Challenges for functional safety
- Traceability and explainability: AI-generated work products are hard to trace back to requirements and design intent, and the reasoning of a neural network cannot be inspected like source code.
- Non-deterministic behaviour: the same input may lead to different outputs, which complicates verification, reproducibility and the argument of freedom from systematic faults.
- Data as a source of faults: for learned components, errors originate in training data that is incomplete, biased or wrongly labelled, not only in code.
- Limits of verification: classic coverage metrics do not apply to machine-learning models, and testing alone cannot show the absence of hazardous behaviour in all situations.
- Tool confidence: AI tools can produce plausible but wrong results, so they must be assessed and their output independently reviewed.
- Change management: retraining or updating a model changes its behaviour and requires an updated safety argument.
ISO/PAS 8800
ISO/PAS 8800:2024, “Road vehicles – Safety and artificial intelligence”, is the first automotive publication dedicated to AI in safety-related systems. It complements ISO 26262 and ISO 21448 (SOTIF) with a safety lifecycle for AI-based elements. It covers AI-specific safety requirements, data and dataset quality, model development and verification, safety analyses, and an assurance argument showing that the AI system is acceptably safe, including monitoring after release.How we can help
We support you in defining the safety concept and the assurance argument for your AI-based functions or AI-assisted development processes, in line with ISO 26262 and ISO/PAS 8800. Contact us to discuss your project.Ready to discuss your project? Tell us about your functional safety needs and we will get back to you within one business day.