Model drift, where the statistical relationship between: Working with data scientists and business owners,

Gradual post-launch accuracy loss driven by a shifting input-target relationship, with stable schemas and clean launch tests, is the signature of model/data drift.

The question

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

1. Model drift, where the statistical relationship between inputs and the target has shifted over time so a once-accurate model no longer reflects current real-world conditions. ✓

   Correct because gradual post-launch degradation tied to a shift in the input-target relationship, with stable schemas and clean launch tests, is the defining signature of model or data drift rather than a build-time defect.
2. Brittleness, where small perturbations or edge-case inputs cause disproportionate and unstable failures that surface as sudden, sharp drops in the model's predictive accuracy.

   Plausible because brittleness is a genuine failure mode, but it produces abrupt, perturbation-triggered failures rather than the steady drift-driven decline described, so it does not fit the pattern.
3. Insufficient testing, where inadequate validation coverage before release allowed latent defects to remain hidden and then emerge as accuracy loss once the system met production traffic.

   Plausible because weak testing does cause post-release failures, but the launch tests passed cleanly and degradation was gradual, pointing to changing conditions rather than an undetected build defect.
4. Lack of quality data, where noisy, incomplete or unrepresentative training data embedded systematic errors that consistently biased predictions from the very first day of deployment.

   Plausible because poor training data is a common root cause, but it would degrade accuracy from launch rather than after eight months of initially strong performance, so it misreads the timeline.

How to remember it

Gradual post-launch accuracy loss driven by a shifting input-target relationship, with stable schemas and clean launch tests, is the signature of model/data drift.

More Understanding How to Govern AI Development questions

• Log the incident in a central register with root cause: Which response best discharges the obligation to →
• Maintain technical documentation: Under the EU AI Act, which combination best satisfies the provider's →
• Specify the intended purpose: Which action most directly fulfills that objective? →
• All 426 Understanding How to Govern AI Development questions →