A Transparency Paradox? Investigating the Impact of Explanation Specificity and Autonomous Vehicle Perceptual Inaccuracies on Passengers

TL;DR

This study interrogates how explanation specificity and autonomous-vehicle perception errors shape passenger responses, framing a transparency paradox where higher transparency can increase anxiety when perception is imperfect. Using a within-subject lab design with immersive driving simulation and a hybrid rule/data-driven explainer, the authors compare Abstract, Specific(5), and Specific(50) explanations across controlled error rates. Key findings show no clear safety advantage for high transparency, while anxiety elevates as perception errors rise; specificity helps safety when errors are low but worsens anxiety when errors are salient, and takeover tendencies are not monotonically tied to error exposure. The work highlights the need for adaptive, context-sensitive explanation interfaces that balance transparency with user comfort, with implications for AV design, regulation, and trust calibration.

Abstract

Transparency in automated systems could be afforded through the provision of intelligible explanations. While transparency is desirable, might it lead to catastrophic outcomes (such as anxiety), that could outweigh its benefits? It's quite unclear how the specificity of explanations (level of transparency) influences recipients, especially in autonomous driving (AD). In this work, we examined the effects of transparency mediated through varying levels of explanation specificity in AD. We first extended a data-driven explainer model by adding a rule-based option for explanation generation in AD, and then conducted a within-subject lab study with 39 participants in an immersive driving simulator to study the effect of the resulting explanations. Specifically, our investigation focused on: (1) how different types of explanations (specific vs. abstract) affect passengers' perceived safety, anxiety, and willingness to take control of the vehicle when the vehicle perception system makes erroneous predictions; and (2) the relationship between passengers' behavioural cues and their feelings during the autonomous drives. Our findings showed that passengers felt safer with specific explanations when the vehicle's perception system had minimal errors, while abstract explanations that hid perception errors led to lower feelings of safety. Anxiety levels increased when specific explanations revealed perception system errors (high transparency). We found no significant link between passengers' visual patterns and their anxiety levels. Our study suggests that passengers prefer clear and specific explanations (high transparency) when they originate from autonomous vehicles (AVs) with optimal perceptual accuracy.

Figures (14)

• Figure 1: Driving simulation setup for the study. The setup included a VR headset, steering wheel, brake and acceleration pedals, screen, and arcade seat. The screen shows a pedestrian crossing at a crosswalk.
• Figure 2: High-level architecture of our simulation software. DReyeVR uses Unreal engine and extends CARLA simulator, which also builds on Unreal engine. DReyeVR extends CARLA by adding VR functionalities, vehicular and ambience sounds, eye tracker data logging, and additional sensors, among others. Our explainer model, which is both rule-based and data-driven, receives ground truth data from CARLA or DReyeVR and generates explanations for predicted actions. The post-processing script allowed us to modify the generated explanations as we desire.
• Figure 3: Scenario routes. Red: Abstract, Green: Specific(5), Blue: Specific(50). Each route is a loop and overlaps with others at some points.
• Figure 4: Sample screenshots and the generated explanations (including observations announcement and causal explanations) from the three driving scenarios. Heatmaps of gaze points from all the participants are plotted over the images, indicating areas of interest. In the Abstract scenario (Figure \ref{['fig:vague']}), all movable/dynamic non-human actors are referred to as 'Vehicle'. Thus, a cyclist was referred to as a vehicle. Figure \ref{['fig:spec5']} depicts a scene from the Specific(5) scenario in which the AV's perception system accurately identified and classified a motorbike and provided a fine-grained explanation for this. In the Specific(50) scene (Figure \ref{['fig:spec50']}), the AV's perception system misclassified a pedestrian as a cyclist. The fine-grained/specific explanation provided exposed this error.
• Figure 5: Study procedure. Eye calibration was done with the VR headset; participants drove for two minutes, participants experienced each of the 4 mins scenarios in counterbalanced order and completed the Feeling of Anxiety, Perceived Safety, and Takeover Feeling Questionnaire (APT Scale) in between each scenario. Participants were debriefed.