Safety and artificial intelligence
AI is changing how safety-related products are built, and it challenges the assumptions behind ISO 26262 and IEC 61508.
Where AI meets functional safety
AI as a development tool
Generating requirements, code and tests. The output must be traceable and independently reviewed.
AI inside the product
A learned component in a safety-related function. Its behaviour depends on data and training, not only on code.
What makes AI hard to assure
Traceability & explainability
AI-generated work products are hard to trace back to requirements, and a neural network cannot be inspected like source code.
Non-deterministic behaviour
The same input may give different outputs, which complicates verification, reproducibility and the argument against systematic faults.
Data as a source of faults
For learned components, errors come from incomplete, biased or wrongly labelled training data, not only from code.
Limits of verification
Classic coverage metrics do not apply to ML models, and testing alone cannot show the absence of hazardous behaviour.
Tool confidence
AI tools can produce plausible but wrong results, so they need qualification and independent review of their output.
Change management
Retraining or updating a model changes its behaviour and needs an updated safety argument.
The first automotive AI safety standard
ISO/PAS 8800, “Road vehicles – Safety and artificial intelligence”, 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 for AI elements
- An assurance argument, including monitoring after release
How we can help
We help you define the safety concept and the assurance argument for AI-based functions and AI-assisted development processes, in line with ISO 26262 and ISO/PAS 8800.
Ready to talk about your safety case?
Email safetyba@safetyba.com or call +421 944 237 841.