Model Drift is the gradual decline in a model’s accuracy over time, as the real-world data it encounters changes and no longer matches the data it was trained on.
What it means in plain English
A model is trained on a snapshot of the world, but the world keeps changing. As patterns shift — new customer behaviour, new fraud tactics, changing language — the live data drifts away from the training data, and the model’s predictions slowly get worse. This is model drift, and it happens quietly, which is what makes it dangerous: a model can degrade without any obvious sign until performance has clearly dropped.
It is why deployed models need ongoing monitoring and periodic retraining.
A simple example
A model trained to detect fraud in 2023 gradually becomes less effective as fraudsters adopt new tactics it never saw. Without retraining, its accuracy drifts downward even though the model itself has not changed.
Why it matters
Model drift is a central challenge in keeping AI reliable in production. Recognising and addressing it — through monitoring and retraining — is a key part of MLOps and the reason a model is never truly “finished.”
Related terms
- MLOps — the discipline that monitors for drift.
- Training Data — what the live data drifts away from.
- Inference — where drift shows up as declining accuracy.
Frequently asked questions
What is model drift?
Model drift is when a deployed model’s performance degrades over time because the real-world data changes and no longer matches what it was trained on.
How is model drift handled?
By monitoring performance in production and retraining or updating the model on fresh data when drift is detected — a core part of MLOps.
