Safety-Aware Machine Learning (ISO 8800)
Engineer AI and ML components that can stand in a safety case: the ISO/PAS 8800:2024 view layered on ISO 26262 and ISO 21448, covering requirements, data, robustness, verification, runtime monitoring and the assurance argument.
- Chapters
- 11
- Chapters
- Figures and interactive visuals
- 14
- Figures and interactive visuals
- Standards connected
- 3
- Standards connected
Quick answers
- What is ISO/PAS 8800?
- ISO/PAS 8800:2024 addresses the safety of AI and machine learning in road vehicles and builds on ISO 26262 and ISO 21448 (SOTIF). It is needed because the behavior of a learned component is set by training rather than fully stated in requirements and code, so insufficient data, unsuitable labels or unfamiliar inputs can produce unsafe outputs even when hardware and software work as implemented. ISO/PAS 8800 adds lifecycle, data, verification and assurance activities for these AI-specific concerns.
- Why is ISO 26262 alone not sufficient for machine learning?
- ISO 26262 assumes failures come from faults that can be analyzed and then prevented, detected or controlled. A learned model can produce a wrong output without any fault, simply by generalizing poorly to an input it was not trained for, and its behavior depends on the data as much as on the code. ISO 21448 covers such performance limitations at system level, and ISO/PAS 8800 adds the AI-specific engineering underneath. The guide maps which standard governs which concern.
Why it pays for itself
Write verifiable ML safety requirements
Derive AI and data requirements from safety goals that state measurable error behavior in safety-relevant categories and conditions, instead of relying on aggregate accuracy.
Make the dataset an argument
Treat data as part of the implemented behavior: show input-space coverage, labeling quality and lineage, and argue why the dataset is sufficient for the claim it supports.
Supervise the model at runtime
Design doer and checker architectures, operating-boundary checks and out-of-distribution detection with defined responses when the model leaves known territory.
What you’ll be able to do
Write verifiable ML requirements
Derive requirements from safety goals that say what must be measured and on which data.
Argue dataset sufficiency
Show input-space coverage, label quality and lineage for the data behind a model.
Verify beyond accuracy
Use safety-relevant metrics and independent test data for ML evidence.
Design runtime supervision
Place monitors and checkers around a model with defined responses.
Integrate with ISO 26262 and SOTIF
Decide which standard governs which concern and how evidence flows between them.
Chapter by chapter
- 01
The AI Safety Gap
See why the classic fault model does not fit learned components and which recurring challenges AI brings to safety.
- The classic fault model
- Recurring AI challenges
- What ISO/PAS 8800 adds
- 02
ISO/PAS 8800 in the Landscape
Place ISO/PAS 8800 among ISO 26262, ISO 21448, ISO/IEC TR 5469 and UL 4600, and see what each contributes.
- Standards ecosystem
- Complementing ISO 26262
- TR 5469 and UL 4600
- 03
The ML Safety Lifecycle
Follow a teaching workflow for ML development, its links to the encompassing lifecycles and the safety gates between phases.
- Why ML needs a lifecycle
- Teaching workflow
- Safety gates
- 04
Safety-Related ML Requirements
Derive ML requirements from safety goals and make them verifiable instead of relying on average accuracy.
- Deriving from safety goals
- Requirement dimensions
- Verifiable requirements
- 05
Data & Dataset Safety
Treat data as part of the implemented behavior, with input-space coverage, labeling quality, lineage and sufficiency arguments.
- Input-space coverage
- Labeling quality
- Data lineage
- 06
Robustness & OOD
Understand distribution shift, out-of-distribution detection, perturbation robustness and calibrated uncertainty.
- Distribution shift types
- OOD detection
- Uncertainty calibration
- 07
V&V of ML
Choose safety-relevant metrics, test by scenario, keep test data independent and know what formal methods add.
- Beyond accuracy
- Scenario-based testing
- Test independence
- 08
Runtime Monitoring
Supervise a model in operation with a doer and checker architecture, operating-boundary checks and defined risk responses.
- Doer and checker
- Operating boundaries
- Runtime OOD detection
- 09
The AI Safety Case
Structure an AI safety case around a small set of concerns and argue residual risk with a GSN-style outline.
- What an AI safety case is
- Argument structure
- Residual risk argument
- 10
Integration with 26262/21448
Map which standard applies where, allocate requirements to AI elements and feed ML evidence into SOTIF validation.
- Applicability map
- ML inside a mechanism
- SOTIF validation evidence
- 11
Anti-Patterns
Recognize the recurring mistakes in ML safety work, from accuracy-only metrics to missing field monitoring.
- Accuracy as the only metric
- Train and test leakage
- Uncontrolled model authority
Read all 11 chapters with Expert.
Start with a free account: the complete Functional Safety Concept guide and 5 practice exams a month, then upgrade when you need this guide.
Expert plan pricingNot just text: the visual toolkit
Feature-space confidence field
Why softmax confidence and distance to the data tell different stories.
ML safety lifecycle pipeline
The development workflow for a learned component.
OOD threshold explorer
An interactive trade-off between false alarms and missed unfamiliar inputs.
Doer and checker architecture
A learned component supervised by a simpler independent checker.
SOTIF scenario model with AI evidence
Where ML evidence enters the SOTIF scenario areas.
Who this guide is for
- ML engineers whose model is about to enter an automotive safety case for the first time
- Functional safety engineers assigned an AI-based component and unsure where ISO 26262 stops
- Safety assessors and reviewers who must judge AI safety evidence at review gates
- Architects designing runtime supervision around a perception or planning network
Frequently Asked Questions
Common questions about Safety-Aware Machine Learning (ISO 8800)
Start the course today
A free account unlocks one full concept guide, 3 work product templates, 1 guided process, the Markov and RBD simulators, and 5 practice exams per month. The Pro and Expert plans unlock more of the 93-guide library. No credit card required.
Continue with
Where this topic leads next.
- Beyond the standardAV Safety: ISO 26262 + SotIF + ML in PracticePractical handbook bridging ISO 26262, ISO 21448 (SotIF), ISO/PAS 8800, and UL 4600 for autonomous vehicles with machine learning. Covers ODD, fail-operational architectures, scenario V&V, and joint safety cases.
- Beyond the standardTrucks & Buses AdaptationHow ISO 26262:2018 extended functional safety to trucks, buses, trailers and semi-trailers. Covers the Part 3 rules (Clause 6.4) for managing variances, multi-stage manufacturing interfaces, heavy-vehicle hazards and worst-case ASIL determination across the affected parts of the standard.
- FoundationsFaults and FailuresLearn the complete fault taxonomy: single-point, residual, latent, and safe faults plus transient vs. permanent classifications. Covers FTTI analysis, dependent failure analysis, FMEDA calculations, and the bathtub curve.