The three types
Data drift — the distribution of inputs shifts. New customer segments, a changed channel mix, seasonal variation. The model sees inputs unlike those it was fitted to.
Label drift — the distribution of the target shifts. The base rate of the thing being predicted changes, which moves calibration even where the input–output relationship is intact.
Concept drift — the relationship between inputs and target changes. The same input should now produce a different answer. This is the type that most reliably defeats naive monitoring, because input distributions can look entirely normal while the correct answer has moved underneath them.
What causes it
Shifts in user behaviour; seasonal and cyclical variation; changes upstream in a data pipeline that alter how features are computed; regulatory or policy changes that redefine the target; and feedback loops, where the model's own decisions reshape the population it subsequently sees. Feedback loops deserve particular attention because they are self-reinforcing and invisible in input monitoring — the model is changing the world it is measured against.
How it differs from a performance regression
Drift is a cause. Regression is a symptom. A regression can be caused by drift, and equally by a model version change, a prompt change, a data pipeline fault, or a change in how the metric is calculated. Treating the two as synonyms leads to retraining as the reflex response, when the actual cause was frequently a deployment. The differential diagnosis belongs on the performance regression entry.
Why it is not the main event in LLM systems
For a model an organisation trained and hosts, classical drift is the right frame. For a system built on a hosted language model it usually is not, because the model does not stay still. Providers update hosted models on their own schedule, sometimes without a version identifier, and that step change is a larger and faster source of behaviour change than the environment moving. An LLM programme that imports a classical drift playbook ends up monitoring the variable that moves least. See model drift and degradation for the register entry covering both cases.
Detecting it
Compare predictions against ground truth as labels arrive; monitor feature distributions against the training baseline; and alert past a threshold agreed in advance rather than interpreted afterwards. Two refinements do most of the work: disaggregate, because aggregate accuracy is the metric most likely to look stable while something specific breaks; and account for label delay, because many models cannot be evaluated against outcomes for weeks or months, and in that gap you are detecting the conditions for degradation rather than degradation itself.