Data drift is a change in the statistical properties of the input data over time. In practical terms, it means the examples a model sees in production are no longer shaped like the data it was trained or validated on. When that happens, performance can degrade even if the model itself has not changed.
Why Drift Matters
Models are built on assumptions about the world reflected in their training data. But the world changes. Customer behavior shifts, fraud patterns evolve, sensors age, product catalogs expand, language changes, and new workflows appear. When the input distribution drifts, the model may become less accurate, less fair, or less well calibrated.
Data drift is one of the main reasons AI systems need ongoing observation after launch rather than one-time approval.
How Teams Detect It
Teams detect drift by monitoring feature distributions, output behavior, confidence changes, and business outcomes. Some use statistical tests. Others use dashboards, alerting thresholds, or anomaly detection pipelines. The goal is not just to know that drift exists, but to understand whether it matters enough to retrain, recalibrate, or redesign the system.
Drift detection works best when tied closely to Model Monitoring. A change in the data is only meaningful if teams can connect it to real model behavior and operational risk.
Why Readers Should Learn It
Data drift is a valuable term because it explains why an AI system that once looked strong can later become unreliable without any dramatic bug. Many AI failures are not sudden breakdowns. They are quiet mismatches between an old model and a changing environment.
For AI literacy, drift is one of the clearest reasons that deployment is not the end of the story.
Related concepts: Model Monitoring, Model Drift, Model Evaluation, Anomaly Detection, and Calibration.