5.4 Clause
Machine learning
The family covers methods whose behaviour is fitted from data: statistical models, classical supervised and unsupervised learning, neural networks and generative models. The descriptor set is meant to say more than "ML", including the learning approach and the data dependency.
What this settles
Machine learning is where most vendor claims and most inventory entries live. The risk is a label so broad that it says nothing about failure modes, data provenance or the evidence a reviewer needs.
Who it binds
The organisation, as a method choice. Data protection, non-discrimination and sector rules may attach to the training and use of the model, but those duties come from those instruments, not from this draft.
What to do in your practice
Record the learning approach, the training data source, the model version and the evaluation set for each model. Where a model is fine-tuned or prompted rather than trained, say so, because the change and evidence story is different.
Evidence an auditor or regulator asks for
| A model card with approach, data source, version and evaluation set. | Gives the reviewer the facts behind the label. |
|---|---|
| A dated evaluation result for each capability claimed. | Connects the claim to a measurement. |
| A change record for a retrain or a model swap. | Shows the description follows the model. |
Which model version produced the output we are discussing, and which evaluation set supports the capability claim?
Common gaps this page catches
- The inventory says "machine learning" and nothing else.
- A model is swapped in production without a change to the description.
- The evaluation set is unknown, so the capability claim cannot be checked.
What to prepare before publication
Adopt a minimum model card and apply it to the three highest-impact models. Include the data source and the evaluation set, not only the algorithm name.
How to test it here
Take one live system rather than a template, find the field or sentence where Machine learning is used, and put this page beside it. Record the answer, the artefact and the date. A glossary entry on its own is an intention; a dated record beside a live artefact is evidence.
What this page does not claim
The page does not reproduce the draft, does not know the descriptor wording that the public preview withholds, and does not certify the system or the organisation. It states the unit at the claim level "methods family" and stops where the public record stops.
Common questions
- What does Machine learning mean in ISO/IEC DIS 42102?
- The family covers methods whose behaviour is fitted from data: statistical models, classical supervised and unsupervised learning, neural networks and generative models. The descriptor set is meant to say more than "ML", including the learning approach and the data dependency.
- What evidence should we keep for 5.4?
- A model card with approach, data source, version and evaluation set. A dated evaluation result for each capability claimed.
- Is Machine learning a requirement?
- No. The draft is a descriptive framework. It sets no management system requirement and no conformity test. It becomes binding only when a contract, a regulator or a management system adopts it.
Related units
Last updated: · Markdown version · Check your system