Christiane Attig, Tim Schrills, Markus Gödker, Christiane Wiebel, Patricia Wollstadt, Thomas Franke, "Enhancing Trust in Smart Charging Agents—The Role of Traceability for Human-Agent-Cooperation", HCI International 2023 – Late Breaking Papers. HCII 2023., vol. 14059, 2023.

Introduction and background: The EU aims for climate neutrality by 2050, which necessitates a comprehensive transformation of the transport sector, including a 90% reduction in emissions [3]. Consequently, the demand for electric vehicles (EVs) will strongly rise within the next years. It has been argued that this demand will pose a challenge for the stability of the power grid [6] – particularly if EVs are charged with renewable electricity, which is subject to strong fluctuations in supply and might not be flexible enough to meet user needs at all times [1]. Conversely, EVs offer a great potential for increasing grid stability through bidirectional charging, that is, EVs can store or provide excess energy to the grid as needed [7]. As a consequence, the complexity of the charging process increases (e.g., in terms of planning, technical understanding). Thus, the collective benefit of grid stability may come at a cost for the individual user, who might face a restriction of personal resources (e.g., time, comfort [6]). Smart charging agents relying on techniques from the field of artificial intelligence (AI) offer one solution to combine user comfort with optimal utilization of renewable energy resources. To realize this solution, smart charging agents need to be perceived as cooperative partners within a joint activity [5] who assist users to achieve not only individual, but also collective goals. Therefore, it is crucial to maximize users’ perception of advantages from cooperating with the system. One core variable for enhancing cooperation between users and an AI system such as a smart charging agent is trust, which can be increased by AI traceability [9]. Objective and significance: The present research aimed at understanding the potential of AI traceability (i.e., transparency, understandability, and predictability [8]) for enhancing trust in the context of smart charging in car sharing fleets. Method: For an online experiment, a basic algorithm was designed to calculate the resource efficiency of booking an EV from a car-sharing fleet based on simulated data. The data was based on 10 features (e.g., time of booking start and end, expected network power demand, likelihood of a peak load). In five subsequent observation blocks, N = 57 participants were asked to observe 10 cost calculations made by the algorithm (i.e., 50 observations in total). After each observation block, participants rated their subjective experience with the algorithm (i.e., trust via the Facets of Systems Trustworthiness scale [4]; traceability with the Subjective Information Processing Awareness scale [8]). To evaluate participants’ ability to predict the algorithm’s results, a performance block followed, in which participants were asked to estimate booking costs based on the disclosed information (20 estimations in total). The traceability of the algorithm was experimentally manipulated by varying the amount of disclosed information that formed the basis of the cost calculation (high, medium, low information; between-factors design). Results: Using planned contrast analyses, it was shown that trust partially varied with the amount of disclosed information (higher amount of information related to higher reported trust). Moreover, traceability was partially higher in the high information group than the medium and low information groups. Analyses of the three subscales of traceability revealed that effects were particularly pronounced for understandability and predictability, while no effect was found for transparency. In addition, participants’ performance in estimating the booking costs did not vary with amount of disclosed information. Discussion: While additional information enhanced subjective experiences of trust, understandability, and predictability of a smart charging agent for EV car sharing, they did not improve transparency ratings and estimation of the algorithm’s output. This pattern of results might reflect an explainability pitfall [2]: Users of smart charging agents might trust these systems more as traceability increases, regardless of how well they understand the system.