Recording data lineage and provenance: Which practice would have prevented this gap?

Data lineage and provenance records trace each dataset's origin and transformations, answering exactly which sources and steps produced the training data.

The question

Preparing for AIGP? Take the free 5-min readiness quiz →

1. Recording data lineage and provenance, so every dataset's origin, transformations, and inclusion can be traced from source to model. ✓

   Correct: lineage and provenance records capture where data came from and how it was transformed, which is exactly what the regulators requested.
2. Retaining only the final training file, on the view that the finished dataset is what really drove the model's behavior.

   Plausible because the final file drives training, but it does not reveal origins or which sources were folded in.
3. Storing model checkpoints at each epoch, which captures how the weights evolved but not where the underlying data originated.

   Plausible as versioning, but checkpoints trace the model, not the provenance of its training data.
4. Keeping the data dictionary that defines each individual field, which documents the schema's meaning but not the source of any given record.

   Plausible since dictionaries aid understanding, but field definitions do not establish where records came from.

How to remember it

Data lineage and provenance records trace each dataset's origin and transformations, answering exactly which sources and steps produced the training data.

More Understanding How to Govern AI Development questions

• Confirm the lawful rights to reuse the data for this new: Which action best satisfies data-governance →
• Bias testing across protected groups together: Beyond validation of predictive performance, which combination →
• Treat the disparity as a material risk and remediate it: Which response best fits risk management during the →
• All 426 Understanding How to Govern AI Development questions →