etc.Six AI safety standards: which one covers what?
AI FUNCTIONAL SAFETY GUIDE · 2026
In short: For road vehicles, add ISO/PAS 8800, which covers AI elements, on top of ISO 26262 (faults) and ISO 21448 SOTIF (performance limitations). For industrial machinery and robots, consult ISO/IEC TR 5469 on top of the IEC 61508 family and prepare for its requirements-level successor, ISO/IEC 22440. If you sell in the EU, check separately whether your product counts as high-risk AI under the EU AI Act.
Six AI safety standards: which one covers what?
In one line: ISO 26262 covers faults, ISO 21448 covers performance limitations without faults, and ISO/PAS 8800 covers the AI elements inside them. In industrial sectors, ISO/IEC TR 5469 and ISO/IEC 22440 play the same role. The EU AI Act is not a standard but a regulation that sits on top of both.
| Name | Type | Scope | Risk addressed | Key concepts |
|---|---|---|---|---|
| ISO 26262 | Int'l Standard | Road vehicle E/E systems | Malfunctions caused by hardware and software faults | HARA, ASIL, safety mechanisms, FMEA · FTA · FMEDA |
| ISO 21448 (SOTIF) | Int'l Standard | Vehicle systems with perception and decision functions | Functional insufficiencies, performance limitations and misuse, even without faults | Triggering conditions, reducing unknown hazardous scenarios, acceptance criteria |
| ISO/PAS 8800 | PAS | AI elements in vehicle safety-related systems | Limits of training data and model performance; impact of AI output errors on safety | AI safety requirements, data management, AI verification, in-operation monitoring, safety argument |
| ISO/IEC TR 5469 | Technical Report | Cross-sector (mainly industrial machinery, robots, process) | Risks of using AI technology in safety functions | AI technology classes and usage levels, wrapping AI in non-AI safety layers |
| ISO/IEC 22440 | In development | Cross-sector | Same area as TR 5469 | Turning TR 5469 recommendations into requirements (check ISO for current status) |
| EU AI Act | EU Regulation | AI systems placed on the market or used in the EU | AI risks to fundamental rights and safety | Risk-based classification; risk management, data, record-keeping and human oversight duties for high-risk AI |
We do not cite clause numbers from the standards. Base your applicability decisions and audit preparation on the original texts.
Which standards apply to your product? Five questions to decide
Answer the five questions below in order and you will know which set of standards to review first. If an answer is unclear, treat it as "yes" and keep it in scope.
- Does a safety-related function of the product contain an AI element such as machine learning or a neural network?
Yes → go to Q2 · No → review sector functional safety first (ISO 26262 or the IEC 61508 family), plus SOTIF if there are perception or decision functions. - Is it a road vehicle (passenger, commercial, two-wheeler) or a component of one?
Yes → ISO 26262 + ISO 21448 + ISO/PAS 8800 · No → go to Q3 - Is it subject to IEC 61508-family functional safety, like industrial machinery, collaborative robots, logistics robots or process equipment?
Yes → IEC 61508 family + ISO/IEC TR 5469, and track ISO/IEC 22440 - Does the AI perform the safety function itself, or does it assist alongside a non-AI safety function?
The more directly it performs the function, the higher the bar. Place it using the TR 5469 usage-level concept, and first consider an architecture that wraps the AI in a non-AI monitoring and limiting layer. - Will it be placed on the EU market or used in the EU?
Yes → check whether it is high-risk AI under the EU AI Act. AI used as a safety component under EU product legislation, such as machinery or medical devices, can be classified as high-risk. For vehicles, confirm separately how and when the obligations flow through type-approval legislation.
AI safety readiness self-assessment (10 items)
Once you know the standards, the next question is "what do we already have?" The ten items below restate activities that ISO/PAS 8800 and TR 5469 both call for, in terms of practical work products. Check each one that applies.
0–3 · Starting
Begin by identifying AI elements and defining the ODD. Without these two, the scope of every later analysis is undefined.4–7 · Strengthening
Hazard analysis and requirements often exist while data, operation and traceability are empty. Connecting the missing items into a trace chain comes first.8–10 · Audit-ready
You have the evidence for a safety argument. Compare against the original texts for missing clauses and prepare for an external assessment.Where does VisualPro help?
VisualPro, VWAY's safety analysis tool, lets you carry out the hazard analysis, requirements and traceability items in the checklist above on a single model. It connects STPA, HARA, FMEA, FTA, FMEDA and TARA through the same data and works with AI assistants over MCP to draft analyses, while engineers make the final call.
AI output error analysis — STPA
Model the AI component as a controller and derive missed detection, false detection, latency and tracking loss as unsafe control actions (UCAs). Hazard paths that arise without faults, as SOTIF and ISO/PAS 8800 describe them, are organized as loss scenarios.
Fault analysis — HARA · FMEA · FTA · FMEDA
Run ISO 26262 activities in the same project. Safety goals and requirements link to STPA results, so the two perspectives do not drift apart.
Traceability and AI integration
Keep the trace chain from hazards to verification and generate assessment reports. Through the MCP server, AI assistants such as Claude can read the model and propose analysis drafts.
Case study white papers
- Applying STPA to Support ISO/PAS 8800 — pedestrian-recognition AEB with a camera neural network (Korean and English)
- Applying STPA to Support ISO/IEC TR 5469 — logistics AMR with a safety LiDAR and a 3D-camera neural network (Korean)
Frequently asked questions
- What is ISO/PAS 8800?
- ISO/PAS 8800 (Road vehicles — Safety and artificial intelligence) is a Publicly Available Specification for AI elements in the safety-related systems of road vehicles. Building on ISO 26262 functional safety and ISO 21448 SOTIF, it sets out how to handle safety requirements, data, verification, in-operation monitoring and the safety argument for AI elements.
- How does ISO/PAS 8800 differ from ISO 21448 (SOTIF)?
- SOTIF covers the whole vehicle function and the risks that arise from functional insufficiencies, performance limitations or reasonably foreseeable misuse, even without faults. ISO/PAS 8800 focuses on AI elements such as machine learning and spells out how to manage training data, model performance and the safety impact of AI output errors. The two are applied together, not one instead of the other.
- Does ISO/IEC TR 5469 apply to automobiles?
- ISO/IEC TR 5469 is a cross-sector Technical Report that guides how to use and assess AI technology from a functional safety perspective. It is advisory rather than normative and is mainly consulted in industrial machinery, robotics and process sectors that use the IEC 61508 family. For automobiles, ISO 26262, ISO 21448 and ISO/PAS 8800 take precedence.
- How is ISO/IEC 22440 different from TR 5469?
- ISO/IEC 22440 is being developed as the successor that turns the recommendations of TR 5469 into requirements usable for certification and assessment. Check the ISO website for its current stage and schedule.
- Does the EU AI Act apply directly to cars and industrial robots?
- The EU AI Act is an EU regulation that governs AI by risk level. AI used as a safety component of products covered by existing EU product legislation, such as machinery or medical devices, can be classified as high-risk. Vehicles are covered through type-approval legislation, so the way and timing of application must be confirmed separately, and the application dates for high-risk obligations may be adjusted by subsequent EU measures.
- Why use STPA for AI safety analysis?
- AI-related accidents more often come from wrong perception or decisions and from interactions between controls than from component faults. STPA looks for unsafe control actions and the loss scenarios behind them regardless of whether anything has failed, so it can systematically analyze how AI output errors such as missed detection, false detection and latency lead to system hazards.
Let us review your self-assessment with you
We will show you, for your own product, which analyses and work products can fill the items you could not check. You can also get a free VisualPro trial and our case study white papers.
Contact: sales@vwaycorp.com · VisualPro trial: www.vwaycorp.com/visualpro-trial
This material is an overview of each standard for reference only. It does not replace the original texts of the standards or the interpretation of certification bodies.