Model Drift
Model drift is the gradual decline in an AI system's performance after deployment as the real world moves away from what it was built on — new inputs, changed user behavior, shifting data. The model didn't change; the world it operates in did, so accuracy quietly erodes.
Also known as: drift, data drift, concept drift
AI ObservabilityAI Evaluation & Reliability
A model is trained and tested on a snapshot of the world. Once it’s live, the world keeps moving — new slang, new products, changed user behavior, shifting input distributions — and the model’s assumptions slowly stop matching reality. That’s drift: performance degrades over time even though the model is byte-for-byte unchanged. It’s dangerous precisely because it’s silent; nothing errors, the answers just get worse.
Catching it is an observability and evaluation problem, not a one-time test. You monitor live inputs and outputs for distribution shifts and quality decline, alert when metrics slide, and re-evaluate against fresh data — the same continuous feedback loop that production AI runs on. Teams that only test before launch are the ones who learn about drift from their users.
