A review on AI Safety in highly automated driving

Wäschle, Moritz; Thaler, Florian; Berres, Axel; Pölzlbauer, Florian; Albers, Albert · 2022 · Crossref

DOI: 10.3389/frai.2022.952773

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This systematic literature review addresses the critical challenge of ensuring safety in Artificial Intelligence (AI) systems, specifically within the context of highly automated driving. The motivation stems from the increasing deployment of data-driven AI solutions, particularly machine learning (ML) and neural networks, in safety-critical applications. While AI offers advantages in handling non-linear system behaviors and unknown environments, it lacks the safety guarantees inherent in classical engineering methods. Current safety standards, such as ISO 26262, are ill-equipped to address the unique failure modes, black-box nature, and data-dependency of ML systems. Consequently, the paper aims to provide a comprehensive overview of AI Safety research, focusing on validation, verification, and testing methods to mitigate risks associated with unintended and harmful behavior. The authors conducted a systematic literature review following established methodologies to identify and categorize relevant research. The review distinguishes between two primary classes of approaches: "classical approaches," which involve adapting existing standards and concepts from outside the ML/AI domain, and "new approaches," which are methods specifically tailored to the unique characteristics of AI systems. The study examines the intersection of highly automated driving and AI safety, analyzing how traditional safety engineering principles apply to AI and where novel solutions are required. The review also considers ongoing standardization efforts, such as those by ISO/IEC JTC 1/SC 42 and UL4600, which are attempting to define frameworks for AI safety, trustworthiness, and certification. Key findings indicate that while classical safety methods provide a foundational framework, they are insufficient for fully addressing the complexities of AI systems. The review highlights that ML systems introduce new hazards through complex human-AI interactions, distinct failure modes, and reliance on incomplete training data. The "new approaches" identified include methods for robust development, adversarial robustness, value alignment, and explainability. The paper notes that significant gaps remain in the scientific basis for AI safety, particularly regarding formal verification of neural networks, architectures for robust deep learning, and tools for generating comprehensible explanations. Furthermore, the review identifies seven main challenges in beneficial AI: fairness, transparency, misuse, security, policy, ethics, and control/alignment. The significance of this work lies in its structured categorization of AI safety research, providing a roadmap for engineers and researchers developing safe autonomous driving systems. By distinguishing between classical and novel approaches, the paper clarifies where existing standards can be leveraged and where new methodologies are essential. The authors conclude that ensuring AI safety requires a multi-faceted approach involving rigorous validation, verification, and testing, alongside ongoing efforts to standardize safety practices for AI. This review underscores the urgent need for further research to bridge the gap between current AI capabilities and the stringent safety requirements of highly automated driving, ultimately aiming to reduce accident risks and build trust in AI-controlled systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• driverless ads
• generative ai voice assistants
• automation surprise
• situational awareness

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Applied Guidance: standards test procedures
• Theoretical Contribution: computational model