A model is trained on data from a particular time, but the world it operates in keeps changing. As habits, prices, language or conditions shift, the patterns the model learned no longer match reality, and its accuracy degrades gradually, a problem known as drift.
The danger is that nothing breaks loudly; the model keeps producing confident answers that are slowly getting worse. Catching it means monitoring real outcomes over time, not just testing once at launch, and retraining when the gap between prediction and reality grows too wide.