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

Researchers identify a critical gap between safety standards for autonomous driving and explainable AI (XAI) methods: current popular XAI techniques like SHAP produce outputs that don't match the evidence types required by ISO and safety standards. The study derives 19 evidentiary criteria across 7 lifecycle stages and determines that causal XAI methods are structurally necessary for hazard identification and incident investigation, while correlational methods suffice elsewhere.

This research addresses a fundamental structural mismatch in autonomous driving safety assurance that has significant implications for how the industry validates ML-based systems. While safety standards like ISO 26262 and ISO/PAS 8800 explicitly require directed cause-and-effect chains and quantified interventional effects as evidence, the dominant XAI literature organizes solutions by output type—saliency maps, feature attribution, and SHAP rankings—without regard for regulatory demands. The paper demonstrates that SHAP, despite being the most-recommended XAI method for autonomous driving, cannot be converted into the evidentiary chains standards require, creating what the authors term an evidence-type gap. The research is methodologically rigorous, deriving testable criteria from multiple ISO standards and safety frameworks (AMLAS, ISO 26262, ISO 21448, ISO/PAS 8800) and validating predictions against real-world driving data across 1,996 clips and 79,840 rows. Causal XAI emerges as structurally required at three critical lifecycle stages—hazard identification, incident investigation, and data management—where it closes gaps of 50-62% compared to alternatives. For autonomous vehicle developers and safety engineers, this work clarifies that XAI method selection cannot be driven by popularity or ease of implementation, but must align with specific regulatory evidence demands at each lifecycle stage. The distinction between structural admissibility and fidelity validation also establishes that passing this rubric is necessary but insufficient for compliance—validation of actual causal relationships remains an open challenge. This research will likely accelerate industry adoption of causal inference methods in safety assurance workflows.

• →Popular XAI methods like SHAP produce outputs structurally incompatible with safety standard evidence requirements for autonomous driving.
• →Causal XAI methods are necessary for three critical ADS lifecycle stages but correlational methods may suffice for others.
• →Safety assurance should be driven by regulatory evidence demands per lifecycle stage, not by XAI method popularity.
• →A derived rubric across 19 criteria and 7 stages provides structured guidance for XAI method admissibility in ADS development.
• →Method admissibility is necessary but not sufficient—validating actual causal relationships remains the open assurance challenge.