What is Model Drift? — Plain-Language AI Definition

What is Model Drift?

Model drift happens when an AI system gets worse over time because the world changes. The model may have worked well at launch, but user behavior, language, input patterns, or business conditions shift, and the old setup no longer fits the current reality.

The model itself may not have changed at all. The environment changed around it.

Why It Matters

A drifting system can fail quietly. Teams may not notice the quality drop until customers complain, errors increase, or business metrics move in the wrong direction.

That is why monitoring matters after launch, not just before it.

Common Causes

• user behavior changes
• new product lines or policies
• seasonal or market shifts
• new language, jargon, or document formats
• changes to upstream data quality

Signs of Drift

Teams may see:

• lower accuracy or acceptance rates
• more user overrides or corrections
• more hallucinations or formatting failures
• safety or refusal behavior changing unexpectedly

How Teams Respond

Typical responses include running updated evals, improving prompts, retraining models, refreshing reference data, or adding stronger review steps.

Key Takeaway

Model drift is not a rare edge case. It is a normal part of operating AI in the real world. If you deploy AI, you need a plan to notice drift and respond to it.