Model drift occurs when an AI system's behavior or measured performance changes over time. Causes can include model updates, altered prompts or tools, new input patterns, changes in user behavior and shifts in the domain where the system operates.
Monitoring and repeated evaluation help detect drift before it causes widespread harm. In clinical settings, checks must also distinguish model changes from updated medical standards, because both can make an older evaluation rubric unreliable.
ELI5
Model drift is a change in how an AI system behaves or performs over time. It can happen because the model, prompts, tools, users, input data or real-world conditions have changed.
For example, a support classifier trained on last year's product questions may become less accurate after a major product launch introduces new terminology. Repeated evaluation can detect the change so the model or process can be updated safely.
What causes AI model drift?
Model updates, changing data, new users, prompt or tool changes and evolving operating conditions can all alter system behavior.
How is model drift detected?
Teams compare current outputs and metrics with validated baselines, investigate new failure patterns and use expert review for contextual changes.
