Data or concept drift: Which continuous-monitoring finding best explains the decline and should trigger
Accuracy declining as real-world patterns diverge from training data is drift, the signal that should trigger scheduled retraining.
The question
A fraud-detection model runs unchanged in production, yet monitoring shows its accuracy steadily declining even though the code is untouched and infrastructure is healthy. Investigation finds that fraud patterns and customer behavior have shifted since training. Which continuous-monitoring finding best explains the decline and should trigger scheduled retraining?
Preparing for AIGP? Take the free 5-min readiness quiz →
- A software regression, where a recent code change silently introduced a defect that degrades the model's outputs in the production environment.Almost right but excluded: the scenario states the code is untouched, so a regression cannot explain the gradual accuracy decline.
- Data or concept drift, where the live input distribution has diverged from the training data, degrading the model's predictive accuracy. ✓Correct: shifting real-world patterns diverging from training data is drift, the classic monitoring signal that should trigger retraining.
- An infrastructure outage, where failing compute or storage resources intermittently disrupt the model's ability to serve predictions.Almost right but excluded: infrastructure is reported healthy, and outages cause availability failures rather than a steady accuracy decline.
- An adversarial data-poisoning attack, where a malicious actor corrupted the training set to embed a hidden backdoor in it.Plausible but wrong: poisoning corrupts training and yields targeted misbehavior, not the gradual, benign drift the shifting patterns describe.
The trap
Assuming declining accuracy means code or infrastructure broke rather than the environment drifting from training data. How to remember it
Accuracy declining as real-world patterns diverge from training data is drift, the signal that should trigger scheduled retraining.
How many of these would you get right?
One of 1581 AIGP questions on Certsqill. Take a free five-minute check and see your score per domain — not one number, but which section to open tonight.
Test your AIGP readiness — freeMore Understanding How to Govern AI Deployment and Use questions
- Deliver mandatory user training on the tool's approved: Which best applies the organization's deployment →
- Threat modeling: Which periodic assessment activity fits this proactive, pre-deployment structuring need? →
- A structured incident and issue log capturing each event: Which documentation practice should the deployer →
- All 424 Understanding How to Govern AI Deployment and Use questions →
Part of the Certsqill AIGP question bank · Understanding How to Govern AI Deployment and Use ·
Every answer, right and wrong, comes with its own explanation.
