Explainability in Automated Driving: From Spatial Attention to Human-Centred Reasoning

Abstract

Automated Vehicles (AVs) are developing rapidly and promise to improve road safety. AV systems are equipped with advanced AI techniques to perceive, learn, decide, and act, yet the decision-making process underpinned by complex AI constitutes a black box for human users. While human understanding of AI-based decision-making is critical to building trust, acceptance, and efficient human-machine cooperation, existing algorithmic approaches provide cause-effect relations in decisions but fail to account for the psychosocial and cognitive dynamic nature of supervising an AV. There is a lack of comprehensive framework addressing this algorithmic transparency with the cognitive and affective dynamics of the human recipient, leaving the human side of the explanation transaction theoretically underdeveloped. We propose a human-centred explainability framework that repositions AV explanation as a cognitive human-machine interaction problem. The framework is operating across spatial, semantic, and cognitive registers. It articulates where a system attends, what it recognises, and why that recognition contributes to better explanation. These levels integrate semantic awareness to enrich causal reasoning, while spatial information contextualise explanations. The preliminary framework is grounded in naturalistic field observations in instrumented vehicles, collecting video data and field notes complemented by post-drive semi-structured interviews. Thematic analysis yielded a driving context-specific explanatory vocabulary, formalising conditions under which users seek, process, and integrate AV explanations. Drawing on cognitive science and human factors theory, we derive design principles for adaptive explanation delivery, leveraging multimodal large language models as the generative engine for contextualised natural language explanations responsive to user expertise and situational urgency. This work demonstrates that the next frontier in AV explainability should be more aligned with human cognition.