Model drift is the phenomenon where a machine learning model's performance degrades over time as the statistical properties of real-world data diverge from the data the model was originally trained on. It is one of the most common reasons deployed ML models fail to deliver consistent results in production environments.
There are two primary types of drift that practitioners monitor. Data drift (also called covariate shift) occurs when the distribution of input features changes, for example, when customer demographics shift or seasonal patterns alter purchasing behavior. Concept drift happens when the underlying relationship between inputs and outputs changes, meaning the patterns the model learned are no longer valid even if the input distribution remains stable.
Detecting model drift requires continuous monitoring of key performance metrics such as accuracy, precision, recall, and prediction distributions. Statistical tests, drift detection algorithms, and automated alerting systems help teams identify when a model's outputs begin to deviate from expected baselines, triggering retraining or recalibration workflows.
Addressing model drift is a core responsibility within MLOps practices. Organizations that proactively monitor and manage drift can maintain the reliability of their AI-powered features, avoid costly errors from stale predictions, and ensure that their machine learning investments continue to deliver value as real-world conditions evolve.