Assessing Driver’s Trust,Compliance,and Reliance in an Automated Flood Warning System: Effects of Error Types and System Reliability

Abstract

The increasing threat of inland flooding due to precipitation changes and floodplain development necessitates efficient real-time flood detection and communication methods. While automated floodwarning systems facilitate such communication, they are susceptible to errors like false alarms and misses, which could undermine drivers’ trust during flood events. This study examined how system accuracy and error type impact perceived system reliability, as well as drivers’ trust and behaviors. Our results showed that both false alarms and misses lowered drivers’ perceived system reliability, and drivers were more inclined to follow recommendations from a system with higher reliability compared to one with low reliability. Misses and false alarms influenced drivers’ reliance and compliance behaviors differently. These findings help predict how system reliability level and error type shape drivers’ responses to automated flood-warning systems, potentially contributing to their design and calibration.

Automated systems help communicate flood risks to human drivers efficiently (Sazara et al., 2022; Wang et al., 2024). However, when predicting floods on the road, these automated systems can make errors such as a false alarm, which falsely claims that there is a flood ahead, or a miss, which fails to report the presence of a flood. Different error types can alter drivers’ perception and interaction with automated systems, such as perceived trustworthiness (Guznov et al., 2016), trust level (Chen et al., 2021), and their compliance-reliance behaviors (Azevedo-Sa et al., 2020b).

There have been mixed results when comparing false alarms and misses in their effects on trust in automation (Azevedo-Sa et al., 2020a; Chancey et al, 2015; Chen et al., 2021; Geels-Blair et al., 2013). For example, Chancey and colleagues (2015) found no significant impact of error type on operators’ trust in the signaling system when they performed a secondary tracking task in addition to a primary flight simulation task. The secondary tracking task required participants to maintain a steady flight altitude in a flight simulation using a joystick. Another study found that false alarms impacted users’ trust more negatively than misses when participants performed an Unmanned Aerial Vehicle (UAV) control task with varying automation reliability levels (Geels-Blair et al., 2013). In the domain of flood warnings, Sawada and colleagues (2022) used statistical modeling to study the relationship between social collective trust and false alarms in the flood early warning system. They did not collect the data empirically but used a mathematical model to simulate the interaction between flood and social collective trust. They found that false alarms reduced social collective trust. However, there has been little empirical research that directly compares the impacts of false alarms and misses on drivers’ trust in flood-warning systems. The current study focused on an automated flood-warning system designed for drivers on the road. We investigated the influence of system reliability and error type on drivers’ trust and their decisions on whether to follow the flood-warning system or not. Given that missing a flood warning from the system can lead to more harmful accidents than a false alarm, we predicted that misses would impair drivers’ trust more than false alarms.

This study used a between-subjects experimental design and manipulated error type (false alarms vs. misses) and system reliability (60% vs. 90%) to investigate their effects on drivers’ trust and their decisions on whether to follow the flood-warning system or not. One hundred thirty-one participants were recruited from an online crowd-sourcing platform named Prolific, and were randomly assigned to one of the four conditions: 60%-false alarms, 90%-false alarms, 60%-misses, and 90%-misses. In each condition, participants experienced 10 different driving scenarios with various flood warnings. Near the end of each drive, we collected drivers’ decisions on whether they would follow the system’s recommendation (i.e., go straight, make a turn, undecided). The system then provided feedback on the actual flood situation to the drivers. Drivers’ subjective trust, perceived risk, and perceived system reliability were then measured.

Our results showed that lower system reliability (60%) led the drivers to report lower subjective trust and was associated with drivers following the system’s recommendations less often, compared to when the system reliability was high (90%). Contrary to our prediction, misses did not reduce trust more than false alarms. Moreover, we found significant differences in participants’ decisions when interacting with miss-prone versus false-alarm-prone systems. Specifically, participants were more likely to be undecided when interacting with miss-prone systems, whereas they tended to not follow the system’s recommendation when interacting with false-alarm-prone systems.

This study provided a theoretical understanding of how error type and system reliability influence drivers’ trust and decisions on the road when interacting with an automated flood-warning system. The results could support the design of human-centered automated flood-warning systems. We recommend developers prioritize higher system reliability, as it leads to higher trust levels and are associated with drivers following the system’s recommendation more. Depending on the road situation, developers can calibrate the ratio of false alarms and misses among all errors produced by the system to promote safer driver behaviors, potentially improving the overall road safety and effectiveness of the flood warning system. For instance, in high-risk areas where the potential danger of missing a flood exceeds the inconvenience caused by false alarms, warnings can be issued more often to ensure critical warnings are not missed. It can be dangerous for drivers to remain undecided in high-risk areas, as this could prevent timely action to avoid flood-related hazards. Thus, the number of misses should be minimized in high-risk areas. Future studies can investigate optimal calibration strategies on the ratio between false alarms and errors to maximize drivers’ safety when experiencing floods.