Stakeholder-Specific Autonomous Vehicle Explanations: Ecologically Valid Evaluation beyond Simulator-Only Studies

Autonomous vehicles rely on deep-learning and foundation-model stacks whose decisions are hard to inspect after safety-critical incidents. Explainable AI for driving spans post-hoc saliency, interpretable scene graphs, and vision-language rationales, yet technical explanations often lack verified faithfulness to perception, planning, and control. Human validation remains fragmented across simulators and short exposures, while real-time foundation-model deployment faces latency and explanation–action alignment risks. This empirical study synthesizes human-centric HMI research, interpretable and post-hoc methods, and foundation-model architectures to characterize these assurance gaps. We propose a four-phase validation protocol, a multi-layer explanation faithfulness index, and a tiered latency–alignment architecture for constrained in-vehicle hardware. The study addresses faithfulness in multi-hazard scenarios, stakeholder-specific trust formats, HMI timing and modality trade-offs, and hybrid architectures that preserve alignment under operational constraints.