Concept Drift is the change over time in the relationship between inputs and the target a model is predicting, a fraud pattern evolves, a customer-intent taxono
Concept drift is change in the relationship between inputs and correct outputs: the same features now mean something different. Fraud tactics evolve so old signals go stale; 'good' content standards shift; market regimes flip predictors' meaning. Detection requires labeled feedback: rising error against fresh ground truth despite stable inputs.
Concept drift breaks models invisibly precisely because inputs look normal: dashboards show familiar traffic while decisions quietly go wrong. Any AI operating against adaptive adversaries or evolving norms, fraud, security, content, markets, must budget for continuous relabeling and model refresh.
Proxy signals: downstream outcomes (chargebacks, complaints), human-review sampling, champion-challenger comparisons, and delayed-label backtesting. Design feedback loops at system inception: bolting them on later is costly.
Remedies differ: data drift may need re-indexing or input handling; concept drift demands fresh labels and retraining, sometimes feature redesign. Misdiagnosis leads to retraining on the wrong thing.
Yes: correctness criteria evolve (policies change, products update), so prompts, retrieval corpora, and eval sets go stale together. Treat eval-set refresh as scheduled maintenance, not a one-time setup.