Model Drift

The degradation of a model's performance after deployment because the world has changed while the model's learned patterns stayed fixed in the past. Drift is a failure of standing conditions rather than of the original training. Concept drift (the input-output relationship changes) and data drift (the mix of inputs changes) are its two main shapes. (Owned in depth by Topic 4.5.) (see Topic 4.5)

The silent degradation of a model's accuracy over time as the real world shifts away from the data it was trained on. Nothing breaks; the model keeps returning confident answers that are slowly getting worse. Caught only by monitoring outputs against real outcomes (ground truth).

The gradual degradation of a learned model's real-world performance as the actual operating environment diverges, over weeks or months, from the conditions its training data represented, with no single moment that resembles a conventional bug report.

The decline of an ML model's performance over time as real-world data patterns diverge from the data it was trained on.

The gap that opens between a model frozen at training time and a world that keeps moving, causing the model's real-world performance to degrade even though its code is unchanged. Drift produces no error message, which is why it is dangerous.