Model drift, where changing real-world fraud patterns: What does this pattern most likely indicate?

Declining live performance despite stable offline metrics is the hallmark of model drift.

The question

Preparing for AIGP? Take the free 5-min readiness quiz →

1. A hardware fault, where the servers running the model have begun returning corrupted numerical results during peak inference load.

   Plausible as an ops issue, but a hardware fault would typically cause errors or outages, not a gradual accuracy decline tied to shifting fraud patterns.
2. Model drift, where changing real-world fraud patterns diverge from the training data and degrade live performance over time. ✓

   Correct because falling live performance with stable offline scores is the classic signature of drift, which continuous monitoring is designed to detect.
3. An expected outcome, where any deployed model naturally becomes more accurate the longer it operates in production environments.

   Plausible-sounding, but models do not automatically improve with age, and declining catch rates signal a problem, not normalcy.
4. A labeling error, where the original training set contained mislabeled transactions that only now affect the model's predictions.

   Plausible as a data-quality concern, but stale labels would have depressed accuracy from the start, not caused a later decline as patterns shift.

How to remember it

Declining live performance despite stable offline metrics is the hallmark of model drift.

More Understanding How to Govern AI Deployment and Use questions

• An issue-management process giving users a clear channel: To apply sound deployment governance, which control →
• An audit, in which an independent reviewer systematically: Which periodic assessment activity best fits this →
• Record it in a risk register with owner and mitigations: To govern this responsibly, what should the company →
• All 424 Understanding How to Govern AI Deployment and Use questions →