AI透明度和可靠性对人-AI协同决策绩效与信任的影响  

The Impact of AI Transparency and Reliability on Human-AI Collaborative Decision-Making Performance and Trust

作　　者：王栩金凤 杨奕诚 李伟涛 张婷茹 WANG Xu-jin-feng;YANG Yi-cheng;LI Wei-tao;ZHANG Ting-ru(Institute of Human Factors and Ergonomics,College of Mechatronics and Control Engineering,Shenzhen University,Shenzhen 518006,China)

机构地区：[1]深圳大学机电与控制工程学院人因工程研究所,广东深圳518060

出　　处：《人类工效学》2024年第2期53-60,共8页Chinese Journal of Ergonomics

摘　　要：目的探索AI透明度和可靠性对人-AI协同中人的决策行为及信任的影响。方法开展一项人-AI协同收入预测实验,实验采用双因素混合设计,其中AI透明度(低、中、高)为组间变量,AI可靠性(60%、75%、90%)为组内变量,共有54名受试者(男、女各27人)参与实验。实验收集了人的决策类型(四种情况:正确接受、错误接受、正确拒绝、错误拒绝)、决策时间、信任、决策正确率和AI依从率数据。结果当AI可靠性较高(75%和90%)时,透明度对人的决策影响不显著;当AI可靠性较低(60%)时,更高的透明度提高了人对AI的依从率。只有在AI给出正确建议时依从率才会增加,AI出错时的依从率(即过度依赖)不受透明度影响。此外,透明度对正确拒绝决策类型没有显著影响。结论透明度有助于提高人对AI的正确依赖,且透明度的增加对人识别AI错误的能力并没有显著的提高作用。Objective To explore the impact of AI transparency and reliability on human decision-making behavior and trust in human-AI collaboration.Methods An experiment on human-AI collaborative income judgment which adopted a two-factor mixed design with AI transparency(low,medium,and high)as the between-group variable,and AI reliability(60%,75%,and 90%)as the within-group variable was conducted.54 participants(27 males,27 females)were enrolled and the following data were collected:decision-making types(four situations:correct acceptance,incorrect acceptance,correct rejection,and incorrect rejection),decision-making time,trust,human accuracy and AI compliance.Results When AI reliability was high(75%and 90%),transparency had no significant impact on human decision-making,while when AI reliability was low(60%),higher levels of transparency increased human compliance rates with AI.But further analysis suggested that the compliance rate only increased when AI made correct advice,but not when AI made mistakes.In addition,transparency had no significant impact on correct rejection.Conclusion Transparency increased appropriate reliance towards AI.However,increasing transparency level did not significantly improve human’s ability to identify AI errors.

关 键 词：机器学习 人机交互 人-AI协同 透明度 可靠性 信任 AI依从率 用户体验 

分 类 号：TB472[一般工业技术—工业设计]