Explainability for High-Stakes Recommendations

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ethics-safety-AI-recommendation-explainability-adults

1. Which of the following best defines a 'high-stakes' AI recommendation?

   1. Any recommendation that involves medical data
   2. Any AI system that processes personal data
   3. AI decisions that significantly affect an individual's employment, financial status, or access to services
   4. AI systems that use deep learning models

2. Why are vague explanations insufficient for high-stakes AI recommendations?

   1. They are preferred by regulatory bodies
   2. They reduce computational costs
   3. They are easier for users to understand
   4. They fail both legal standards and user trust requirements

3. What does it mean for an AI explanation to be 'actionable'?

   1. The explanation tells users what specific factors led to the decision
   2. The explanation is provided immediately after the decision
   3. The explanation provides steps the user could take to potentially change the outcome
   4. The explanation can be appealed to a human reviewer

4. What is the primary purpose of maintaining an audit trail for AI explanations?

   1. To reduce the cost of compliance
   2. To provide users with historical access to their data
   3. To improve the accuracy of the AI model
   4. To allow regulators to review how decisions were made and explained

5. Which of the following is a legal requirement that AI cannot substitute with good explanations alone?

   1. Offering customer support
   2. Sending adverse-action notices when decisions are unfavorable
   3. Providing user-friendly interfaces
   4. Maintaining model interpretability

6. What is the difference between model interpretability and explanation quality?

   1. Quality is irrelevant if the model is interpretable
   2. Interpretability is only needed for deep learning models
   3. Interpretability refers to how the model works internally; quality refers to whether explanations actually reflect real decision factors
   4. They are essentially the same concept

7. What is required to validate that explanations actually reflect model behavior?

   1. Switching to simpler algorithms
   2. Using more training data
   3. Testing explanations with actual users and comparing to model decision logic
   4. Hiring more engineers to review the model

8. Which scenario would require an adverse-action notice under typical financial services regulations?

   1. An AI suggests a movie the user will probably enjoy
   2. An AI recommends a restaurant for dinner
   3. A credit card company AI declines a loan application
   4. An AI suggests a job candidate to interview

9. What governance element is needed for AI explanations to remain effective over time?

   1. A process for updating explanations as models and regulations evolve
   2. Fixed explanations that never change
   3. Removal of all user-facing explanations
   4. Conversion to technical-only explanations

10. Why should explanations avoid technical jargon when presented to users?

    1. Technical jargon confuses users and defeats the purpose of building trust
    2. Technical explanations are always more accurate
    3. Jargon is required by most regulations
    4. Jargon is illegal in AI systems

11. What is the relationship between user trust and explainability in high-stakes AI?

    1. Specific, actionable explanations build trust; vague ones erode it
    2. AI systems should prioritize accuracy over explaining decisions
    3. Trust is irrelevant when decisions are legally required
    4. Trust only matters for commercial AI, not regulated systems

12. When designing explainability for a hiring AI system, which component is essential for regulatory compliance?

    1. An audit trail of all explanations provided to candidates
    2. A chatbot interface
    3. A video explaining how the AI works
    4. A way to modify the AI's decision criteria

13. A healthcare AI recommends a specific treatment. What makes this a high-stakes scenario requiring strong explainability?

    1. Healthcare is always high-stakes regardless of impact
    2. The recommendation directly affects patient health outcomes and could be life-altering
    3. Patients prefer detailed explanations for medical decisions
    4. Healthcare AI always uses complex neural networks

14. What should an organization do if user testing reveals explanations are misunderstood by the target audience?

    1. Revise the explanations to use simpler language and test again
    2. Remove explanations entirely to avoid confusion
    3. Continue using the same explanations since the content is technically accurate
    4. Report the users as difficult to work with

15. Which of these represents a 'specific' reason rather than a generic category in AI explanations?

    1. Your debt-to-income ratio exceeded 43%
    2. Your credit score was too low
    3. Your application lacked sufficient qualifications
    4. Your profile did not meet our requirements