Model drift occurs when the statistical relationship between model inputs and outputs changes over time, causing previously accurate predictions to become less reliable. Regular retraining with updated, relevant data recalibrates the model to current real-world patterns.

Why A is Correct: According to ISACA AAIR model maintenance guidance, regular retraining with new relevant datasets is the most direct mitigation for model drift. By periodically retraining on current data, the model learns the latest patterns and relationships—counteracting the drift that accumulates as real-world conditions diverge from the original training data. This is the standard industry practice for maintaining production AI models in dynamic environments.

Why B is Wrong: Restricting automated data validation to low-risk models creates a governance double standard that leaves high-risk models more vulnerable. If anything, high-risk models require more rigorous automated validation, not less. This approach increases rather than mitigates drift risk for critical applications.

Why C is Wrong: Maintaining existing dataset variance during preprocessing preserves statistical characteristics from a historical snapshot. If drift has occurred in real-world data, deliberately maintaining old variance levels prevents the model from adapting to new conditions.

Why D is Wrong: Role-based access controls protect model parameters and data from unauthorized modification. While important for security, access controls do not address model drift, which is driven by changing real-world conditions rather than unauthorized changes.