中文    English

Journal of library and information science in agriculture

   

Verification Friction in Checking Generative-AI Hallucinations and a Support Framework for University Libraries: Interviews with 25 Researchers at a Comprehensive University

XIAO Bo1, WU Xiaoqing2   

  1. 1. Party School of the CPC Changzhi Municipal Committee, Changzhi 046000
    2. Zibo Polytechnic University, Zibo 255300
  • Received:2026-08-11 Online:2026-11-16

Abstract:

[Purpose/Significance] Generative artificial intelligence (AI), is moving from single-turn question answering to retrieval-connected and multi-tool systems. It, plays a role in literature discovery, research writing, and data interpretation. Hallucinations in research settings have expanded to include not only obvious fabricated references to real sources that do not support the claims attributed to them, but also conclusions that go beyond the available evidence, leaps of inference, and uncertain information presented as fact. Because generated content must eventually enter an evidence chain that is traceable, verifiable, and attributable to a responsible researcher, turning an apparently complete AI answer into usable evidence demands further work that existing studies have seldom examined. Adapting the notion of friction from research on scientific data, this study defines verification friction as the extra conversion cost that researchers bear after obtaining an AI output and before using it in a formal research product. It asks which hallucinations researchers meet in different tasks, how they become verification friction, and where university libraries can reduce it and where their intervention should stop. [Method/Process] Using information ecology as a sensitizing perspective, the study draws on semi-structured interviews with 25 researchers in formal teaching or research roles at a comprehensive university. The interviews were conducted across the humanities, social sciences, natural sciences and engineering from January to March 2025. Participants had used generative AI in a research task within the previous three months and were recruited purposively through library service channels, supplemented by snowball sampling. Interviews suit the question because researchers' judgment and handling of AI output are seldom visible in usage logs. Transcripts of about 330 000 Chinese characters were coded in NVivo 14 in four steps: open coding (412 meaning units and 126 initial codes), constant comparison (35 secondary codes), thematic aggregation (nine primary themes) and relation building across 97 verification episodes. Two coders jointly coded three transcripts and independently coded the other 22 on a segmentation fixed beforehand by the author (agreement 87.4% and Cohen's kappa 0.84). Disagreements were settled by discussion, the corpus was re-examined to limit coding drift, four participants with stable routines were examined as negative cases, and saturation was tracked in batches of five interviews. [Results/Conclusions] Three types of hallucination emerged: traceability distortion (nonexistent references reported by 21 of 25 participants and misattributed findings by 19), semantic and boundary distortion (overgeneralized conclusions by 20 and misread statistics by 16) and reasoning-structure distortion (flawed reasoning chains by 17). Researchers started verifying when output conflicted with prior knowledge, when key evidence cannot be located, or when stated certainty outruns the evidence. Four kinds of verification friction followed: source, tool, method, and responsibility friction. These differ from the difficulty of finding information because they arise once an answer has been found and concern converting it into traceable, reproducible evidence to which the researcher can respond. Libraries can appropriately intervene in evidence tracing, verification-method training, disclosure and process support, whereas disciplinary judgment, methodological adjudication and the attribution of research responsibility rest with researchers, disciplinary experts and research administrators. A three-tier framework links tools (a single entry point for checking generated references together with reusable verification records), capabilities (discipline-specific training and task-based checklists) and governance (disclosure templates and graded verification). The findings describe 25 researchers at one university, and recruitment through library channels may have raised their familiarity with library services. Future work should compare institutions and add verification logs.

Key words: university library, generative AI, verification friction, AI hallucinations, research support

CLC Number: 

  • G258.6

Fig.1

Analytical path of the study"

Table 1

Distinguishing verification friction from adjacent concepts"

概念 主要对象 核心问题 与验证摩擦的区别与联系
信息搜寻成本 信息获取行为 找到信息需要多少成本 关注获取信息的投入;核验中的检索在操作上与之相同,但以确认AI给出的内容为目的,计入验证摩擦
信息过载 信息数量与复杂度 如何处理过量信息 可提高验证成本,但不是验证摩擦的必要条件
认知负荷 个体认知资源 任务是否超出处理能力 可伴随各类摩擦出现,在方法判断中较为集中
人机信任校准 对AI的依赖程度 信任是否匹配实际可靠性 验证摩擦影响继续依赖、抽查或弃用的选择
验证摩擦 AI输出进入科研证据链的转换 如何使生成内容达到科研使用标准 涵盖来源、工具、方法和责任4个维度

Table 2

Characteristics of interview participants"

维度 类别 人数/人 占比/%
学科 人文学科 7 28
学科 社会科学 8 32
学科 理学 5 20
学科 工学 5 20
职称 教授 8 32
职称 副教授 10 40
职称 讲师/助理研究人员 7 28
AI使用频率 高频(每周≥3次) 12 48
AI使用频率 中频(每周1~2次) 9 36
AI使用频率 低频(每月数次) 4 16

Table 3

Major forms of generative-AI hallucinations in research tasks"

一级类型 二级表现 人数/人 比例/%
可追溯性失真 不存在的文献 21 84
可追溯性失真 标识符与文献错配 18 72
可追溯性失真 作者身份混淆 15 60
可追溯性失真 研究成果归因错误 19 76
可追溯性失真 数据来源或数据虚构 14 56
语义与边界失真 统计结果误读 16 64
语义与边界失真 方法适用条件错误 12 48
语义与边界失真 结论过度概括 20 80
语义与边界失真 术语语义偏移 15 60
推理结构失真 推理链问题 17 68
推理结构失真 内部矛盾 13 52
推理结构失真 无证据概念关联 11 44

Fig.2

Verification process for AI outputs in research"

Table 4

Examples of the coding process"

访谈片段(受访者) 初始代码 二级代码 一级主题
“题目、作者都有,搜了几个库也找不到”(P03) 引文不存在;跨库检索失败 不存在的文献 可追溯性失真
“DOI能打开,但不是它说的文章”(P12) DOI指向其他文献 标识符与文献错配 可追溯性失真
“p值大于0.05,它还说显著”(P08) 显著性判断与p值不符 统计结果误读 语义与边界失真
“每一步像有道理,连起来少了前提”(P17) 隐性前提缺失;推理跳跃 推理链问题 推理结构失真
“它说数据出自统计年鉴,翻了两个年份的年鉴都对不上”(P09) 数据出处无法核对;年鉴比对未果 数据来源难追溯 来源摩擦
“几个数据库来回找,每个入口都不一样,40多分钟”(P06) 跨库切换;入口各异;时间成本 平台入口分散 工具摩擦
“回归结果它解释得头头是道,手头没有现成的核对步骤,只能去问做统计的同事”(P14) 缺少核对步骤;求助同事 统计与方法检查程序缺失 方法摩擦
“一处错了,后面还要查到什么程度心里没底,只好把剩下几条引文也挨个去查”(P21) 错误引发核验范围扩大;核验深度无把握 核验深度不清 责任摩擦

Table 5

Batchwise record of thematic saturation"

访谈批次 新增一级主题/个 新增二级代码/个 新的机制关系
P1~P5 6 18 出现
P6~P10 2 10 出现
P11~P15 1 5 出现
P16~P20 0 2 未出现
P21~P25 0 0 未出现

Table 6

Major verification-support needs reported by participants"

支持需求 人数/人 比例/% 主要指向
AI生成内容快速验证工具 19 76 引文存在性、元数据与来源定位
专业数据库与AI/检索流程衔接 18 72 减少跨库切换和重复检索
AI批判性使用培训 17 68 形成分任务核验方法和披露意识

Table 7

Dimensions, manifestations, and consequences of verification friction"

维度 主要表现 典型核验动作 可能后果
来源摩擦 文献、数据、版本或出处难定位 检索原文、核对元数据、追溯数据源 查证耗时;证据链中断
工具摩擦 平台分散、权限差异大、记录不可复用 跨库切换、链接解析、保存查证记录 重复劳动;核验中止
方法摩擦 缺少统计、方法与逻辑检查程序 复算、比对适用条件、咨询同行 误判或全盘弃用
责任摩擦 披露、核验深度与责任主体不清 确认规范、保存过程、分级核验 过度信任或过度保守

Table 8

Verification priorities across disciplinary tasks"

学科任务 主要核验对象 常见核验方式 对应支持重点
人文学科 版本、作者、引文语境 原文溯源、多版本比对 文献溯源与版本辨析
社会科学 数据、统计解释、结论边界 数据源核对、复算、方法条件检查 统计解释与模型适用性
理学 公式、参数、原理一致性 数学复核、原始论文比对 关键推导与参数来源
工学 代码、标准、技术约束 运行测试、版本核对、条件复现 代码与标准记录

Table 9

Division of responsibilities and library boundaries in AI verification"

问题 首要责任主体 图书馆角色
文献是否存在 研究者与图书馆 可直接支持
DOI、作者、版本是否匹配 研究者与图书馆 可直接支持
数据来源是否可追溯 研究者 提供数据发现与溯源支持
统计模型是否正确 研究者与方法专家 提供基础核验指南,不作裁决
理论判断是否成立 研究者与学科专家 不作实质裁决
AI使用如何披露 科研管理部门与期刊 提供模板和培训
科研不端如何认定 学术委员会等 不承担认定职能

Table 10

Disciplinary tasks and the three-tier service modules"

学科任务 工具层 能力层 治理层
人文学科 版本、作者、文献跨库核验 文献溯源、版本辨析 引文与AI使用披露
社会科学 数据源、DOI、统计资料连接 统计解释、模型条件、结论边界 数据与AI使用披露
理学 原始论文、公式与参数来源 原理一致性与数学复核 关键推导记录
工学 代码、标准与技术资料 可运行性、复现条件、工程约束 版本与工具记录

Fig.3

Framework for university-library support of researchers in verifying generative-AI hallucinations"

[1]
教育部, 中央网信办, 国家发展改革委, 等. 教育部等九部门关于加快推进教育数字化的意见: 教办〔2025〕3号[EB/OL]. (2025-04-16)[2026-09-29].
[2]
Ji Ziwei, Lee N, Frieske R, et al. Survey of hallucination in natural language generation[J]. ACM Computing Surveys, 2023, 55(12): 1-38.
[3]
Maynez J, Narayan S, Bohnet B, et al. On faithfulness and factuality in abstractive summarization[C]//Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Kerrville: Association for Computational Linguistics, 2020: 1906-1919.
[4]
Walters W H, Wilder E I. Fabrication and errors in the bibliographic citations generated by ChatGPT[J]. Scientific Reports, 2023, 13: 14045.
[5]
Kalai A T, Nachum O, Vempala S S, et al. Evaluating large language models for accuracy incentivizes hallucinations[J]. Nature, 2026, 653(8116): 1047-1051.
[6]
Hüllermeier E, Waegeman W. Aleatoric and epistemic uncertainty in machine learning: An introduction to concepts and methods[J]. Machine Learning, 2021, 110(3): 457-506.
[7]
Adel A, Alani N. Can generative AI reliably synthesise literature? Exploring hallucination issues in ChatGPT[J]. AI & Society, 2025, 40(8): 6799-6812.
[8]
Lewis P, Perez E, Piktus A, et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]//Proceedings of the 34th International Conference on Neural Information Processing Systems. New York: ACM, 2020: 9459-9474.
[9]
Manakul P, Liusie A, Gales M. SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models[C]//Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. Kerrville: Association for Computational Linguistics 2023: 9004-9017.
[10]
Cox A M, Pinfield S, Rutter S. The intelligent library: Thought leaders' views on the likely impact of artificial intelligence on academic libraries[J]. Library Hi Tech, 2019, 37(3): 418-435.
[11]
张晓林. Library-Inside: AI赋能图书馆新质生产力的一种基础模型[J]. 中国图书馆学报, 2024, 50(3): 4-16.
Zhang Xiaolin. Library-inside: A foundation model for AI-empowered new quality productive forces for library services[J]. Journal of Library Science in China, 2024, 50(3): 4-16.
[12]
Long Duri, Magerko B. What is AI literacy? Competencies and design considerations[C]//Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. New York: ACM, 2020: 1-16.
[13]
University L M, Archambault S. Expanding on the frames: Making a case for algorithmic literacy[J]. Communications in Information Literacy, 2023, 17(2): 530-553.
[14]
沈洪杰, 沈洪伟, 王均莉. 生成式AI赋能数字图书馆信息素养教育:路径探索、挑战分析与应对策略[J]. 农业图书情报学报, 2025, 37(7): 50-60.
Shen Hongjie, Shen Hongwei, Wang Junli. Generative AI empowering information literacy education in digital libraries: Path exploration, challenge analysis, and response strategies[J]. Journal of Library and Information Science in Agriculture, 2025, 37(7): 50-60.
[15]
Ali M Y, Richardson J. AI literacy guidelines and policies for academic libraries: A scoping review[J]. IFLA Journal, 2025, 51(3): 588-599.
[16]
Nardi B A, O'Day V L. Information Ecologies: Using Technology with Heart[M]. Cambridge: MIT Press, 1999: 49-50.
[17]
Edwards P N, Mayernik M S, Batcheller A L, et al. Science friction: Data, metadata, and collaboration[J]. Social Studies of Science, 2011, 41(5): 667-690.
[18]
Pirolli P, Card S. Information foraging[J]. Psychological Review, 1999, 106(4): 643-675.
[19]
Sweller J. Cognitive load during problem solving: Effects on learning[J]. Cognitive Science, 1988, 12(2): 257-285.
[20]
Bawden D, Robinson L. The dark side of information: Overload, anxiety and other paradoxes and pathologies[J]. Journal of Information Science, 2009, 35(2): 180-191.
[21]
Lee J D, See K A. Trust in automation: Designing for appropriate reliance[J]. Human Factors, 2004, 46(1): 50-80.
[22]
Guest G, Bunce A, Johnson L. How many interviews are enough?: An experiment with data saturation and variability[J]. Field Methods, 2006, 18(1): 59-82.
[1] WAN Qiao. Ecological Transformation, Risk Assessment, and Value Orientation of University Libraries in the Development of Digital and Intelligent Technology [J]. Journal of library and information science in agriculture, 2026, 38(9): 93-102.
[2] LI Dan, FENG Danran. Diffusion of Generative Artificial Intelligence Technology Based on Complex Network Evolutionary Game [J]. Journal of library and information science in agriculture, 2026, 38(8): 93-107.
[3] CHENG Cheng, ZHOU Jie, WANG Han. Transformation Logic and Practical Pathways for University Libraries Integrating into the Trusted Data Space Construction [J]. Journal of library and information science in agriculture, 2026, 38(6): 43-58.
[4] MAO Kaiyan. Promotion of Chinese Classical Literature for Children's Reading: Applications and Initiatives of Sora-Type Video Generation [J]. Journal of library and information science in agriculture, 2026, 38(2): 90-103.
[5] JIANG Jingze, ZHOU Tianmin, LI Mei, CHENG Cheng, CHEN Haiyan. A study of the Core Competence Model of Compound AI Librarians in the Intelligent Transformation of University Libraries [J]. Journal of library and information science in agriculture, 2025, 37(9): 97-109.
[6] JIANG Yumeng. An Investigation and Enlightenment on the Current Status of AI Literacy Education in Nordic University Libraries [J]. Journal of library and information science in agriculture, 2025, 37(7): 19-34.
[7] ZHANG Tao, LYU Qianhui. Generative AI Governance Practices in Europe and the United States and the Enlightenment for China [J]. Journal of library and information science in agriculture, 2025, 37(4): 12-23.
[8] TAN Miao, DAI Mengfei. Empowered Digital Reading Promotion of Historical Documents with Generative AI [J]. Journal of library and information science in agriculture, 2025, 37(4): 83-93.
[9] SUN Zhimei, LIU Yan, LIU Wenyun, LI Ruiqin. Perception and Utilization of Digital Literacy Knowledge by University Library Users from Representational Pers Perspective [J]. Journal of library and information science in agriculture, 2025, 37(2): 23-36.
[10] GUO Limin, LIU Yueru, FU Yaming. Design and Transformation Pathways of Library Systems Driven by Generative Agents [J]. Journal of library and information science in agriculture, 2025, 37(12): 36-47.
[11] HE Cong, YANG Jing, XIAO Xiong, SUN Wenwen, LI Chenggang. Construction of Efficiency Evaluation System for University Library Resource Construction under the Background of AI Era: Practical Exploration of Beijing Institute of Technology Library [J]. Journal of library and information science in agriculture, 2025, 37(10): 96-111.
[12] Fan YUAN, Jia LI. Opportunities, Challenges, and Future Directions for Generative Artificial Intelligence in Library Information Literacy Education: A Scoping Review [J]. Journal of library and information science in agriculture, 2024, 36(9): 44-57.
[13] Ru WU, Ruhan ZHANG, Chenxing DAI, cheng LU. Intellectual Property Literacy Education in University Libraries Based on Innovative Talent Cultivation [J]. Journal of library and information science in agriculture, 2024, 36(6): 79-87.
[14] Feiyan HUANG. Practices for New Research University's Collaborative Collection Development of Print and Electronic Books: A Case Study of SUSTech Library [J]. Journal of library and information science in agriculture, 2024, 36(11): 92-103.
[15] Xiao SHI. Analysis of Strategic Planning Content and Development Trends of American University Libraries [J]. Journal of library and information science in agriculture, 2024, 36(10): 38-52.
Viewed
Full text


Abstract

Cited

  Shared   
  Discussed   
No Suggested Reading articles found!