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.

• ✗

  Remove the ZIP code feature from the model

  Why it's wrong here

  Removing ZIP code discards a legitimate predictive feature while leaving the underlying disparity unaddressed; the high false positive rate stems from biased training data or proxy variables, not the column itself. Dropping a feature is tempting as a quick fairness fix, but it is the right choice only when that feature is itself the protected attribute and no proxies remain.

• ✗

  Increase the decision threshold for those ZIP codes

  Why it's wrong here

  Raising the threshold shifts the precision-recall trade-off globally for that group; it does not remove the ZIP code proxy encoding historical bias, so disparity persists. It tempts because threshold tuning is a legitimate calibration step, and it would be correct when overall error rates, not group-specific fairness, are the concern.

• ✗

  Discontinue use of the model for those ZIP codes

  Why it's wrong here

  Withdrawing the model from those ZIP codes denies credit rather than repairing the biased feature, leaving the proxy variable in place elsewhere. It tempts because it immediately halts harm in affected areas, and it would be correct as a temporary containment measure while the model is retrained without the ZIP code feature.

• ✓

  Retrain the model with fairness constraints

  Why this is correct

  Retraining with fairness constraints directly targets the ZIP-code disparity by adding a regularisation term that penalises disparate false positive rates across protected groups during optimisation. This satisfies the audit finding: the model's error rates vary by geography, so the constraint forces the learner to equalise them.