What is the difference between data drift and concept drift?
Data drift means the inputs changed: a new customer segment arrives, a sensor is recalibrated, a field’s meaning shifts upstream, exactly the silent change a data contract exists to catch. Concept drift means the relationship changed: the same inputs now imply a different correct answer, the way fraud patterns mutate precisely because a model started catching the old ones. Data drift you can see in the inputs alone; concept drift only shows up when outcomes are compared against predictions.
Why is model drift dangerous rather than just annoying?
Because nothing crashes. A drifting model returns well-formed, confident outputs that are increasingly wrong, and the error surfaces far downstream: approval rates shift, forecasts miss, complaints rise. The failure mode is silence, so the control must be active measurement, not waiting for alarms that deterministic software would have raised. IBM’s working definition captures the cause: performance degradation from changes in data or in the relationships between input and output variables (IBM, read 12 October 2026).
How do you detect model drift before users do?
Watch two layers: input distributions and outcome quality. Inputs: compare live feature distributions against the training baseline, which is statistical anomaly detection on your own telemetry. Outcomes: score predictions against ground truth as it arrives, on a dashboard someone owns, as part of ordinary AI observability. Set the thresholds when the model ships, while model validation still has everyone’s attention. A model without a drift monitor has an expiry date nobody wrote down.
Does model drift apply to LLMs and agents?
The mechanism generalizes even where you never retrain. A prompt tuned for one model version behaves differently after a provider update; retrieval grounded in stale documents answers from a world that moved; an agent’s tools change under it. Anything whose quality depends on a snapshot of the world decays with the world, so version what you depend on, evaluate on a schedule, and treat every upstream change as a reason to re-measure.