A fraud model learns that certain purchase patterns signal stolen cards. Then fraudsters change tactics. The transactions look about as they did, with similar amounts, merchants and times, but the patterns that once meant fraud now mostly mean ordinary customers, and new fraud looks normal. Accuracy falls even though nobody touched the model. The inputs barely moved; what they mean for the outcome did.
Researchers separate this from changes in the inputs alone, often called data drift or covariate shift, such as a sudden wave of customers from a new country. Some writers use “concept drift” for both kinds of change and call the input-only kind “virtual drift,” so check which sense a tool or paper means. The practical difference matters. Input changes can be spotted by watching the data itself; a changed input-outcome relationship usually shows up only when real outcomes arrive, such as confirmed fraud cases weeks later.
Fixing concept drift usually means collecting recent labeled examples and retraining or adapting the model; redeploying the same model on the same old data doesn’t help. It is one cause of model drift, the broader term for prediction quality falling after deployment. Watching for it is part of AI observability.
Sources
- Lu et al.: Learning under Concept Drift: A Review — Surveys drift detection, understanding and adaptation; separates changes in the input distribution, sometimes called virtual drift, from changes in the input-outcome relationship that move the decision boundary; checked October 24, 2026.
Go deeper
- Rabanser et al.: Failing Loudly paper
Compare methods for detecting dataset shift and the study's evidence that some detected shifts do not harm predictions.