Technical
Model Drift
A change in an AI model's behavior or performance over time, often without explicit notification.
Model drift occurs when an AI model's outputs change from its original or expected behavior. For enterprises using third-party AI tools, drift typically manifests in two ways:
- Provider-side drift: The vendor updates, retrains, or replaces the underlying model (e.g., a new GPT version), changing behavior for all users
- Data drift: The distribution of inputs the model receives changes over time, causing performance degradation on certain types of queries
Drift is particularly challenging because it's often invisible — the tool keeps working, but its outputs subtly change in ways that may affect accuracy, compliance, or risk posture.
Why it matters
A tool that was reviewed and approved six months ago may behave differently today. Continuous monitoring and periodic reassessment are necessary to catch drift before it causes problems.