AML Ethical AI Practices & Data Governance 2 — Questions and Answers
Question 1: A company trains a facial recognition model and discovers it has 95% accuracy on light-skinned individuals but only 78% accuracy on dark-skinned individuals. This is an example of which type of bias?
- Confirmation bias
- Representation bias (Correct answer)
- Measurement bias
- Aggregation bias
Correct answer: Representation bias
Representation bias occurs when training data underrepresents certain demographic groups, causing disparate model performance across subpopulations.
Question 2: Under GDPR, what is the maximum timeframe within which a data controller must notify a supervisory authority of a personal data breach that risks individuals' rights?
- 24 hours
- 48 hours
- 72 hours (Correct answer)
- 7 days
Correct answer: 72 hours
GDPR Article 33 requires controllers to notify the supervisory authority of a reportable breach within 72 hours of becoming aware of it.
Question 3: Which governance principle ensures that AI decisions can be traced back to their data sources, model parameters, and logic used?
- Accountability
- Traceability (Correct answer)
- Fairness
- Robustness
Correct answer: Traceability
Traceability is the governance property that enables auditors to follow the chain from inputs, through model logic, to outputs and decisions.
Question 4: A model trained to predict loan defaults is found to use zip code as a proxy for race. Under US fair lending law, this is classified as:
- Disparate treatment
- Disparate impact (Correct answer)
- Redlining proxy discrimination
- Statistical parity violation
Correct answer: Disparate impact
Using a facially neutral variable that correlates with a protected characteristic and produces discriminatory outcomes constitutes disparate impact discrimination.
Question 5: Which data anonymization technique adds calibrated statistical noise to query results to prevent re-identification while preserving aggregate statistical utility?
- K-anonymity
- Data masking
- Differential privacy (Correct answer)
- Pseudonymization
Correct answer: Differential privacy
Differential privacy provides a mathematically rigorous privacy guarantee by injecting carefully calibrated noise, bounding any single record's contribution to query outputs.
Question 6: In the context of AI governance, a 'model card' is primarily designed to:
- Document model hyperparameters for reproducibility
- Communicate intended uses, limitations, and fairness metrics to stakeholders (Correct answer)
- Store model weights in a portable format
- Define the API contract for model serving
Correct answer: Communicate intended uses, limitations, and fairness metrics to stakeholders
Model cards, introduced by Mitchell et al., are structured documentation artifacts that disclose a model's intended use, performance metrics across subgroups, and known limitations.
Question 7: The 'right to explanation' in automated decision-making, as referenced in GDPR Article 22, most directly supports which AI ethics principle?
- Beneficence
- Non-maleficence
- Transparency (Correct answer)
- Justice
Correct answer: Transparency
The right to explanation supports transparency by requiring that individuals subject to automated decisions be given meaningful information about the logic involved.
AML Ethical AI Practices & Data Governance 2
A company trains a facial recognition model and discovers it has 95% accuracy on light-skinned individuals but only 78% accuracy on dark-skinned individuals.
This is an example of which type of bias?