Definition
Model drift is a broad operational term for model behavior or performance moving away from an accepted baseline over time. The observed change might come from the environment, the model, or the system wrapped around it.
Operational use should define the baseline, the measured behavior, the affected slices, the threshold, and the investigation path. Preserve model and harness versions so a detected change can be traced to a deployment rather than guessed from aggregate metrics.
Why it is contested
The disagreement is not whether deployed behavior can change. It is whether model drift names a cause, a measured effect, or an umbrella covering both. Input data can shift, the relationship between inputs and targets can change, a provider can update a model, or a team can change prompts, retrieval, tools, thresholds, and infrastructure. Calling every movement model drift can hide the component that needs correction. A useful claim must identify the baseline, the changed measurement, and the suspected mechanism.
Distinguish it from nearby terms
Data drift changes input distributions. Concept drift changes the relationship between inputs and correct outcomes. Performance degradation is the measured effect. Model drift is often used as an umbrella across these causes and effects.
Check your understanding
Answer quality falls after a retrieval-index update while the model version stays fixed. The system drifted from baseline, but blaming the model would send the investigation to the wrong component.