The world a model trained on keeps changing
A model learns statistical patterns from the data it was trained on, and that training happens at a specific moment in time. The problem is that the real world doesn't hold still after that moment. User behavior shifts, market conditions change, competitors launch new products, and entirely new categories of input can appear that simply didn't exist when the training data was collected. This mismatch between the data a model was trained on and the data it now sees in production is called data drift.
Drift doesn't require anyone to touch the model. The model's weights stay exactly the same, but the meaning of its inputs has shifted, so its predictions quietly become less reliable on the new reality even though nothing about the model itself changed. A fraud model trained before a new payment method existed, or a recommendation model trained before a shift in user tastes, are both still running the same math, just on a world that has moved on.