Presentation
Trust and Distrust in AI: Examining Trustworthiness of Large Language Models in the Nuclear Human Factors Domain
DescriptionThe current study examined how trust and distrust differ and relate to three attributional abstractions of trust in a signal detection task in the nuclear power plant (NPP) domain. With the growth of powerful technologies such as artificial intelligence, integration of such technologies may yield efficiency and financial benefits, as well as supporting an aging workforce. In this study, participants were tasked with judging statements about the circulatory water system in NPPs as true or false, while assisted by an NPP-trained LLM. System transparency was manipulated with color coding statements based on their Factual precision in Atomicity Score (FActSCORE), and explainability was manipulated with the addition of retrieval-augmented generation (RAG). Trust significantly predicted all three attributional abstractions, while distrust only predicted performance-based trust. Distrust was negatively associated with trust, though the presence of RAG removed the correlation. These findings suggest that trust and distrust are separate mechanisms, and the presence of RAG may dampen the effect of distrust on performance. Furthermore, trustworthiness features should be incorporated in the design and implementation of AI technologies.
Event Type
Lecture
TimeTuesday, October 20th4:50pm - 5:10pm PDT
Location