Quality of data for AI systems

Define and document data quality requirements and check that data used in building and running the AI system satisfy them. In practice the record is data quality requirements for training, validation, test and production data. For quality of data for ai systems the evidence starts with the provenance record.

What it asks

Define and document data quality requirements and check that data used in building and running the AI system satisfy them.

Where generative AI changes the answer

Quality checks have to cover duplication, staleness and the presence of personal data, because each one changes how the model behaves.

Who it binds

The data owner for the AI system and the team that builds or buys it. For A.7.4 the named role is the owner of data quality requirements for training, validation, test and production data.

Evidence an auditor normally samples

• Data quality requirements for training, validation, test and production data
• Data quality measurement results
• Bias analysis and adjustments made for performance and fairness

The test you can run

Compare data quality requirements for training, validation, test and production data with the use as it runs today; the gap between the two is the finding to record.

Common failures

• Quality requirements defined for training data only
• Bias impact on fairness not assessed

Before the draft publishes

For A.7.4, trace one output back to the data behind it. If you cannot, provenance is the first control to build. The first artefact to produce is data quality requirements for training, validation, test and production data.

Where this sits against the draft

The draft's published scope names controls for the risks it identifies, and A.7.4 (quality of data for ai systems) is the published ISO/IEC 42001 control closest to that work. The draft does not replace it, and an audit still tests it against the evidence listed here.