How is concept drift different from data drift?

Data drift is the inputs changing; concept drift is the meaning of the inputs changing, even if inputs look the same.

How often should models retrain?

As often as drift demands: monthly for fast-moving domains, quarterly for stable ones, always with monitoring to catch the exceptions.

Can drift be predicted?

Rarely in advance, but its symptoms, declining live accuracy and shifting statistics, can be caught early.

Does drift affect generative AI too?

Yes; assistants grounded on old documents drift as facts change, which is why retrieval from current sources matters.

AI Glossary / AI and machine learning / Beginner

Concept drift

Definition

Concept drift is the gradual or sudden change in the relationship between a model's inputs and its target, so that patterns the model learned stop matching the world it now predicts for.

What it means

A model trained last year learned last year's rules: what a normal transaction looks like, which customers churn, what demand follows a price cut. The world does not stand still.

When the underlying relationship changes, the model's accuracy decays even though nothing in the model broke. Fraudsters adapt, customer tastes shift, a competitor changes prices, a law changes behavior.

Drift comes in speeds. Gradual drift is a slow slide, such as language evolving in reviews. Sudden drift is a cliff, such as a pandemic rewriting shopping habits in a week.

It differs from data drift, where the inputs themselves change, such as a new device mix. Concept drift is deeper: the same inputs now mean something different.

For a business, concept drift is the reason every model has a maintenance budget. The question is never whether drift will come, but whether you will notice it before it costs you.

The term covers both the slow kind, where yesterday's model is a little worse each month, and regime change, where the old relationship simply stops existing.

How it works

Monitor live performance where outcomes arrive: accuracy on last month's confirmed cases, calibration of recent scores.

Watch input and output statistics too: shifts in the feature mix or in the predicted-class balance are early smoke signals.

When drift is detected, retrain on recent data, possibly weighting newer examples more heavily so the fresh pattern dominates.

For sudden drift, sometimes only post-event data should be used, because the old world is simply gone.

Set a review cadence: even without alarms, schedule periodic retraining, because slow drift is easy to sleep through.

Three practical examples

Fraud models needing constant refresh as criminals adapt their tactics to whatever the model learned.
Demand forecasting breaking when a competitor's promotion changes how customers respond to your prices.
Spam filters retraining as spam language evolves to evade the words filters learned last quarter.

Think of it this way

Think of a sailor who memorized the channel markers years ago. The seabed shifts, the markers are moved, but the sailor's map is unchanged. Every trip still works, right up until one does not.

The limit: you cannot see the seabed. You notice drift by watching where the boat actually goes, which is why monitoring outcomes beats admiring the map.

Worked example

the churn model that quietly expired

Suppose a telecom's churn model scored 85 percent accuracy at launch. Nobody monitors it; it is part of the furniture.

Over 18 months, accuracy slides to 70 percent as a price war changes why customers leave. The retention team still acts on its scores, wasting calls on the wrong customers. Figures are illustrative.

A new analyst plots monthly accuracy, sees the slide, and retrains on the last six months of data. Accuracy recovers within a week.

The standing fix: accuracy on the dashboard, a drift alert threshold, and scheduled retraining every quarter.

Illustrative case study

A fictional insurer's claims model, trained on years of stable data, fails after a regulatory change alters how customers claim. Losses accumulate for two quarters before the cause is found.

The insurer adopts champion-challenger: a fresh model retrains monthly and must beat the incumbent on recent data before taking over.

The invented lesson: drift response should be a standing process, not a post-mortem. By the time someone asks why results slipped, the slip has been compounding for months.

The insurer also starts logging why each retraining happened: scheduled, alert-driven or incident-driven. Within a year the log shows drift arriving like weather, regularly, and budgeting for it stops being an argument.

Common misconceptions and mistakes

If the model breaks, you will see an error. Models fail silently; they keep producing confident scores while the world moves on.
Drift means bad data. The data can be perfect; it is the relationship between inputs and outcomes that changed.
Retraining once fixes it. Drift is continuous, so the response must be continuous too: monitoring, scheduled retraining, alerts.

Business use

Put live accuracy monitoring and a retraining schedule into every model's budget from day one; a model without them is a depreciating asset with no service plan.

Ask vendors how their model detects and responds to drift, and who owns that response after you sign.

Contracts for bought models should say who watches for drift and who pays for retraining; a model sold as finished is a model sold without its maintenance plan.

Check your understanding

Why does a model's accuracy decay even when the model itself is unchanged?

Answer: Because the relationship between inputs and outcomes in the world has moved on, while the model still applies the patterns it learned from the old world.

Related terms

Learn next

AI for Business Owners covers how to keep bought and built models honest as your market moves: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S2] Google, Machine Learning Glossary; [S3] NIST AI Resource Center glossary. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Concept activation vector Next: Concept map