Drift is the real world changing until yesterday's model stops fitting today's data.
DefinitionWhat it means
Drift describes the gradual divergence between the data a model sees in production and the data it was originally trained on, whether because customer behavior changes, a new product line launches, or an upstream data source is altered. Data scientists distinguish data drift, a shift in input distributions, from concept drift, a change in the actual relationship between inputs and the correct output, both of which quietly erode accuracy over time.
Why it mattersWhy you should care
Drift is why a model that performed well at launch can silently degrade months later without any code change, making drift monitoring a standard part of production ML operations alongside accuracy dashboards and retraining triggers. For teams buying or building AI systems, planning for drift, not just for initial accuracy, is often the difference between a model that stays useful and one that requires costly firefighting.
At a glanceSee it
A taxonomy that separates a shift in the inputs from a shift in the input-to-label rule, then splits concept drift by how fast it arrives.
The operational response loop — monitor, decide against a threshold, and either keep serving or retrain and redeploy.
Where you see itIn the wild
- Monitoring dashboards tracking input distribution over time
- Retraining triggers set off by a drift-detection alert
- Post-mortems explaining why a model's accuracy quietly dropped