Glossary
Behavioral Drift
A gradual, unintended change in how a model behaves over time — often across model updates, prompt changes, or accumulating context — that wasn't explicitly planned.
Behavioral drift is the slow erosion of intended behavior that can happen without anyone noticing until it becomes a problem. It might occur when a model is updated and subtly changes how it handles a category of requests, when a prompt is iterated on and an important constraint gets dropped, or when user interactions shift the distribution of inputs in ways that expose behavioral inconsistencies. Drift is dangerous because it’s gradual — no single change looks alarming — and because it can happen in any direction, including becoming more or less cautious, more or less verbose, or more or less accurate in specific domains. For behavior architects, continuous monitoring and regression testing are the primary defenses against undetected behavioral drift.