AML Ethical AI Practices & Data Governance 3 — Questions and Answers
Question 1: A data governance team needs to classify customer transaction records under a tiered sensitivity framework. Which classification level would MOST appropriately apply to full credit card numbers?
- Public
- Internal
- Confidential
- Restricted (Correct answer)
Correct answer: Restricted
Full PAN (Primary Account Numbers) fall under the most sensitive 'Restricted' or equivalent tier due to PCI-DSS requirements and high re-identification risk.
Question 2: Which fairness metric requires that the proportion of positive predictions be equal across protected groups, regardless of actual base rates?
- Equalized odds
- Calibration
- Demographic parity (Correct answer)
- Predictive parity
Correct answer: Demographic parity
Demographic parity (also called statistical parity) requires the model's positive prediction rate to be equal across all demographic groups.
Question 3: An organization implements a policy requiring all ML models affecting employment decisions to be reviewed by a diverse panel before deployment. This control primarily addresses which risk?
- Model overfitting
- Algorithmic bias and discriminatory outcomes (Correct answer)
- Data pipeline failures
- Adversarial attacks
Correct answer: Algorithmic bias and discriminatory outcomes
Diverse review panels are a governance control specifically designed to detect and mitigate embedded biases before a model causes discriminatory employment decisions.
Question 4: Under the NIST AI Risk Management Framework (AI RMF), which core function is responsible for identifying, analyzing, and prioritizing AI risks?
- GOVERN
- MAP (Correct answer)
- MEASURE
- MANAGE
Correct answer: MAP
The MAP function in the NIST AI RMF focuses on categorizing and identifying AI risks in their organizational and societal context.
Question 5: A practitioner uses SHAP (SHapley Additive exPlanations) values to explain a credit scoring model's individual predictions. This falls under which category of explainability?
- Global model interpretability
- Local post-hoc explainability (Correct answer)
- Intrinsic transparency
- Counterfactual explanation
Correct answer: Local post-hoc explainability
SHAP produces local, post-hoc explanations by attributing each feature's contribution to a specific individual prediction after the model is trained.
Question 6: Which principle of data governance ensures that data is accessible to authorized users in the right format at the right time?
- Data integrity
- Data availability (Correct answer)
- Data provenance
- Data stewardship
Correct answer: Data availability
Data availability ensures that authorized stakeholders can access accurate, complete data when needed for decision-making or model training.
Question 7: An AI system used in healthcare automatically flags patients for high-risk interventions. Ethically, the MOST critical requirement for this system is:
- Maximizing predictive accuracy above 95%
- Ensuring meaningful human oversight before acting on flags (Correct answer)
- Processing all patient data in real-time
- Using only open-source algorithms
Correct answer: Ensuring meaningful human oversight before acting on flags
In high-stakes healthcare settings, human-in-the-loop oversight is ethically required to prevent automated errors from directly harming patients.
AML Ethical AI Practices & Data Governance 3
A data governance team needs to classify customer transaction records under a tiered sensitivity framework.
Which classification level would MOST appropriately apply to full credit card numbers?