ISO/PAS 8800

ISO/PAS 8800 provides a structured framework for integrating AI into road vehicle safety systems, addressing risks such as unpredictability, data reliance, and complex decision-making throughout the AI lifecycle. It aims to prevent undesired safety-related behavior at the vehicle level caused by inadequate performance, systematic failures, and random hardware faults in AI elements.

The "PAS" in its designation indicates a "Publicly Available Specification". It is a preliminary standard that is still under development, but provides valuable guidance.

The importance of ISO/PAS 8800 for automotive AI development stems from the inherent challenge of guaranteeing safety in systems that learn and adapt. Traditional functional safety standards primarily address known failure modes and predictable system behavior.

However, AI introduces new types of risks, such as those arising from:

• Biased training data

• Lack of robustness to novel inputs

• Uncertainties in predictions

ISO/PAS 8800 provides the necessary framework to manage these specific AI-related safety concerns, helping manufacturers develop a compelling safety argument for AI-driven vehicle functions. By offering clear guidelines, it supports the safe and responsible deployment of AI in critical automotive applications, ultimately enhancing public trust and accelerating the adoption of advanced vehicle technologies.

ISO/PAS 8800's key principles revolve around a holistic, lifecycle-based approach to AI safety in road vehicles. It recognizes that AI systems require specific considerations beyond traditional safety engineering.

The objectives include defining what constitutes an "AI element" in a vehicle context, establishing methods for deriving safety requirements based on AI technology properties, and addressing data management, which is crucial for AI performance. A central principle is the creation of a strong "assurance argumentation" for the safety of AI elements within vehicle systems, meaning a documented case that provides confidence that the AI system is acceptably safe. This involves an iterative approach to risk reduction, considering the unique challenges posed by AI's learning and adaptive capabilities.

ISO/PAS 8800 is designed to complement existing automotive functional safety standards, rather than replace them. It integrates seamlessly with the established safety framework, most notably the standards for functional safety (ISO 26262) and for the safety of the intended functionality (ISO 21448).

ISO/PAS 8800 extends these standards by providing specific requirements for systems incorporating AI. It bridges the gap by addressing how AI-specific properties, such as:

• Bias

• Prediction accuracy

• Robustness

• Generalization

• Safety Impact

For instance, it takes requirements derived from ISO 26262 and ISO 21448 and expands them with AI‑specific considerations. These standards are primarily focused on the automotive domain. However, the overarching principles of managing safety in complex, AI‑driven systems also resonate with other safety‑critical industries. These industries are governed by standards such as IEC 61508 for general industrial safety, DO‑178C for aerospace software, and EN 50128 for railway applications. All of these domains are increasingly facing challenges posed by AI integration.

ISO/PAS 8800 gives recommendations such as “A justification shall be provided that the selected AI technologies and AI methods are capable of fulfilling the AI safety requirements.” This is not as concrete as traditional software standards, which recommend specific methods like using equivalence classes and code coverage during unit testing.

However, it provides initial guidance, for example by listing properties of data such as accuracy and completeness, and by describing methods to achieve them. It does not include AI‑specific verification methods and metrics like those found in VDE‑AR 2842‑61, a German application rule that outlines structured procedures, quality criteria, and test methods for assessing the reliability and performance of AI systems.

Demonstrating confidence in the use of software tools plays a vital role in achieving compliance with ISO/PAS 8800, much like it does for ISO 26262. AI development heavily relies on sophisticated software tools for data handling, model training, validation, and deployment. These include data annotation tools, machine learning frameworks, simulation environments, and verification tools.

The principles for tool confidence, as detailed in ISO 26262 Part 8, Clause 11, are directly applicable and applied within the context of ISO/PAS 8800. This means that organizations must ensure the confidence in these tools, i.e.

• Classify their toolchains including the AI development and verification tools

• Qualify the tools classified as critical (TCL2 or TCL3)

This includes, for example, qualifying data handling tools, training platforms, and AI model verification tools, as the quality and integrity of the data and the behavior of the trained models are paramount to the safety of AI-driven automotive functions.

Validas specializes in tool and library classification and qualification for safety-critical systems, with deep expertise in ISO 26262 and cross-domain safety standards. This expertise directly translates to ISO/PAS 8800 projects, where tool confidence is a key enabler of AI safety assurance.

Validas AG supports organizations by:

• Structuring and classifying AI development and verification toolchains

• Providing reusable qualification kits for tools and libraries where applicable

• Developing tailored qualification strategies for AI-specific tools and frameworks

• Ensuring alignment with ISO 26262 Part 8 while addressing AI-specific risks

• Strengthening AI safety arguments by supplying robust, audit-ready tool confidence evidence

By bridging the gap between abstract AI safety requirements and concrete, defensible evidence, Validas enables OEMs and suppliers to integrate AI technologies with confidence, safely, efficiently, and in compliance with evolving standards such as ISO/PAS 8800.