What Is AI Model Drift
Data drift vs concept drift — why your validation programme must account for it.
A validated AI system can fail without any code changes. This is model drift — gradual performance degradation as the real-world environment diverges from the training environment. EU GMP Annex 22 and the GAMP AI Guide both require ongoing monitoring programmes to detect it.
Data Drift vs Concept Drift
Data drift occurs when the statistical characteristics of input data change — a new raw material supplier, a new manufacturing line, a different patient population. Concept drift occurs when the relationship between inputs and outputs changes — the learned pattern is no longer valid. Both can invalidate a previously validated model without any code changes.
What Your Monitoring Programme Needs
Defined performance KPIs with alert thresholds, monitoring frequency proportionate to risk, a procedure for investigating alerts, and clear revalidation triggers — documented in a monitoring plan as part of the six-document AI inspection package.