Model Drift

Model drift is the gradual degradation of an AI system's performance caused by changes in the data it encounters in production relative to the data it was trained on. Drift can take many forms: the statistical distribution of inputs shifts, new patterns emerge that the model was not designed for, or the relationship between inputs and outcomes changes. 

Because drift is gradual and often invisible without systematic monitoring, organizations typically discover it only after measurable harm has occurred. At Eticas.ai, drift detection is a core component of post-deployment monitoring. We track metrics automatically against benchmarks established during the initial evaluation, using both statistical tests for distribution shift and performance metrics disaggregated by population group, so that drift affecting specific communities is detected early.