ISO/IEC 23053:2022, the system framework
23053 gives a way to describe an AI system that uses machine learning, naming its parts and what each part does in the wider ecosystem.
The requirement in plain English
23053 gives a way to describe an AI system that uses machine learning, naming its parts and what each part does in the wider ecosystem.
What the public record says
The public view gives the scope and terms. The scope says it covers organisations of every type and size that build or use AI systems.
Why it matters to you
A robustness claim needs a system boundary. 23053 supplies the component model: data, model, training, inference, and the actors around them. If your evidence pack describes a model but not the system it sits in, the stability definition at 3.1 cannot be applied, because the subject is the AI system. The framework also helps a reviewer see which component a perturbation reaches and which component a metric measures.
Evidence a reviewer asks for
- The system description using the components named in 23053.
- The boundary diagram, with data, model and deployment marked.
- A note for each perturbation saying which component it targets.
Checks to run this month
- Does the evidence describe a system or only a model?
- Is the boundary drawn the same way in the risk register and the test plan?
- Which component does each metric measure?
Dates
23053 published in 2022. FDIS registered 2026-08-21.
Cross-references
- Clause 2, Normative references
- Clause 3.1, Stability
- High-dimensional input and why exhaustive testing fails
What this page does not state
The public view carries the scope and terms, not the full framework chapters.