Date of Award
2023
Document Type
Thesis
Degree Name
Master of Arts (MA)
Department
Psychology
Committee Chair
Lisa Vangsness
Committee Member
Nathan Tenhundfeld
Committee Member
William Mackzie
Subject(s)
Trust, Explanation, Artifical intelligence, Automation--Psychological aspects, Automation--Human factors, Explainable AI (XAI)
Abstract
Automation has been implemented in a range of machinery. Providing supplementary information about system processes (i.e., explainability) could mitigate over-reliance and enhance operator awareness of potential anomalies. Trust plays a critical role in human- automation collaboration, as over-trust can lead to misuse or over-reliance, while under-trust can result in disuse or the failure to engage automation when it could enhance performance. Dynamic learned trust and situational trust fluctuate during an operator's interaction with a system and can be influenced by the rate of system failures or workload respectively. Design features like explainability can impact perceived usefulness and help users identify system errors and competencies. This study investigates the impact of explainability on user performance in a quality control task. Participants were randomly assigned to either receive training on system failures or not, with varying quality control inspection quotas to simulate various taskloads. The study used a mixed design and measured participants' use of system recommendations and accuracy over time. The results revealed that explainability enhanced accuracy in moderate to lower-quota environments, but this effect was contingent on participants’ receiving training. Explainability also increased reliance levels in lower-quota situations; however, as workload intensified, the time it took for users to determine suitable reliance diminished. While using explainability to assist users in fine-tuning reliance strategies and enhancing accuracy is advisable, it should not serve as a substitute for training, particularly for individuals in high workload environments. Automation has been implemented in a range of machinery. Providing supplementary information about system processes (i.e., explainability) could mitigate over-reliance and enhance operator awareness of potential anomalies. Trust plays a critical role in human- automation collaboration, as over-trust can lead to misuse or over-reliance, while under-trust can result in disuse or the failure to engage automation when it could enhance performance. Dynamic learned trust and situational trust fluctuate during an operator's interaction with a system and can be influenced by the rate of system failures or workload respectively. Design features like explainability can impact perceived usefulness and help users identify system errors and competencies. This study investigates the impact of explainability on user performance in a quality control task. Participants were randomly assigned to either receive training on system failures or not, with varying quality control inspection quotas to simulate various taskloads. The study used a mixed design and measured participants' use of system recommendations and accuracy over time. The results revealed that explainability enhanced accuracy in moderate to lower-quota environments, but this effect was contingent on participants’ receiving training. Explainability also increased reliance levels in lower-quota situations; however, as workload intensified, the time it took for users to determine suitable reliance diminished. While using explainability to assist users in fine-tuning reliance strategies and enhancing accuracy is advisable, it should not serve as a substitute for training, particularly for individuals in high workload environments.
Recommended Citation
Chesser, Amber F., "Learning to calibrate trust through explainability" (2023). Theses. 617.
https://louis.uah.edu/uah-theses/617