What is data quality in an AI context?
Data quality describes how well data fit a specific use case. For AI it is decisive: biased, incomplete or stale inputs still get learned. “Garbage in, garbage out.”
DEFINITION
Data quality is the degree to which data are fit for a stated purpose—it is never an absolute score. Great marketing analytics data can be useless for clinical triage.
Why AI multiplies the stakes: models learn from whatever is present and absent. Errors, skewed sampling and blind spots become behaviour the system reproduces—often amplified. The old systems adage still applies: garbage in, garbage out.
Five practical dimensions
- Completeness – are required fields actually populated? Missingness warps many algorithms.
- Correctness – are facts right? Wrong master data produces wrong outputs.
- Consistency – do sources agree? Contradictory feeds confuse training signals.
- Currency – is freshness adequate for the decision being automated? Stale training mirrors yesterday’s world.
- Representativeness – do examples cover the populations and edge cases you must serve? Undercoverage is a primary driver of bias.
Quality is not a one-off cleanse—data rot, processes drift and use cases evolve, so stewardship is continuous.
CONNECTIONS
Leadership
Psychological safety around reporting bad records accelerates detection: people flag suspect rows without fearing blame games.
Agility
For data-heavy AI increments, include data QA inside the Definition of Done—inputs validated before outputs are celebrated.
Project management
Treat data risk explicitly in the register: owners, mitigations and monitoring—most model failures trace to inputs, not algorithms.
KEY POINTS
- Models inherit whatever statistical story the dataset quietly tells—including silence.
- The five-dimensional checklist anchors conversations between domain, analytics and engineers.
- Fit-for-purpose beats abstract “perfect” warehouses.
- Bias diagnoses often rewind to upstream representation—not mystical model malice.
- Stewardship rhythms beat heroic weekend clean-up sprints.
EXAMPLE
A lender trains approvals on decade-old officer decisions unknowingly skewed toward certain neighbourhoods. The classifier mirrors that skew at scale—not because engineers desired harm but because training material encoded it. Remediation merges statistical parity work with lineage discipline and fairness reviews as part of data quality—not an afterthought flag.
MISCONCEPTIONS
Won’t sheer volume outweigh messy fields?
Volume multiplies flawed patterns confidently. Thousands of audited rows routinely beat millions of polluted ones.
Isn’t cleansing purely engineering’s job?
Pipelines enable enforcement, yet semantic truth—what “good” means—lives with domain stewards collaborating through shared definitions and escalation paths.