The Necessity of a Holistic Safety Evaluation Framework for AI-Based Automation Features

AI components previously considered "low-risk" can pose serious safety threats: The paper reveals that Quality Management components—traditionally excluded from rigorous safety evaluations because they lack safety requirements—can actually introduce hazardous risks when AI is involved, particularly in perception systems used for autonomous driving.

Current safety standards are insufficient for AI-based automation: Existing safety frameworks (SOTIF and Functional Safety) were not designed to address the unique risks posed by AI systems. The paper argues that compliance with newer AI safety standards like ISO/PAS 8800 requires fundamentally rethinking how we evaluate these components.

Perception algorithms in autonomous vehicles need comprehensive safety analysis: The case studies demonstrate that deficiencies in AI-driven perception systems can lead to unintended behaviors with "critical safety implications," even when those systems weren't originally classified as safety-critical.

A holistic approach is essential, not piecemeal evaluation: Rather than evaluating AI components in isolation using traditional safety methods, the paper emphasizes the need for comprehensive analysis that integrates Functional Safety, SOTIF, and AI-specific safety standards together.

Safety frameworks must evolve to keep pace with AI deployment: The paper's core message is that as AI becomes more integrated into safety-critical systems like autonomous vehicles, our safety evaluation processes must be revised and updated to address these "evolving challenges."