A shift over time in the data a model sees in production, away from the data it was trained on, which can wear down its accuracy without raising any error.
Data Drift degrades deployed AI models gradually, often without obvious errors. A model trained on pre-pandemic supply chain data, for example, will make poor predictions in post-pandemic conditions. Continuous monitoring and automated retraining pipelines, run through MLOps, are needed to detect and correct Data Drift before it affects the decisions that depend on the model.
See also MLOps, Large Language Model, Fine-Tuning and Model Inference.