AI0-001 AI Concepts and Foundations 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.

• ✓

  Provide adverse-action explanations for declined applicants that identify the principal factors influencing the decision.

  Why this is correct

  Adverse-action explanations are legally required in many credit contexts and directly support transparency. By identifying the principal factors behind a decline, the firm helps applicants understand and potentially contest decisions. This practice also forces the model to be interpretable at the individual level, which is a key responsible-AI requirement when using alternative data that applicants may not expect to influence credit outcomes.

• ✓

  Conduct disparate-impact testing across protected groups and document the results before deployment.

  Why this is correct

  Disparate-impact testing measures whether the model's outcomes disproportionately disadvantage protected groups, which is a core fairness safeguard. Documenting results creates an audit trail for regulators and internal governance. In credit scoring with alternative data, this testing is essential because variables like rental history can correlate with protected characteristics, so measuring impact across groups directly addresses the compliance team's concern.

• ✗

  Use only the most predictive features and remove any feature that reduces overall accuracy.

  Why it's wrong here

  Optimizing solely for accuracy ignores fairness and transparency obligations. Removing features that reduce accuracy could eliminate important protected-group indicators or fairness constraints, worsening disparities. In a regulated credit context, this approach conflicts with responsible AI principles because it treats predictive performance as the only objective and provides no mechanism to detect or mitigate discriminatory impact.

• ✗

  Retrain the model monthly on all new applications without human review to keep pace with changing data.

  Why it's wrong here

  Fully automated monthly retraining without human review can introduce or amplify bias as data distributions shift, and it undermines transparency and accountability. Responsible deployment requires oversight, validation, and documentation of changes. In a regulated credit environment, uncontrolled retraining could violate model-risk-management requirements and make it impossible to explain why a given applicant was declined.

• ✗

  Encrypt the training data at rest and restrict access to the data science team.

  Why it's wrong here

  Encryption and access control are important security measures, but they address data protection rather than fairness or transparency. The compliance team's concern is about discriminatory outcomes and explainability, not confidentiality. While these controls are necessary for privacy compliance, they do not test for disparate impact or provide explanations to applicants, so they do not satisfy the responsible-AI practices described.