How do you audit data quality for an AI system?

To audit data quality for an AI system, check where the data came from, whether labels are right, whether it covers the cases you will meet in use, whether test data leaked into training, whether it contains data it should not, and whether anyone could have tampered with it. Write the answers in a datasheet that travels with the dataset.

Why does data quality decide AI outcomes?

A model learns patterns from its examples. If the examples are wrong, old, narrow or biased, the model reproduces that with the same fluent confidence as when it is right. NIST's AI Risk Management Framework is voluntary guidance meant to help organizations build trustworthiness into the design, development, use and evaluation of AI systems, and data is one of the places where that trustworthiness is won or lost.

A 2018 paper, Datasheets for Datasets, made the point from the machine learning side. Its authors observed that the community had no standardized process for documenting datasets, which can lead to severe consequences in high stakes domains, and proposed that every dataset carry a datasheet covering its motivation, composition, collection process and recommended uses, in analogy with the datasheets that accompany electronic components.

What does an audit check?

Provenance. Where did each source come from, under what terms, and who can vouch for it? Data of unknown origin is a risk to quality, legal standing and security.

Label quality. For supervised data, sample the labels and have two people check them independently. Disagreement rates tell you how reliable the ground truth is. Labels made in a hurry teach the model to be hasty.

Coverage and balance. Does the data reflect the cases and groups the system will face? A fraud model trained on last year's patterns may miss this year's. Compare the data's distribution to the deployment setting, and look for groups that are missing or thin.

Duplicates and leakage. Duplicate records inflate scores. Worse, if examples from the test set appear in training data, the evaluation measures memory, not skill. Check for exact and near duplicate overlap between splits.

Freshness and drift. Record dates. Plan for how the model will be re-evaluated when the world changes.

Sensitive content. Search for personal data, credentials and confidential material that should not be in training or reference data. Decide the legal basis and retention for any personal data that stays.

Integrity and poisoning. NIST's adversarial machine learning report, AI 100-2, provides a taxonomy of attacks across the life cycle of AI systems, and poisoning is one of its listed attack types. Ask who can write to your training and reference data, how changes are logged, and whether you can roll back. For systems that retrieve documents at answer time, the same questions apply to the document store.

A constructed example

A support team plans an assistant trained on past tickets. The audit finds that 30 percent of tickets are duplicates created by an email gateway; that resolution labels were set by a script that marked any ticket closed within an hour as "resolved", including spam; that 12 percent of the text includes customer phone numbers and a few passwords pasted into messages; and that anyone in support can edit the shared knowledge base the assistant also reads. The team removes duplicates, relabels a sample by hand, strips sensitive strings, restricts edit rights with an approval step, and writes the findings into a datasheet before training.

How to check or apply this

1. Create a datasheet for each dataset: purpose, sources, collection method, time range, known gaps, intended and unintended uses, owner.
2. Verify provenance and license or consent terms for each source.
3. Sample and double label a few hundred items, measure disagreement, and fix systematic errors.
4. Compare data distributions to the real deployment and list missing cases.
5. Deduplicate, and test for overlap between training, validation and test sets.
6. Scan for personal data, secrets and confidential content, and remove or protect them.
7. Lock down write access to training and retrieval data, log changes and keep versions so you can roll back.
8. Re-run the audit on a schedule and when the data source changes, and record results next to model evaluations.

Quick check

Sources

• NIST AI Risk Management Framework
• Gebru et al., Datasheets for Datasets (arXiv 1803.09010)
• NIST AI 100-2 E2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations

Last reviewed: 18 October 2026.