Output Type Before Quality: A Standards-Derived XAI Admissibility Rubric for Autonomous-Driving Safety

Safety standards require specific types of evidence that most popular XAI tools cannot provide: The paper identifies a critical "evidence-type gap" where SHAP, the most commonly recommended explainability method for autonomous driving, produces ranked feature lists that cannot be converted into the cause-and-effect chains that safety standards (ISO 26262, ISO 21448, AMLAS) actually require for assurance cases.

Causal explanation methods are structurally necessary for certain critical stages of autonomous vehicle safety: Causal XAI approaches are required to meet derived safety criteria at three key lifecycle stages—hazard identification, incident investigation, and data management—where they close 50-62% more of the evidence gap compared to other method types, while simpler correlational or language-based methods may suffice at other stages.

Meeting structural admissibility standards is necessary but not sufficient for real safety compliance: The paper's rubric can identify whether an XAI method could theoretically satisfy safety requirements, but cannot verify whether the actual explanations it produces are correct—meaning organizations must still validate that the cause-and-effect relationships identified are genuinely accurate, not just plausibly formatted.

Current XAI evaluation practices are misaligned with autonomous driving safety requirements: The XAI research community organizes methods by technique type (saliency maps, feature attribution, counterfactuals), but safety standards organize requirements by evidence type and lifecycle stage, creating a fundamental mismatch that practitioners must navigate.