Glossary
Data & Model Drift
When the world changes and a deployed model's accuracy degrades.
Definition
Drift is the phenomenon where a deployed AI model's accuracy degrades over time because the world it was trained on has changed. Data drift = input distribution changes. Concept drift = the relationship between input and correct output changes.
Context
All production models drift. Monitoring for drift — via input-distribution checks, output-quality metrics, and downstream business KPIs — is a core MLOps responsibility. LLM-based systems are somewhat more robust to drift than narrow ML models, but grounding data (embeddings, retrieval indexes) still needs regular refresh.