Drift occurs when a system's performance degrades relative to its validated baseline. Data drift arises when production inputs diverge from those the system was evaluated on, while model drift can follow vendor updates that change behavior.
Detection requires ongoing measurement against a stable reference.
Alternative Names:
Performance Drift, Model Degradation
Why it Matters?
In legal applications drift most often arrives through the vendor rather than through changing documents, since software-as-a-service providers update underlying models without customer approval. A system validated in March may behave differently in September with no notice. That is why change notification provisions belong in the contract and why periodic revalidation against a golden dataset is a supervision obligation rather than an optional practice.
Frequently Confused with
Related terms
Frequently asked questions
What causes drift in legal AI systems?
How is drift detected?
