Let's Talk
Let's Talk
logotype
logotype
Digging into the ISO/IEC 5469:2024 – Artificial intelligence: Functional safety and AI systems standard
04/28/24

Digging into the ISO/IEC 5469:2024 – Artificial intelligence: Functional safety and AI systems standard

This article offers an in-depth look at topics related to Functional Safety Autonomous Systems and Responsible AI.

For expert-level training—including certification-based programs—on these topics and more, explore our Automotive trainings and Responsible AI trainings. To learn how we support product development, compliance, and organizational safety goals with consulting support, visit our Functional Safety & Cybersecurity, Autonomous Product Development and Responsible Artificial Intelligence pages—or contact us directly.

We recently published an article on the soon-to-be released ISO/CD PAS 8800 – Road Vehicles: Safety and artificial intelligence standard. Within this article we highlighted seven standards referenced in the ISO/CD PAS 8800 standard, one being ISO/IEC 5469:2024 – Artificial intelligence: Functional safety and AI systems. The ISO/IEC 5469:2024 standard was released in January of 2024, and, unlike ISO/CD PAS 8800, it is not specific to just automotive. It is focused on the use of AI inside safety related functions, the use of non-AI safety related functions to ensure the safety of AI controlled equipment, and the use of AI systems to design and develop safety related functions, agnostic to industry. This means from a functional safety perspective, it references common standards such as the IEC 61508, ISO 26262, IEC 62061, ISO 13849 and IEC 61511 functional safety standards. That said, the terminology and interdependence between AI technologies and functional safety generally is referenced to IEC 61508. In this context it is helpful for those in automotive applying the ISO/IEC 5469:2024 standard to have a familiarity with IEC 61508, so that a relationship to the ISO 26262 standard can be better understood.
One of the main concepts in the ISO/IEC 5469:2024 standard is deriving acceptance criteria for the three-stage realization principle as defined by the ISO/IEC 22989:2022 standard. The three stages are:
  • Data acquisition (input tasks);
  • Knowledge induction from data and human knowledge (learning tasks including training);
  • Processing and generation of outputs (processing tasks including inference)
For those of you familiar with functional safety standards, you might have noticed a strong similarity between the three-stage realization principle and the definition of a system, which relates a sensor, a controller and an actuator with one another. For each of the three stages it is important to define the desirable properties, detail the methods and techniques to be used, and to establish acceptance criteria. This piece-by-piece approach is like how we address functional safety within a system.
ISO/IEC 5469:2024 details out verification and validation techniques (Clause 9), control and mitigation measures (Clause 10), and processes and methodologies (Clause 11). As an example of mitigation measures, the standard shows three possible solutions:
  • Having a back-up function designed with “non-AI” techniques;
  • Using a supervisor function with constraints and limits;
  • Having Machine Learning (ML) redundancy with output voters.

From a process perspective, ISO/IEC 5469:2024 relates the AI lifecycle with a functional safety lifecycle. The AI lifecycle is referenced to the ISO/IEC 5338 – AI system life cycle processes standard whereas the functional safety lifecycle is referenced to IEC 61508. Within the AI system lifecycle it is necessary to have an AI management system, as defined in the ISO/IEC 42001:2023 standard.

In functional safety standards like ISO 26262, all the requirements are baked into the standard itself. As we are seeing in the AI safety landscape, it is more complicated, and knowledge of multiple standards is required. At SRES, along with our partner SGS-TUV, we provide detailed training on AI safety, with the option of obtaining the AI Safety Professional (AISP) certificate. More information can be found here. Additionally, SRES provides a training and workshop related to Responsible AI (RAI) aligned to the ISO/IEC 42001:2023 standard.

Have insights or questions? Leave a comment below—we welcome thoughtful discussion from our technical community.

Interested in learning more about our services? Find all upcoming trainings here and all consulting offerings here.

AI System Lifecycle Processes – ISO/IEC 5338

AI System Lifecycle Processes – ISO/IEC 5338

04/30/24

Could ISO/SAE 21434 Have Stopped the TeslaLogger Exploit?

09/18/24
Could ISO/SAE 21434 Have Stopped the TeslaLogger Exploit?