Toward Safe AI in Autonomous Vehicles: Challenges, Standards, and an Integrated Safety Engineering Framework
Amr Alsayed Mohamed, Heba Kamal Aslan · Procedia Computer Science · 2025
Ensuring the safety of AI in autonomous vehicles (AVs) poses unique challenges due to the non-deterministic and adaptive behavior of learning-enabled components. While ISO 26262 and ISO 21448 provide foundations for functional and intended safety, they fall short in addressing AI-specific risks. ISO/PAS 8800:2024 fills this gap with the first dedicated guidance for AI-based automotive systems. This paper surveys key safety assurance challenges in AI components—including explainability, traceability, robustness, and life cycle adaptation—and outlines the structure and principles of ISO/PAS 8800. We propose a practical engineering framework that integrates AI safety across five development stages: from requirements definition to post-deployment monitoring. The framework operationalizes ISO/PAS 8800, complements legacy standards, and supports continuous assurance workflows. By bridging emerging standards with real-world engineering practices, this work offers actionable guidance for researchers, developers, and certification stakeholders seeking to design safe, auditable, and certifiable AI systems for AVs.