Concept Drift
The relationship between input and outcome changes while the inputs look the same — the drift you cannot detect without labels, and therefore the one that reaches production undetected.
Reviewed July 15, 2026Stable
When not to use it
- As a diagnosis before ruling out a pipeline change, which produces identical symptoms and is far quicker to fix.
- As a synonym for data drift. They need different detection and different responses.
Reach for something else instead
- Data drift monitoring where labels are unavailable, accepting that it cannot see this failure.
- Scheduled retraining where verification latency makes detection impossible.
- Online learning where labels arrive quickly and noise is controlled.
Read more on the blog
- AI in finance: the regulator looked and stepped backBanking has had formal model risk regulation since 2011. In April 2026 the successor framework arrived and deliberately placed generative AI outside its scope. That decision is the finding.
- Zillow Offers: $304 million, in the audited filingA pricing model moved from advising consumers to committing capital. The write-down appears in a quarterly SEC filing, which makes this the best-documented AI failure in the record.
- Where AI has not landed: 77% report no use caseTransportation reports 7.5% AI use, construction 9.5%, against 73% for large information-sector firms. The most common reason given is not cost or skills. It is that no use case applies.
Further reading
- Gama et al. (2014), A Survey on Concept Drift Adaptation — the standard taxonomy of drift types and detection methods.
- Bifet & Gavaldà (2007), Learning from Time-Changing Data with Adaptive Windowing — the ADWIN detector.
- Lu et al. (2019), Learning under Concept Drift: A Review — a more recent synthesis including the verification-latency problem.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Monitoring inputs only, which cannot detect it by construction.
- Assuming stable accuracy means stable relationships, when the test set was frozen before the drift began.
- Retraining on recent data without checking whether the recent labels are themselves affected.
At a glance
FieldApplied AI
ChangesP(Y|X) not P(X)
Detectable fromlabels only
Warning timeusually none
Worst caseverification latency of months
DifficultyPractical
Where this sits
A destination. 5 concepts lead here, and nothing in the corpus depends on it.
5Levelsteps in
5Needs firstconcepts
0Opens upnothing further
1Areastays here
Learn these firstData Drift
Computed from the prerequisite graph, not assigned. How this works