AI glossary · Basic · Data engineering · Data & evaluation
Data quality
Data quality is the degree to which data is accurate, complete, consistent and up to date enough to be fit for its purpose. Analyses or AI models built on low-quality data produce results that are hard to trust.
About this term
- Level
- Basic · Technical terms you often meet in AI news
- Category
- Data, ops & platforms
- Added
- 2026-10-08
How it works
Data quality is usually broken down into a few criteria. The common ones are whether values match the facts (accuracy), whether there are no empty fields (completeness), whether values are the same across systems (consistency), whether data is recent enough (timeliness), whether the same record is not duplicated (uniqueness), and whether data follows the defined format (validity). These rules are checked automatically at points in the data pipeline, and when data falls outside the standard, an alert is sent or the data is not passed on to the next stage.
Quality criteria and common problems
| What it checks | Common problem | |
|---|---|---|
| Accuracy | Does it match the facts? | Wrongly entered values |
| Completeness | Are there no blanks? | Missing fields |
| Consistency | Is it the same across systems? | Different definitions in each department |
| Timeliness | Is it recent enough? | Retired old documents |
| Uniqueness | Are there no duplicates? | Multiple copies of the same record |
| Validity | Does it follow the format? | Mixed date formats |
Why it matters
As the saying 'garbage in, garbage out' goes, models learn the errors and biases in data as they are. In LLM training especially, as the view spread that data quality matters more for performance than quantity, deduplication and filtering out low-quality documents became important steps. In business AI too, if wrong documents are retrieved, the answers sound plausible but are wrong. The later a quality problem is found, the more it costs to fix, so it is best to catch it early, when data comes in.
Things to know
Perfect data hardly exists, so it is realistic to set a 'good enough quality' standard for each use. Data that was high quality at first can drift away from reality over time; this is called data drift. So quality checks are not one-off and need to be repeated continuously. Setting quality standards and owners goes hand in hand with data governance.
Example
Teams running internal knowledge search chatbots often run into a problem where the old and new versions of the same policy are both retrieved, and the chatbot points users to a retired policy. To prevent this, they add quality-control steps to the pipeline: attaching creation dates and validity periods to documents, filtering out duplicate documents, and removing outdated documents from the index.