先验信念和任务知识对人机协同决策的影响研究

doi:10.16381/j.cnki.issn1003-207x.2024.1385

摘要/Abstract

摘要：

如何在现实环境中创建高效的人机团队是人机协同决策的关键问题。本文以AI辅助的人类决策场景为研究情境，通过一项随机对照实验探究决策者的先验信念和任务知识对人机协同决策的影响。研究表明，当决策者预先持有对预测标签分布的先验偏见时，这一信念状态导致决策者出现更多的假阳性错误，从而降低人机系统的协同程度；任务知识对于消除偏见的影响有积极作用，具有更高任务知识水平的决策者能够减少由先验偏见所诱发的假阳性错误决策；但任务知识对人机协同具有双向作用，即人机系统的协同程度并不是随着决策者任务知识的增加而无限增加，而是先增后减，呈现倒U型结构，这是由于在人机交互背景下，过高的任务知识会带来假阴性错误风险。基于研究结论，本文对现实环境中人机团队的部署与管理提供了科学指导。

关键词: 人工智能, 人机协同, 先验信念, 任务知识

Abstract:

The growing integration of artificial intelligence （AI） into human workflows has given rise to a new paradigm of human–AI collaborative decision-making， in which human judgment and AI recommendations are combined to achieve superior performance. However， effective human–AI collaboration remains challenging because people often use AI advice inappropriately. Hence， understanding how to build effective human-AI teams has become a central question for both researchers and practitioners. Existing research has primarily approached this issue through the lens of trust calibration， seeking to improve collaboration by encouraging people to rely on AI systems to an appropriate degree. These approaches typically focus on subjective psychological factors， such as trust in AI， self-confidence， and perceived usefulness. While insightful， such affective factors are difficult to manage through organizational interventions， and therefore their practical value for AI deployment may be limited. Drawing on the knowledge-belief framework in cognitive science， we argue that two task-related factors—prior beliefs and task knowledge—play a fundamental role in shaping how individuals process and integrate external information， including AI advice. Unlike affective factors， these cognitive factors can be influenced through training and task design， making them particularly relevant for organizational practice. As such， the present study examines how prior beliefs and task knowledge influence human–AI collaboration. To investigate these relationships， we conducted a between-subjects experiment on academic performance prediction， in which prior bias was manipulated （biased vs. unbiased）， and task knowledge was measured as a continuous variable. The results yield three main findings. First， biased prior beliefs impair human-AI collaboration. This effect arises because individuals with biased priors are more likely to make false-positive errors， causing them to incorrectly reject valid AI recommendations. Second， greater task knowledge mitigates the negative effect of prior bias. As individuals have more task knowledge， they become better able to recognize and avoid false-positive errors induced by distorted prior beliefs. Third， the relationship between task knowledge and the effectiveness of collaboration follows an inverted U-shaped pattern. Collaboration improves as task knowledge increases， but only up to a certain point. Beyond that point， additional task knowledge reduces overall team performance. This decline is driven by an increase in false-negative errors： individuals with very high levels of task knowledge become more likely to accept incorrect AI recommendations. We argue that this effect reflects overconfidence， whereby highly knowledgeable decision-makers become less vigilant in monitoring AI outputs. Based on these findings， we provide practical guidance for the design， formation， and management of human-AI teams in organizational settings.

Key words: artificial intelligence, human-machine synergy, prior beliefs, task knowledge

中图分类号:

C934

汪子昊,徐选华,王宗润, 等. 先验信念和任务知识对人机协同决策的影响研究[J]. 中国管理科学, 2026, 34(8): 92-103.

Zihao Wang,Xuanhua Xu,Zhongrun Wang, et al. Research on the Effects of Prior Beliefs and Task Knowledge on Human-AI Collaborative Decision-Making[J]. Chinese Journal of Management Science, 2026, 34(8): 92-103.

图/表 7

图1

表1

先验偏见与人机协同的关系"

变量	HMD $(N = 140)$
变量	M1	M2	M3	M4
第一步：基本变量
性别	0.028	0.027	0.025	0.021
年龄	0.001	0.001	0.000	0.000
学历	-0.029	-0.027	-0.038^*	-0.040^*
第二步：先验偏见		-0.059^**	-0.054^**	-0.054^**
第三步：任务知识水平			0.359^*	0.355^*
第四步：先验偏见✕任务知识水平				0.522^*
$R 2$	0.030	0.092	0.136	0.161
$Δ R 2$	0.030	0.062^**	0.044^*	0.025^*

表1

表2

任务知识与人机协同的倒U型关系"

	HMD $(N = 140)$
变量	M1	M2	M3
第一步：基本变量
性别	0.027	0.025	0.027
年龄	0.001	0.000	0.001
学历	-0.027	-0.038^*	-0.031
先验偏见	-0.059^**	-0.054^**	-0.053^**
第二步：任务知识水平		0.359^*	0.345^*
第三步：任务知识水平²			-3.025^*
$R 2$	0.092	0.136	0.166
$Δ R 2$	0.092^*	0.044^*	0.030^*

表2

图2

图3

图4

图5

参考文献 46

[1]	王红卫，李珏，刘建国，等. 人机融合复杂社会系统研究［J］. 中国管理科学， 2023， 31（7）： 1-21.
	Wang H W， Li J， Liu J G， et al. Research on human-machine integration complex social system［J］. Chinese Journal of Management Science， 2023， 31（7）： 1-21.
[2]	刘伦. 面向人机协同的公共决策转型：内涵、议题与框架［J］. 中国行政管理， 2023， 39（9）： 142-151.
	Liu L. Towards human-AI collaborative public decision making： Concepts， issues， and framework［J］. Chinese Public Administration， 2023， 39（9）： 142-151.
[3]	Lucas G M， Becerik-Gerber B， Roll S C. Calibrating workers’ trust in intelligent automated systems［J］. Patterns， 2024， 5（9）： 101045.
[4]	Wiczorek R， Meyer J. Effects of trust， self-confidence， and feedback on the use of decision automation［J］. Frontiers in Psychology， 2019， 10： 519.
[5]	Tolmeijer S， Gadiraju U， Ghantasala R， et al. Second chance for a first impression？ Trust development in intelligent system interaction［C］//Proceedings of the 29th ACM Conference on User Modeling， Adaptation and Personalization， Utrecht，Netherlands， June 21-25， ACM， 2021： 77-87.
[6]	Waggoner P， Kennedy R. The role of personality in trust in public policy automation［J］. Journal of Behavioral Data Science， 2022， 2（1）： 106-123.
[7]	Yin M， Wortman Vaughan J， Wallach H. Understanding the effect of accuracy on trust in machine learning models［C］//Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems， Glasgow， Scotland， May 4-9， ACM， 2019： 1-12.
[8]	Bansal G， Wu T， Zhou J， et al. Does the whole exceed its parts？ The effect of ai explanations on complementary team performance［C］//Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems， Yokohama， Japan， May 8-13， ACM， 2021： 1-16.
[9]	高彧，王聪，王翀. 点估计与区间估计：算法预测呈现方式对人机共融决策效果的影响［J］. 管理工程学报， 2024， 38（1）： 46-59.
	Gao Y， Wang C， Wang C. Point and interval estimation： The influence of algorithmic prediction presentation on human-algorithm-integrated decision-making［J］. Journal of Industrial Engineering and Engineering Management， 2024， 38（1）： 46-59.
[10]	Buçinca Z， Lin P， Gajos K Z， et al. Proxy tasks and subjective measures can be misleading in evaluating explainable AI systems［C］//Proceedings of the 25th International Conference on Intelligent User Interfaces， Cagliari，Italy， March 17-20，ACM， 2020： 454-464.
[11]	Yaniv I， Kleinberger E. Advice taking in decision making： Egocentric discounting and reputation formation［J］. Organizational Behavior and Human Decision Processes， 2000， 83（2）： 260-281.
[12]	Snow T. From satisficing to artificing： The evolution of administrative decision-making in the age of the algorithm［J］. Data Policy， 2021， 3： e3.
[13]	Smith G F， Benson P G， Curley S P. Belief， knowledge， and uncertainty： A cognitive perspective on subjective probability［J］. Organizational Behavior and Human Decision Processes， 1991， 48（2）： 291-321.
[14]	Leyer M， Schneider S. Decision augmentation and automation with artificial intelligence： Threat or opportunity for managers？［J］. Business Horizons， 2021， 64（5）： 711-724.
[15]	Shrestha Y R， Ben-Menahem S M， von Krogh G. Organizational decision-making structures in the age of artificial intelligence［J］. California Management Review， 2019， 61（4）： 66-83.
[16]	李忆，喻靓茹，邱东. 人与人工智能协作模式综述［J］. 情报杂志， 2020， 39（10）： 137-143.
	Li Y， Yu L R， Qiu D. Review of cooperation mode between human and artificial intelligence［J］. Journal of Intelligence， 2020， 39（10）： 137-143.
[17]	Hou T， Li M， Tan Y， et al. Physician adoption of AI assistant［J］. Manufacturing Service Operations Management， 2024， 26（5）： 1639-1655.
[18]	Selten F， Robeer M， Grimmelikhuijsen S. ‘Just like I thought’： Street-level bureaucrats trust AI recommendations if they confirm their professional judgment［J］. Public Administration Review， 2023， 83（2）： 263-278.
[19]	Kappes A， Harvey A H， Lohrenz T， et al. Confirmation bias in the utilization of others’ opinion strength［J］. Nature Neuroscience， 2020， 23（1）： 130-137.
[20]	Amsel E， Brock S. The development of evidence evaluation skills［J］. Cognitive Development， 1996， 11（4）： 523-550.
[21]	Alon-Barkat S， Busuioc M. Human–AI interactions in public sector decision making： “automation bias” and “selective adherence” to algorithmic advice［J］. Journal of Public Administration Research and Theory， 2023， 33（1）： 153-169.
[22]	Choudhury P， Starr E， Agarwal R. Machine learning and human capital complementarities： Experimental evidence on bias mitigation［J］. Strategic Management Journal， 2020， 41（8）： 1381-1411.
[23]	Veenman M V J， Van Hout-Wolters B H A M， Afflerbach P. Metacognition and learning： Conceptual and methodological considerations［J］. Metacognition and Learning， 2006， 1（1）： 3-14.
[24]	Rollwage M， Fleming S M. Confirmation bias is adaptive when coupled with efficient metacognition［J］. Philosophical Transactions of the Royal Society B， 2021， 376（1822）： 20200131.
[25]	Bandura A. Self-efficacy： The exercise of control［M］. Worcestershire： Worth Publishers， 1997.
[26]	Molenberghs P， Trautwein F M， Böckler A， et al. Neural correlates of metacognitive ability and of feeling confident： A large-scale fMRI study［J］. Social Cognitive and Affective Neuroscience， 2016， 11（12）： 1942-1951.
[27]	Heart T， Zucker A， Roshtov S S I， et al. Investigating physicians' compliance with drug prescription notifications［J］. Journal of the Association for Information Systems， 2011， 12（3）： 235-254.
[28]	Hutton R J B， Klein G. Expert decision making［J］. Systems Engineering；The Journal of The International Council on Systems Engineering，1999， 2（1）： 32-45.
[29]	Adomavicius G， Bockstedt J C， Curley S P， et al. Do recommender systems manipulate consumer preferences？ A study of anchoring effects［J］. Information Systems Research， 2013， 24（4）： 956-975.
[30]	Pescetelli N， Yeung N. The role of decision confidence in advice-taking and trust formation［J］. Journal of Experimental Psychology： General， 2021， 150（3）： 507-526.
[31]	Lu Z， Yin M. Human reliance on machine learning models when performance feedback is limited： Heuris tics and risks［C］//Proceedings of the 2021 CHI Confer ence on Human Factors in Computing Systems， Yokohama， Japan， May 8-13， ACM， 2021： 1-16.
[32]	Schemmer M， Kuehl N， Benz C， et al. Appropriate reliance on AI advice： Conceptualization and the effect of explanations［C］//Proceedings of the 28th International Conference on Intelligent User Interfaces， Sydney，Australia， March 27-31，ACM， 2023： 410-422.
[33]	Dey I， Gnesdilow D， Passonneau R， et al. Potential pitfalls of false positives［C］//Proceedings of the 25th International Conference on Artificial Intelligence in Education， Recife， Brazil， July 8-12， Springer，2024： 469-476.
[34]	Logg J M， Minson J A， Moore D A. Algorithm appreciation： People prefer algorithmic to human judgment［J］. Organizational Behavior and Human Decision Processes， 2019， 151： 90-103.
[35]	Lehmann C A， Haubitz C B， Fügener A， et al. The risk of algorithm transparency： How algorithm complexity drives the effects on the use of advice［J］. Production and Operations Management， 2022， 31（9）： 3419-3434.
[36]	Ning X， Lu Y， Li W， et al. How transparency affects algorithmic advice utilization： The mediating roles of trusting beliefs［J］. Decision Support Systems， 2024， 183： 114273.
[37]	Bonaccio S， Dalal R S. Advice taking and decision-making： An integrative literature review， and implications for the organizational sciences［J］. Organizational Behavior and Human Decision Processes， 2006， 101（2）： 127-151.
[38]	Cortez P， Silva A M G. Using data mining to predict secondary school student performance［C］//Proceedings of the 5th Annual Future Business Technology Conference， Porto， Portugal， April 9-11 ， INSTICC Press， 2008： 5-12.
[39]	Fügener A， Grahl J， Gupta A， et al. Will humans-in-the-loop become borgs？ Merits and pitfalls of working with AI［J］. MIS Quarterly， 2021， 45（3）： 1527-1556.
[40]	Abid L， Masmoudi A， Zouari-Ghorbel S. The consumer loan’s payment default predictive model： An application of the logistic regression and the discriminant analysis in a Tunisian commercial bank［J］. Journal of the Knowledge Economy， 2018， 9（3）： 948-962.
[41]	Tolles J， Meurer W J. Logistic regression： Relating patient characteristics to outcomes［J］. JAMA， 2016， 316（5）： 533-534.
[42]	Bansal G， Nushi B， Kamar E， et al. Updates in human-AI teams： Understanding and addressing the performance/compatibility tradeoff［J］. Proceedings of the AAAI Conference on Artificial Intelligence， 2019， 33（1）： 2429-2437.
[43]	John D R， Scott C A， Bettman J R. Sampling data for covariation assessment： The effect of prior beliefs on search patterns［J］. Journal of Consumer Research， 1986， 13（1）： 38-47.
[44]	Chong L， Zhang G， Goucher-Lambert K， et al. Human confidence in artificial intelligence and in themselves： The evolution and impact of confidence on adoption of AI advice［J］. Computers in Human Behavior， 2022， 127： 107018.
[45]	Tang M， Liao H. From conventional group decision making to large-scale group decision making： What are the challenges and how to meet them in big data era？ A state-of-the-art survey［J］. Omega， 2021， 100： 102141.
[46]	Vuorre M， Metcalfe J. Measures of relative metacognitive accuracy are confounded with task performance in tasks that permit guessing［J］. Metacognition and Learning， 2022， 17（2）： 269-291.