Trust and Uncertainties: Characterizing Trustworthy AI Systems Within a Multidimensional Theory of Trust

Abstract Trust is a basic feeling and attitude. It shapes human relations as the glue that holds groups and even societies together. Now that AI is increasingly prevalent in our daily lives, the extent to which we can trust these systems has become a key question. It can be discussed from a psychological person perspective (“Under which conditions are we inclined to trust AI systems?”) or from an objective systems perspective (“Under which conditions is a system worthy of trust, and to what degree?”). To offer a general framework for comparing different systems, we adopt a system-level perspective, abstracting from subjective psychological conditions. An especially innovative aspect of this comparative framework is its integration of two dimensions: uncertainty (currently a hot topic in AI research) and commitment (a rather new one for AI systems). This allows us compare certain AI systems, like ChatGPT or autonomous cars, to more familiar systems, like classical (non-autonomous) cars and (prototypical) democratic institutions. We can therefore clarify in which dimensions they differ. This overview can be used both to understand specific features of AI systems and to reveal deficits in their trustworthiness that must be overcome to make AI systems acceptable. Despite intense and widespread discussion of whether and to what degree we can trust AI systems, we still lack a general framework for any systematic comparison of trustworthiness. Our account is supposed to develop further and improve the famous analysis of Glikson & Wooley (2020) by proposing a multidimensional framework of trustworthiness, with three central steps. First, drawing on noteworthy articles in the existing literature, we identify six central dimensions of trust from a general perspective: objective functionality, transparency, uncertainty (quantification), embodiment, immediacy behaviors, and commitment. Second, we develop a more detailed perspective, partially characterizing each dimension by detailing several of its specific features. Finally, we show how we can evaluate each feature of any dimension (implemented as low, medium, or high) and thereby calculate an average value for each dimension a system has. This results in a multidimensional account of trust that allows us compare different systems’ trustworthiness as a basis for the future development of AI systems.