GCP Generative AI Leader Practice Question 77: Drift Monitoring
Google Cloud Generative AI Leader · Test 8 · question 7 of 10
A bank deployed a gen AI model six months ago to flag potentially fraudulent transactions. Its accuracy on that launch day was strong, but customer spending patterns have shifted since then and the operations team suspects the live input data no longer resembles what the model was trained and evaluated on, which could be quietly eroding accuracy without anyone noticing.
Which recommended practice directly addresses this concern after deployment?
Show answer and explanation
Answer:
Drift monitoring compares the distribution of the data a deployed model is actually seeing against the distribution it was trained or last evaluated on, and flags when that changes significantly. It is one of Google Cloud's recommended continuous monitoring and evaluation practices, and it is exactly the tool for catching the kind of gradual shift in customer spending patterns described here.
A is wrong because temperature controls output randomness at inference time and has nothing to do with detecting whether input data has shifted. B is wrong because a one-time pre-launch evaluation only tells you accuracy at launch; it cannot catch degradation that happens afterward. C is wrong because adding prompt examples is a prompt engineering technique for improving output quality, not a way to detect that the underlying data distribution has changed.
Reference: Monitor feature skew and drift