Overview:Many companies' first reaction to ISO/PAS 8800 is: "Aren't we already doing ISO 26262 and SOTIF — why do we need another one?" This article puts the four standards that are often confused into a single table:what each one governs, where the boundaries lie, and how they work together in real projects.
1. How the Four Standards Divide the Work
| Standard | In one line | What risk it governs |
|---|---|---|
| ISO 26262 Functional safety | Functional safety of electrical/electronic systems | "Deterministic faults" such as random hardware failures and systematic software failures |
| ISO 21448 Safety of the intended functionality (SOTIF) | Scenarios in which hazards can arise even without a fault | Limits of perception capability, insufficient scenario coverage, environmental conditions beyond the design domain |
| ISO/SAE 21434 Cybersecurity | Automotive cybersecurity engineering | External attacks, exploitation of vulnerabilities, malicious tampering, data theft |
| ISO/PAS 8800 AI safety | Safety engineering for in-vehicle AI / machine learning systems | Data deficiencies, insufficient model robustness, black-box inexplicability, performance drift after deployment |
The key to understanding this table is to remember two dividing lines:"safety" and "security" are two different things— the former is about whether people and vehicles can be harmed, the latter is about whether the system can be breached from outside.ISO/PAS 8800 sits on the first line, filling the gap left by AI.
2. Why AI Crosses Both Lines at Once
Once AI systems enter the vehicle, risk comes from both sides at once:
- On the safety side:the training data does not cover a particular real-world scenario, the model misjudges it, and the vehicle takes a dangerous action — not because hardware failed or because it was attacked, but because of deficiencies in the AI's own capability and data;
- On the security side:model weights are tampered with, training data is poisoned, sensors are disturbed by adversarial samples — the AI's behavior is altered maliciously from outside.
In practice, then, the three frameworks must work together:ISO 26262 controls failures, SOTIF controls capability limits, ISO 21434 controls external attacks, and ISO/PAS 8800 controls the AI's own data, models and operational evolution.They are not substitutes for one another — they stack layer upon layer.
3. Good News: One Audit Can Serve Multiple Standards
In the case of the world's first ISO/PAS 8800 certification, the issuing body TÜV Rheinland stated explicitly that an ISO/PAS 8800 auditcan be combined with audits of ISO 26262 (functional safety) and ISO 21448 (SOTIF),allowing a company to cover several standards with a single system audit and reduce duplicated investment. The audit focus includes:
- whether AI-safety-relatedmanagement processesare established and running under control;
- whether the evidence coversdatasets, models, training processes, testing activities and operational monitoring;
- whether the safety claims and the test evidence arecomplete and traceable.
For small and medium-sized companies this is a very practical insight:you do not have to wait to "do 8800 separately later" — add the AI-specific requirements into your existing functional safety system and go through audit together.
4. Recommended Implementation Path: Three Steps
- Step 1 · Current-state review and gap analysis:take your existing functional safety, SOTIF and cybersecurity process documents and check them clause by clause against AI characteristics — which clauses apply directly, which need AI-specific supplements, and which are entirely blank. This step usually produces conclusions within one to two weeks and is the lowest-cost, most accurate way to set direction.
- Step 2 · Fill the three shortest planks:① data governance (dataset suitability, scenario coverage analysis, records of collection and labeling); ② model validation (normal and edge-case scenarios, robustness, behavior in unknown scenarios); ③ change and monitoring (version management, post-launch performance monitoring, trigger conditions and handling processes).
- Step 3 · Build the AI safety file and prepare for audit:thread the outputs of the five stages — requirements, design, data, testing and monitoring — into a complete evidence chain forming an auditable safety argument, and choose a route that audits jointly with your existing ISO 26262 / ISO 21448 system.
5. Three Common Misconceptions
- Misconception one: "We don't do L4, so it doesn't apply to us."As long as a safety-related function on the vehicle uses machine learning (for example vision-based perception or model-based prediction), it falls within the standard's scope — there is no necessary link to the level of automation.
- Misconception two: "High accuracy equals safety."Model accuracy is a performance metric, not a safety metric. What the standard requires is traceable safety evidence — where the data came from, what scenarios it covers, how abnormal situations are handled, and how the system is monitored after launch.
- Misconception three: "This is the R&D department's own business."The standard requires management processes and an evidence chain spanning data, development, testing, quality and after-sales; without organizational backing endorsed by management, the evidence chain will inevitably break at some departmental boundary.
6. Final Thoughts
From UN R155/R156 to GB 44495/44496, and from ISO 26262, ISO 21448 and ISO/SAE 21434 to today's ISO/PAS 8800, the automotive industry's safety compliance system is expanding from "mechanical and electronic" to "software, networks and artificial intelligence". And the pattern becomes clearer each time:first the standard appears, then certification cases follow, and finally it becomes a market-access condition.
ISO/PAS 8800 has completed the first step and is now in the second. For companies, doing the gap analysis now costs far less than being forced to produce materials later by customers or certification bodies.
Related Articles
- ISO/PAS 8800 Is Published: Automotive AI Safety Now Has an International Standard, and the World's First Certification Has Been Issued
- What Is ISO/SAE 21434: A Full Overview of the Automotive Cybersecurity Engineering Standard
- SOTIF, ODD and MRM: Core Concepts of Automated Driving Safety of the Intended Functionality
- The "Triple Triangle" of Automated Driving Safety: Functional Safety, Safety of the Intended Functionality and Cybersecurity
Need AI Safety and Compliance Support?
If your organization is working on intelligent driving AI system safety, ISO/PAS 8800, or ISO 26262 / ISO 21448 implementation, contact us for a free compliance diagnostic. QinGuard specializes in automotive safety and compliance consulting, providing end-to-end services from gap analysis and process development to safety argumentation and certification support.