Model Drift
A decline or change in model behavior over time as conditions shift.
The model stayed fixed. Did its field performance change?
A registrar checks exact accession-number matches in two independently reviewed sets of 100 scans collected in different months. The deployed recognizer is the same version.
Earlier sample · 94% exact
The recognizer gets 94 complete identifiers right and six wrong.
Later sample · 82% exact
It gets 82 right and eighteen wrong. The observed exact-match rate fell by twelve percentage points.
Find the cause before retraining
Record the sampling method and uncertainty. Compare like materials across months, replay older scans through the current pipeline, and check camera settings and transcription rules. A changed input mix, changed meaning of a target, mislabeled evaluation data or a pipeline fault can look similar in the headline metric.
How Teams Respond
Teams respond to drift through monitoring, re-evaluation, refreshed data, retraining, threshold changes, and sometimes human fallback. The key is to detect changes before they quietly damage decisions or user trust.
Model drift is a reminder that a model is not a finished truth machine. It is a living system component operating in a moving environment.
Follow the idea
Data drift measures shifts in the inputs, one possible cause of a performance change.
Model monitoring keeps the records needed to investigate the change.
Related concepts (5)
Examples in the AI guides
Source · Sculley and colleagues on feedback and monitoring in machine learning systems
Return to Model Drift in the A–Z glossary