Contents
StandardISO

ISO/PAS 8800

Road vehicles — Safety and artificial intelligence

ISO/PAS 8800 explains how engineers build evidence that AI used in road vehicles is acceptably safe throughout development and operation.

What it is

ISO/PAS 8800 is an ISO Publicly Available Specification for assuring the safety of automotive artificial intelligence, particularly machine learning. It addresses how AI behavior can contribute to unsafe vehicle behavior and how developers can support a convincing safety argument. “Publicly Available Specification” names the document type; the full text is paid.

Its scope includes inadequate AI outputs, development errors and hardware faults affecting AI elements. It does not prescribe a particular neural-network architecture or make a model safe simply because it achieves a high average accuracy. The official scope describes its focus on vehicle safety and machine learning. ISO scope and publication record.

Who it applies to

The document concerns safety-related E/E systems using AI in series-production road vehicles, excluding mopeds. Relevant teams include vehicle manufacturers, system suppliers and developers supplying AI elements and their supporting evidence.

Interactions with external AI can matter to the vehicle safety argument, although development of those external elements is outside its scope. AI used only as a development tool is a separate consideration from AI executing a vehicle function.

For project planning, identify which AI outputs influence safety and who owns each assumption, requirement and evidence package. Applicability follows the function and its safety relevance, rather than whether the software runs on a dedicated ECU or a central computer.

What it covers

The central engineering concern is assurance across data, models, training, integration and operation. A useful way to organize the work is:

Concern Practical engineering question
Data Does the evidence represent relevant conditions, including difficult cases?
Model Which limitations could produce unsafe outputs?
Integration How do preprocessing, runtime execution and downstream decisions affect behavior?
Operation How will teams detect problems and assess changes after deployment?

These questions connect AI performance to vehicle consequences. A missed object, for example, matters differently depending on vehicle speed, available supervision and the response of the braking function. Testing therefore needs context, documented assumptions and acceptance criteria, alongside aggregate metrics.

How it fits with other documents

ISO 26262 provides the functional-safety framework for automotive E/E systems. ISO 21448 addresses safety of the intended functionality, including hazards arising from functional insufficiencies. ISO/PAS 8800 adds AI-focused assurance considerations that teams should integrate with those existing safety activities.

For an AI perception function, engineers should connect model limitations and validation evidence to the system-level hazard analysis and safety argument. AI assurance also needs coordination with cybersecurity and software-update processes when attacks or deployed changes can affect safety. One document does not replace the others.

In practice

Consider a camera-based pedestrian detector running on an HPC accelerator. Start with its intended use, operating conditions and the consequences of missed or incorrect detections. Then connect those concerns to data selection, labeling checks, scenario testing and integration evidence.

Evaluate the deployed combination of model, preprocessing, inference runtime and hardware. Changes to quantization or accelerator execution can justify renewed assessment even when the model’s intended function stays the same.

Maintain traceability between released configurations, known limitations and supporting safety evidence. When field observations reveal an unexpected failure pattern, use that evidence to reassess assumptions and decide what validation is needed before releasing a correction.

What it means for your discipline

  • Functional safetyCore

    Own the AI safety argument and connect AI limitations to vehicle hazards, safety requirements and validation evidence.

  • HPC & application softwareCore

    Develop and verify data pipelines, trained models and inference software with traceable evidence. Assess how model or data changes affect safety.

  • E/E architecturePartial

    Allocate responsibilities between AI elements, supervision and fallback functions. Define operating assumptions and interfaces that make the safety argument credible.

  • Hardware & siliconPartial

    Provide evidence about the compute platform supporting AI execution, including accelerator faults, numerical behavior and resource constraints.

  • Software update & OTAPartial

    Include AI safety evidence in change-impact assessments for model, preprocessing and runtime updates. Keep deployed versions connected to their validated configurations.

No direct stake: Cybersecurity · Base software · Networks & communication · Diagnostics & lifecycle