Data drift, where production input distributions diverge: Which factor does the evidence most directly

A steadily shifting input distribution against unchanged weights, with good launch performance, is the signature of data drift.

The question

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

1. Data drift, where production input distributions diverge over time from the training data the model learned from. ✓

   Correct because III.C.5 lists model or data drift among incident causes, and a steadily shifting input distribution with fixed weights is drift.
2. Model brittleness, where small perturbed or adversarial inputs cause disproportionate, unstable failures.

   Almost right because brittleness is a listed cause, but the evidence shows gradual distribution shift, not perturbation sensitivity.
3. Insufficient pre-release testing, where evaluation coverage failed to exercise the conditions the model would later face.

   Plausible but wrong: the model performed well at launch, indicating adequate initial testing rather than a coverage gap.
4. Poor training-data quality, where mislabeled or noisy records degraded the model's learned decision boundaries from the start.

   Plausible but wrong: strong launch performance argues against foundational data-quality defects as the primary cause.

How to remember it

A steadily shifting input distribution against unchanged weights, with good launch performance, is the signature of data drift.

More Understanding How to Govern AI Development questions

• Contain the harm: Which approach best satisfies incident management and documentation obligations under →
• Technical documentation: Which set best matches the requirement? →
• Define the business context and use case: Which activity should anchor the effort before architecture →
• All 426 Understanding How to Govern AI Development questions →