AI Glossary / Governance, ethics and safety / Beginner
Data quality claim
Definition
A data quality claim states how well data meets a specified quality requirement. A useful claim identifies the data, the requirement and the evidence supporting the stated level of quality, rather than merely describing a dataset as clean or reliable.
What it means
A claim is something that can be examined, challenged and updated. “These records are high quality” leaves too much unspecified. “In this release, 99% of eligible records contain a valid event timestamp under the documented rule” is more precise, though it still needs evidence.
The requirement determines what the claim means. Completeness, factual accuracy, timeliness and coverage are different concerns. Passing one test does not establish all of them. A data quality claim should therefore be interpreted within its stated purpose and scope, not as a universal guarantee about a dataset.
How it works
Begin with a requirement tied to an intended use. Identify the dataset version, included population, quality characteristic, measurement method and acceptance threshold. Perform the relevant checks or review, retaining enough detail for someone else to understand the result.
Then state what the evidence supports. A full-dataset rule check and a sample-based accuracy audit support different kinds of conclusions. Sampling may require uncertainty estimates and careful attention to how records were selected. Known exceptions and exclusions should be visible.
The claim needs an owner and a review trigger. A new data source, changed schema or revised use case can make an earlier claim inapplicable. These are practical documentation suggestions, not a substitute for the detailed requirements of a standard or agreement.
Three practical examples
- Forecasting data: A supplier states the proportion of daily observations received before a defined reporting deadline, with the observation period and excluded outages recorded.
- Labeled images: A team reports agreement between independent reviewers on a sampled set of labels, explaining the review procedure and unresolved cases.
- Customer records: A data steward claims that specified mandatory fields pass completeness checks in a named release, while making clear that completeness does not verify the fields' truth.
Think of it this way
Think of a product label that makes a specific, testable statement rather than saying “excellent.” The statement is useful only if its terms and evidence are clear. Data differs from a packaged item because it can change continuously and may be reused for purposes far beyond the original claim.
Worked example
A fictional dataset contains ten thousand eligible event records. A documented check finds that 9,850 contain timestamps matching the permitted format and range rules. The observed pass rate is 98.5%.
If the requirement is at least 99%, the dataset does not meet that requirement: it falls short by 0.5 percentage points. A truthful claim might state the measured rate, rule version and release identifier, together with the planned remediation.
It would be misleading to round the result to “about 99% compliant” and imply acceptance. It would also be wrong to claim 98.5% factual timestamp accuracy, because the test assessed permitted structure and range, not whether the recorded events occurred at those times.
Illustrative case study
Fictional case: Data Meadow supplies records for a service-demand model. Its initial handover says the data is fully checked, but the receiving team cannot determine what was tested.
The supplier replaces that phrase with separate claims for field completeness, duplicate identifiers and sampled source agreement. Each points to a dataset version and evidence file. An unresolved regional coverage gap remains visible instead of being absorbed into one average score. This invented case demonstrates clearer communication, not certification or measured business results.
Common misconceptions and mistakes
A claim is not the same as proof. Its credibility depends on the evidence, methods and scope. A machine-generated validation report can still use a flawed rule or exclude problematic records.
Do not broaden a sampled finding to every record without justification. Do not confuse a supplier's self-declaration with independent assessment. Another mistake is carrying a claim forward to a new dataset release without checking whether relevant conditions changed. A high overall percentage can also hide severe failures in a small group important to the application.
Business use
Clear quality claims improve handovers between data providers, model developers and decision owners. They make it easier to decide whether data is suitable, what additional checks are needed and who should resolve exceptions.
Use claims that connect to actual acceptance criteria and preserve their supporting evidence. Avoid marketing language that outruns the measurement. When data supports several applications, maintain separate suitability assessments where requirements differ. A claim about forecasting readiness, for example, does not automatically establish suitability for individual-level decisions.
People also ask
Is a data quality claim always numerical?
No. It can be qualitative, but the requirement, scope and supporting evidence should still be clear.
Does a completeness claim establish accuracy?
No. A populated field can contain an incorrect value.
Can a claim become outdated?
Yes. Data changes, revised requirements or new uses can invalidate the original scope.
Check your understanding
A provider says every record passed a date-format check. Can a user conclude that every date reflects the correct real-world event?
Answer: No. Format validity and factual accuracy require different evidence.
Related terms
- Data quality measure
- Evidence
- Data quality plan
- Data quality requirement
- data quality assessment
- dataset version
- data provenance
- acceptance criterion
Learn next
Rewrite a vague statement such as “the training data is reliable” into two narrow claims. For each, identify the population, metric or review method, threshold, evidence and exceptions. Ask a colleague to determine whether the claim could be reproduced without additional explanation.
More in Governance, ethics and safety
- Data poisoning - Data poisoning is deliberate interference with data used by an AI system in order to cause unwanted behavior.
- Data privacy - Data privacy concerns how information about people is collected, used, shared, retained, and otherwise processed, and how those activities affect the people...
- Data proxy - A data proxy is observed information used to stand in for another quantity or characteristic that is difficult, unavailable, or impossible to measure directly.
- Data quality model - A data quality model organizes the characteristics used to define and evaluate data quality.
- Data quality planning - Data quality planning is the activity of setting data quality objectives and deciding which processes and resources are needed to achieve them.
- Data record - A data record is a collection of related data items handled as one unit.
Sources
ITU-T Y Supplement 97 (11/2025), term 3.1.32, consulted directly in the official PDF: https://www.itu.int/rec/T-REC-Y.Sup97-202511-I/en. The glossary attributes this definition to ISO/IEC 5259-3. ISO publisher overview consulted: https://www.iso.org/standard/81092.html; the complete ISO standard was not accessed. The examples and workflow are original explanatory material, not prescribed thresholds or a conformity assessment.
Previous: Data Quality Next: Data quality dimensions