Data Drift
Data drift is when the live data flowing into a deployed model gradually stops resembling the data it was trained on.
Customer behavior shifts, sensors age, markets move, and the model keeps predicting against a world that no longer exists, so its accuracy slips quietly without any error or crash to warn you. Also known as: Concept Drift, Distribution Shift.
What this topic covers
- Foundations — Start here to understand what data drift actually is: how the statistical distribution of live inputs slowly parts ways with the training data, and why a model can keep running flawlessly while its predictions quietly lose touch with reality.
- Implementation — These guides show you how to catch drift before users do: instrumenting a monitoring pipeline, choosing detection thresholds that fire on real shifts rather than noise, and deciding when a drift signal should actually trigger a retrain.
- What's changing — Drift monitoring is consolidating fast, with separate observability tools merging detection, alerting, and root-cause analysis into single platforms.
- Risks & limits — Before you trust a deployed model, consider what drift hides: a system that fails silently, degrading decisions about credit, health, or hiring long before anyone notices, and a real question of who is accountable when no alarm ever sounds.
This topic is curated by our AI council — see how it works.