Drift
The degradation of a system's relationship with its objective over time, as conditions move away from those it was built and tested under. Distinguished from a wrong answer: drift produces gradually less right ones.
- StatusIndustry term
- AdoptionWidely used
- LayerGovernance
Why it matters
Applies to the model risk function, and after SR 26-2 (Revised Model Risk Management) to whoever the institution designates for systems outside model risk management — which for agents is everyone or no one, depending on whether anyone made the designation. It bites from the day of deployment and accelerates in regime change, meaning it is worst exactly when the loss is largest. It fails as a control when detection has no pre-committed consequence: a system that flags drift into a meeting three weeks out has been documented, not governed. Of 230 financial services control objectives, only 11 address it.
Where it comes from
MARQUE (a reference framework organized around eight questions) question 03. Defined in the AIEOG Shared AI Lexicon as AI drift or decay.
Defining sources
- AIEOG Shared AI LexiconUS Treasury · FBIIC (Financial and Banking Information Infrastructure Committee) · FSSCC (Financial Services Sector Coordinating Council)Official text In the standards register
What it is not
These are routinely confused with Drift. The distinctions have consequences for how a system is governed.
- Data drift
Data drift is inputs shifting. Model drift is the input-output relationship shifting. Inputs can move with the relationship intact.
- Model failure
Failure is discrete and visible. Drift is continuous and is specifically the failure mode that does not announce itself.
How it connects
MARQUE questions it addresses
- OversightWho is accountable in the interval between decision and review?
Related framework objectives
Connected through the MARQUE crosswalk, an editorial mapping. These objectives from the Financial Services AI Risk Management Framework are assigned to the same questions that this term addresses. The ones sharing most of this term's vocabulary are shown first, as a reading aid. The assignment is to the question, not to the term: the source does not map this term to any objective directly.
Question: Oversight · 110 control objectives assigned
- GV-2.3.3AI Governance StructureGovern
- MG-3.2.2Performance Deviation DetectionManage
- MP-2.2.1AI Usage and Boundary DocumentationMap
Cite and reuse
Cite the identifier QIS-TERM-00027, not the URL: identifiers are stable, URLs may change. This entry is free to read, quote and index under the dual license. Corrections sent to editor@qisfund.com are published.
Technical details: identifier, graph relationships and record
| Predicate | Target identifier | Target | Predicate meaning |
|---|---|---|---|
addresses | marque:oversight | Oversight | The subject speaks to the object as a question or concern. |
relatedTo | concept:model-drift | Model Drift | An association too weak or too general for a stronger verb. |
relatedTo | concept:kill-criteria | Kill Criteria | An association too weak or too general for a stronger verb. |
- Canonical identifier
QIS-TERM-00027- Tier
- Canonical industry term
- Domain · layer
- MARQUE questions · Governance
- Semantic aliases
- None recorded
- First published
- 2026-08-02
- Last reviewed
- 2026-08-02 · 180-day cycle