The data vocabulary covers class, cluster, feature, distance and unlabelled data, which are the words a data statement needs to be readable.
What the base framework establishes
A class is a category a human defines, with shared attributes and usually a name. A cluster is a category the method induces, which need not have a name. A feature is a property of an object or event that can be measured, and the notes explain that features are the machine-readable description the model actually sees, so their design carries the information the model can use. Distance is the measured proximity of two points, ordinarily Euclidean in this context. Unlabelled describes a sample with no target variable. Together the group is the minimum vocabulary for describing what went into a model and what the model was asked to predict.
Why it matters to an AI management system
Data statements fail for a simple reason: they describe the source system rather than the dataset the model saw. These five terms force the description down to the rows, columns and labels that were actually used. That is also the level at which bias, leakage and provenance questions can be answered.
The public draft summary names no change to 3.3 Data. The final text could still touch it.
Evidence a reviewer samples
| Category | Artefact |
|---|---|
| Data statement | The dataset name, the classes or target, the features and the labelling method. |
| Feature register | Each feature, its source, its type and the reason it is used. |
| Label record | Who or what produced the labels, the instructions and the quality check. |
One question to prepare for. Does your data statement describe the dataset the model saw, or the system it came from?
Common gaps
- The data statement names a source table but not the extract or the date range.
- A feature is described by its column name only, with no definition of what it measures.
- Labels are described as ground truth without saying who assigned them.
Prepare before publication
Create a data statement template with fields for dataset, extract date, classes or target, features, labels and de-identification. Apply it to the highest-risk system first.
Dates that matter
No separate date attaches to 3.3; the record dates are the stage date, the planning estimate and the CEN expectation stated above.
Cross-references
6.4 Data · 8.2 Data acquisition · 8.3 Data preparation
Source
Source for 3.3: ISO/IEC 23053:2022, clause 3.3. The statement is read from the public base sample at 3.3 and is written in our words. The two official records are linked on this page so any line can be checked.
Not legal advice; does not certify. A reading of the framework and the evidence a reviewer commonly samples, not a substitute for the standard or your auditor's rules.