:institutional sources of economic structure.Research in organizational behavior,1986(8):53-111.
[27] 齐玥，陈俊廷，秦邵天，等.通用人工智能时代的人与AI信任.心理科学进展，2024,32(12):1-13.
[28] 张洪忠，任吴炯.超越“第二自我”的人机对话——基于AI大模型应用的信任关系探讨.新闻大学，2024(3):47-60.
[29] 孙彦，李纾，殷晓莉.决策与推理的双系统——启发式系统和分析系统.心理科学进展，2007(5):721-726.
[30] GILOVICH T,GRIFFIN D W,KAHNEMAN D.Heuristics and biases:the psychology of intuitive judgement.Cambridge:Cambridge university press,2002.
[31] 林爱珺，陈亦新.信息熵、媒体算法与价值引领.湖南师范大学社会科学学报，2022,51(2):125-131.
[32] HARDY D.Creativity:Flow and the psychology of discovery and invention.Personnel Psychology,1998,51(3):794-797.
[33] 姜宇辉.“相信此世”——在与克尔凯郭尔的对话之中重释德勒兹的“否定主义”.四川大学学报(哲学社会科学版),2023(5):135-145+197.
[34] 吉登斯.社会的构成：结构化理论大纲.李康，李猛，译.北京：中国人民大学出版社，2016.
[35] 刘大年，曹月.知识的幻象：人工智能与知识变迁.现代出版，2024(9):53-67.
[36] ELLUL J.The technological society.New York:Alfred A.Knopf,1964.
[37] KALAI A T,VEMPALA S S.Calibrated language models must hallucinate.(2024-03-20)[2024-10-29].http://arxiv.org/abs/2311.
[38] 经羽伦，张殿元.生成式AI幻象的制造逻辑及其超真实建构的文化后果.山东师范大学学报(社会科学版),2024,69(5):113-126.
[39] 赵晓伟，祝智庭，沈书生.教育提示语工程：构建数智时代的认识论新话语.中国远程教育，2023,43(11):22-31.
[40] CLARK P A.AI’s rise generates new job title:prompt engineer.(2023-02-22)[2024-09-29].https://publicservicesalliance.org/wp-content/uploads/2023/02/AIs-rise-generates-new-job-title-Prompt-engineer.pdf.
[41] 叶继红，雷德森.科学技术的社会变迁：一个社会学的分析.科学管理研究，2004(5):51-55.
基本信息:
DOI：10.14086/j.cnki.xwycbpl.2026.01.002
中图分类号:G206
引用信息:
[1]喻国明,金丽萍,苏芳.大模型幻觉的表现特征、效应争议及潜在价值[J].新闻与传播评论,2026,79(01):14-24.DOI:10.14086/j.cnki.xwycbpl.2026.01.002.
基金信息:
教育部哲学社会科学研究重大专项项目(2025JZDZ014)