AI0-001 AI Security, Ethics and Governance Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Add more features

  Why it's wrong here

  Adding features does not address the input distribution itself, so the model still receives data unlike its training set; feature engineering only helps once the shift is characterised. It is tempting because extra features can lift accuracy, but that applies to underfitting, not to covariate or concept drift.

• ✓

  Retrain model with new data

  Why this is correct

  Retraining on recent data is the first practical step because distribution shift means the learned mapping no longer matches current input statistics; refreshing training data realigns the model with production. Monitoring alone would not restore performance, and the stem already identifies shift as the cause.

• ✗

  Use a different algorithm

  Why it's wrong here

  Swapping the algorithm leaves the training distribution unchanged, so the new model still learns the old mapping and fails on shifted inputs. Algorithm selection suits baseline comparison or capacity tuning, not correcting drift; the first step is to detect and characterise the shift, then retrain on representative data.

• ✗

  Reduce model complexity

  Why it's wrong here

  Reducing complexity targets variance, not the input-feature distribution change causing the drop. It cannot realign training and production data. Complexity reduction suits overfitting, where training scores far exceed validation on identically distributed data. Here, retraining on recent production data or monitoring drift addresses the actual shift.