Data verification
Checking that data is not just correctly formatted but actually correct, confirming the values match what is true in the real world.
The Simple Version
Checking that data is not just correctly formatted but actually correct, confirming the values match what is true in the real world.
Detailed Explanation
Data verification addresses the accuracy dimension of data quality by comparing data values against an authoritative reference source (e.g. a government register, a physical measurement, an expert assessment). In AI training data, verification commonly involves ground-truth annotation comparison (do annotator labels match expert reference labels?), duplicate-source cross-checking, and statistical sampling against known distributions. Verification is more resource-intensive than validation but essential for high-stakes AI applications. The EU AI Act's data governance requirements implicitly call for verification where inaccuracies in training data could lead to safety or rights harms.
Key Characteristics
- Assesses accuracy against external truth, not just conformance to rules
- More resource-intensive than validation; often involves sampling and expert review
- Critical for annotation quality in supervised learning datasets
- Supports the accuracy dimension of ISO/IEC 5259 data quality frameworks
Why It Matters
For high-risk AI applications in healthcare or law enforcement, verification of training labels by domain experts is an essential quality control, and a defensible element of Annex IV documentation.
Real-World Analogy
Like an auditor checking that financial statements accurately reflect actual transactions, not just that they are formatted according to accounting standards, verification goes beyond format to confirm underlying truth.
Common Misconceptions
- Verification is a one-time activity completed during data collection, data accuracy can degrade over time, requiring periodic re-verification for time-sensitive domains.
- Automated tools can fully substitute for expert verification, automation can flag anomalies, but domain-expert judgment is often irreplaceable for ground-truth accuracy assessment.