CAIC Ethical Considerations in AI 3 — Questions and Answers
Question 1: The EU AI Act classifies AI systems into risk tiers. Which category includes biometric surveillance in public spaces?
- Minimal risk
- Limited risk
- High risk
- Unacceptable risk (Correct answer)
Correct answer: Unacceptable risk
The EU AI Act places real-time remote biometric identification in public spaces in the 'unacceptable risk' category and generally prohibits it.
Question 2: A 'model card' is a documentation tool that primarily communicates:
- Licensing fees for commercial use
- Model performance, intended uses, and known limitations (Correct answer)
- Source code and training scripts
- Hardware requirements for inference
Correct answer: Model performance, intended uses, and known limitations
Model cards provide structured documentation about a model's intended purpose, evaluation results, and ethical considerations for transparency.
Question 3: Which term describes the risk that an AI system optimizes for a measurable proxy goal instead of the intended real-world objective?
- Overfitting
- Concept drift
- Reward hacking / Goodhart's Law (Correct answer)
- Catastrophic forgetting
Correct answer: Reward hacking / Goodhart's Law
Reward hacking (related to Goodhart's Law) occurs when an AI exploits loopholes in the specified metric rather than achieving the true underlying goal.
Question 4: An AI consultant recommends deploying a facial recognition system at a client's office. What ethical obligation must be disclosed to employees?
- The vendor's pricing model
- That their biometric data is being collected and how it will be stored (Correct answer)
- The model's architecture details
- Competitor systems evaluated during procurement
Correct answer: That their biometric data is being collected and how it will be stored
Employees have the right to know when biometric data is collected, its purpose, and retention policies under informed consent and many data protection laws.
Question 5: Which approach helps ensure fairness by requiring that an AI's error rates are equalized across demographic groups?
- Individual fairness
- Equalized odds / demographic parity (Correct answer)
- Calibration weighting
- Federated learning
Correct answer: Equalized odds / demographic parity
Equalized odds is a group fairness criterion requiring that true positive and false positive rates be equal across protected demographic groups.
Question 6: A client asks you to build an AI system that automatically rejects loan applications from certain zip codes. Your best response as an ethical AI consultant is to:
- Build it with extra logging for audit trails
- Refuse and explain that zip codes often serve as proxies for race, violating fair lending laws (Correct answer)
- Build it but cap the rejection rate
- Use a black-box model so the pattern is less detectable
Correct answer: Refuse and explain that zip codes often serve as proxies for race, violating fair lending laws
Using zip codes as loan criteria can constitute illegal redlining by acting as a proxy for protected characteristics under the Equal Credit Opportunity Act.
Question 7: The principle of 'data minimization' in ethical AI means:
- Compressing model weights to reduce storage
- Collecting only the data strictly necessary for the specified purpose (Correct answer)
- Using synthetic data exclusively during training
- Minimizing the number of training epochs
Correct answer: Collecting only the data strictly necessary for the specified purpose
Data minimization limits privacy risk by ensuring organizations collect, process, and retain only data truly necessary for their declared purpose.
CAIC Ethical Considerations in AI 3
The EU AI Act classifies AI systems into risk tiers.
Which category includes biometric surveillance in public spaces?