Degradation of an AI model's performance over time as real-world conditions change from what it was trained on. Like skills becoming outdated as the world changes.
A fraud detection model experiences drift as scammers develop new tactics that weren't in the training data, requiring periodic retraining.
All four clouds provide managed monitoring to detect drift (data/feature drift and/or prediction/performance drift), alert on changes, and support retraining workflows. Capabilities differ by metric depth, supported data types, and how easily you can connect monitoring outputs to pipelines for retraining.