Skip to main content
All glossary terms

AI governance

AI Model Drift

AI model drift is a change in the relationship between a model’s inputs and the outcomes it is expected to predict or classify. It can appear as changed input patterns, changed target behavior, or declining performance, and must be distinguished from random variation in observed results.

Also known as model drift.

Why it matters in practice

Monitoring can reveal that a model’s original validation no longer reflects the environment where it operates. Teams choose signals tied to the intended task, investigate meaningful changes, and decide whether recalibration, retraining, or withdrawal is appropriate. A drift alert is evidence to review, not proof that the model is failing.

Example

A support team notices that incoming ticket topics have shifted, checks label quality and error rates, and compares the live sample with its validation data.

Educational note

This definition is general information, not legal advice. It may omit facts that matter to a specific situation and does not establish or guarantee compliance.