Safety design concepts for statistical machine learning components toward accordance with functional safety standards

Current safety standards don't adequately address machine learning risks: The paper notes that international functional safety standards for electrical and programmable systems "do not recommend to use SML in safety critical systems so far," highlighting a significant gap between AI deployment and existing safety frameworks.

Five technical safety concepts are needed for ML in critical systems: The authors propose organizing five distinct technical safety concepts (TSCs) specifically designed to help machine learning components comply with functional safety standards, suggesting that one-size-fits-all approaches are insufficient.

Explainability is essential for safe ML deployment: The paper emphasizes that development processes should incorporate eXplainable AI (XAI) methods alongside established automotive development standards (Automotive SPICE) to improve both explainability and reliability before systems are deployed.

Real-world incidents demonstrate the urgency of this problem: The authors reference "crucial incidents and accidents" caused by machine learning misjudgments in safety-critical systems, underscoring that this is not a theoretical concern but an active problem requiring immediate solutions.

Trade-offs exist between safety approaches: The paper compares different technical safety concepts across cost and difficulty factors, suggesting that organizations must carefully evaluate which safety methods are practical for their specific applications and constraints.