社会与工程心理学研究室知识库

随着自动化系统和人工智能的日益普及，人是否能与自动化发生和谐互动变得越来越重要。其中，自动化信任在人一自动化交互中扮演了重要角色。不合理的自动化信任会造成对自动化系统的滥用或拒用，影响人机系统的整体表现，损害系统使用者的健康。因此，研究哪些因素会影响自动化信任具有重要的现实意义和理论意义。大量基于实验室的研究发现，自动化系统本身的可靠性是影响人一自动化交互的基础因素，对自动化信任起到重要的预测作用。但是，在真实世界发生的人机交互过程中，系统的实际可靠性通常既不透明也难以更改。因此，研究者开始关注那些可以对自动化信任产生影响但又独立于实际可靠性之外的设计性因素，以此来达到校准自动化信任的目的。现有研究对自动化系统的外形特征(如外表)或交互性特征(如情感反馈)所造成的影响己经做了大量研究，但对自动化系统所具有的群体身份特征的研究则刚刚起步。从现实角度来说，随着自动化系统的智能性程度越来越高，人一自动化高度协作的场景越来越多，自动化设备会被自然地赋予更多的社会身份特征，成为不同群体的成员，群体身份特征是否会影响人一自动化信任的产生与发展是有待研究的重要问题。己有研究初步发现，个体对于和自己属于同一群体的人工智能的评价更积极，但现有研究尚未深入探讨在长期接触中，群体身份对信任的影响。

本研究将进一步探索群体身份特征对自动化信任的影响，通过使用最小群体范式对自动化设备的群体身份进行操控，同时使用TNO信任任务对人一自动化辅助设备之间的信任问题进行探索。实验一中，采用了2(辅助系统的群体身份特征:内群体身份、外群体身份)X3(辅助系统的可靠性水平:95%, 70%, 45%)的混合设计，结果首先验证了辅助系统可靠性的作用，可靠性水平越高，被试的信任程度越高，进而我们发现群体身份会影响使用者对自动化系统的信任，相比于具有外群体身份的自动化辅助系统，被试更加信任和自己处于相同群体的自动化辅助设备。实验二进一步探讨群体身份对信任的影响是否依赖于智能系统本身的外形特点以及是否能够在可靠性下降时仍然具有保护作用。实验二采用了2 (辅助系统的群体身份特征:内群体身份、外群体身份)X3(辅助系统的可靠性水平:95%, 70%, 45%) × 2(外形特征:拟人形象，非拟人形象)的被试内设计，其中拟人形象与实验一相似，非拟人形象则为台式计算机的外形;此外，每个被试均经历了可靠性水平从高到低的变化。结果发现:首先，群体身份的主效应依然显著，被试更信任具有内群体身份的辅助系统;其次，拟人性特征与群体身份没有交互作用，说明群体身份的影响不依赖与智能系统的外形特征;最后，群体身份与辅助系统的可靠性水平的交互作用不显著，说明群体身份的影响并不随可靠性下降而减少。实验三进一步考察了在可靠性降低到低谷后重新恢复时，被试信任的变化情况。实验三采用了2(辅助系统的群体身份特征:内群体身份、外群体身份)X4(辅助系统的可靠性水平:95%, 70%, 45%, 95%)×X2(外形特征: 拟人形象，非拟人形象)的被试内设计，结果表明:首先，群体身份显著的主效应依然显著;其次，外形特征与群体身份仍然没有交互作用，可靠性水平与群体身份没有交互作用;最后，相比与具有外群体身份的智能系统，面对具有内群体身份的智能系统，用户在可靠性降低后重新恢复的过程中，能够更快达到高水平的信任程度。综合以上实验，本研究发现辅助系统群体身份对自动化信任存在鲁棒影响。首先，被试对具有内群体身份辅助系统的信任水平，无论是在主观信任还是行为信任上，都显著高于具有外群体身份的辅助系统的信任水平。其次，这一影响并不依赖于系统拟人形象特征而存在;最后，即便是经历辅助系统的可靠性水平连续降低，升高的变化，被试依然更信任具有内群体身份的辅助系统。本研究为推进人一自动化交互提供了新的研究角度，并为人一自动化信任的校正提供了新的方法和支撑。

Trust in automation is a key element in human-automation interaction, since inappropriate trust is linked inappropriate use of automation, misuse or disuse, cause damage to human-automation performance and user's health. Therefore, it is of great realistic and theoretical significance to study factors that could affect tic development of trust in automation. Laboratory research have found that the reliability of automation is the key factor of automation trust, however, in the actual human-computer interaction process, the actual reliability of the system is often non-transparent and difficult to change. Therefore, researchers started to focus on design factors that could influence trust in automation. And lots of researches on the impact of the personalized features (such as appearance) or interactive features (such as emotional feedback) on automation system have been conducted, but the research on the group identity of automation system has just started. From a practical point of view, since the intelligence degree of automation improved rapidly, there will be more human-automation collaboration, automation system will naturally be given more social identity features and become members of different groups. It's of great importance to study the issue if group identity of automation system will affect operator's trust in automation. Some studies have found that individuals have more positive evaluations of artificial intelligence which share the same group membership with himself. This study will further explore the influence of group identity on automation trust. The minimum group paradigm was adopted to control the group identity of automation system, and TNO trust task was adopted to investigate trust in human-automation interaction. In Experiment 1，a mixed design of 2 (group identity of the decision support system: in-group, out-group) X3 (reliability level of the system: 95%, 70%, 45%) was used, group identity as within-subject factor. The effect of reliability on automation trust were verified, the higher the reliability of the system, the higher trust of the system. .And results also showed that in-out group identity have significant influence on automation trust, participants were more likely to trust in-group automation system. In order to further investigate the effect of automation group identity on trust in automation, automation without human feature was added in Experiment 2, a 2 (group identity of the decision support system: in-group, out-group) X3 (reliability level of the system: 9S%, 70%, 45%) X2 (anthropomorphism: humanoid, non-humanoid) within-subjects design was used. In the humanoid condition, we used the same material that used in experiment 2, and in the non-humanoid condition9 we designed two desk computers without human feature. And the reliability of the system was presented in fixed order from 95% to 70%, and then 45%. Results showed a significant main effect of group idcntity9 and participants have higher trust level on automated system with in-group identity. And we did not end significant interaction between group identity, reliability and anthropomorphism which means that the effect of group identity does not depend on anthropomorphic of automated systcm9 and not reduce when the reliability of system gradually declines. In experiment 3, after gradually decline from 9S% to 45%, the reliability of the system then directly backs up to 95%. We conducted a 2 (group identity of the decision support system: in-group, out-group) X 4 (reliability level of the system: 95%, 70%, 45%, 95%) X 2 (Anthl0opomorphism: humanoid, non-humanoid) within-subjects design. Results showed significant main effects of the group identity and reliability of the automated system. And interaction between them was not significant. Interaction between group identity and anthropomorphism was not significant either. Participants showed significant higher trust in automation with in-group identity. Post hoc test of block level showed that when the reliability back up to 9S%, trust in automation with in-group identity recovered faster than out-group automation. In conclusion, studies showed a robust effect of group identity on automation trust. First of all, participants' trust in automation system with in-group identity were significantly higher than automation with out-group identity, both in subjective trust and behavioral trust. Secondly, the effect of group identity does not depend on the system anthropomorphic feature. Thirdly, even if the reliability level experienced a continuously decrease and then directly back up to original level, the subject still has higher trust on the system with the in-group identity. This study provides a new perspective for the advancement of human-automation interaction and a new method and support for the calibration of trust in automation.