Glossary

What is model drift?

Model drift is when an AI system’s results change or get worse over time even though nobody on the team changed anything. In traditional machine learning, it happens when real-world data moves away from the data the model was trained on, such as new customer behaviour or new prices. With AI services, it also happens when the provider updates or replaces the model behind the same name.

Why it matters for a business

Drift does not announce itself. A brief that was right in May comes back subtly different in November, or a report fills with stray headings because a new version lays out its answers differently, and nothing errors. The defences are to request a specific dated model version and keep it, which is called pinning, to keep a fallback model ready for when that version is retired, and to re-run a fixed set of test cases before every deliberate change.