Building Trust in Safeguards Voice User Interfaces: An Experimental Approach

: The research community is exploring multiple potential applications of artificial intelligence and its enabling technologies for international nuclear safeguards monitoring and verification. Based on successful implementation in other domains or early safeguards development, some applications hold significant promise. However, those benefits can only be realized if users place appropriate levels of trust – that is, reliance in the system – when delegating tasking or accepting recommendations. Prior work by members of this team has examined user trust in visual-based systems focused on model performance, but voice user interfaces such as those in digital assistants pose unique opportunities. While some safeguards challenges for voice user interfaces – noise in the operational environment, nonnative English speakers, etc. – have been addressed in other domains, there remain aspects of voice user interfaces that are unique for trust in safeguards applications. In this work, we use human performance testing to evaluate the factors that impact trust in voice user interfaces within the safeguards context. Our aim is to provide actionable recommendations to the software development community to optimize system performance of voice user interfaces for safeguards. In this paper, we describe our task analysis and safeguards trust factors. We detail our experimental prioritization, focusing on two safeguards tasks: seal checking and nuclear material measurement. We describe early results and our experimental plans moving forward.