Skip to content

Stockertim Dictionary · Technology · Human resources

Model drift

Also called concept drift, data drift, model decay

In short

The decline in a model's accuracy or fairness over time as the people, jobs or conditions it sees move away from the data it was built on.

  • Advanced
  • Concept
  • Checked

In detail

A model is a snapshot of patterns in its training data. When the world moves, the snapshot ages. Data drift means the inputs change, for example applicants start using a new CV template or a new graduate scheme brings younger candidates. Concept drift means the relationship itself changes, for example remote work alters which factors predict whether people stay.

Drift is quiet. A screening or attrition model keeps producing confident scores while their link to reality weakens, and fairness can decay even when overall accuracy holds, because the change may affect one group more than others. Vendor updates to a tool can have the same effect overnight, which is why buyers ask for advance notice of material changes.

The answer is monitoring with thresholds set in advance: compare the mix of inputs with the training data, track accuracy against real outcomes as they arrive, recalculate selection rates and impact ratios by group at regular intervals, and retrain, recalibrate or withdraw the model when a threshold is crossed. NIST's AI Risk Management Framework treats ongoing monitoring as part of managing an AI system, not an afterthought.

In practice

HR and analytics teams schedule quarterly checks of each live model, record results in the AI use case register, and require a re-test after any vendor change, new job family or significant shift in the applicant pool.

Watch out

  • Stable overall accuracy does not prove stable fairness: check results group by group.
  • Retraining on recent decisions made with the model's help can lock in its errors.

Learn it properly

Official sources

Facts last checked on 3 November 2026.

Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.

English