Model drift

An AI model learns what the world looked like when it was trained. If the world changes (new customers, a different economy, other products), its predictions gradually get worse. That is called drift.

There are two types: data drift, when the characteristics of the input data change (for example, the average age of applicants), and concept drift, when the relationship between the data and the outcome changes (what used to signal a good payer no longer does).

That is why a model is not validated once and forgotten. You need to monitor performance and data-distribution metrics, set alert thresholds and decide when to retrain. For high-risk systems, this monitoring is part of the post-market monitoring the AI Act requires.