Building Privacy, Trust, and Safety in AI Systems CRAGE Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   What is the primary role of human oversight in AI safety?

   • ATo increase the speed of AI decision-making
   • BTo automate all safety checks without human involvement
   • CTo monitor AI behavior and intervene when necessary to prevent harmful outcomes
   • DTo replace AI systems entirely with human decision-making
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A hospital is deploying an AI system that predicts patient deterioration. The system will be used in intensive care units (ICUs) and will alert clinicians to intervene. The hospital must balance the need for early intervention with the risk of alert fatigue and false alarms. The risk assessment team has identified that the system has a high sensitivity but a low specificity, leading to many false positives. Which mitigation strategy best balances the trade-off between sensitivity and specificity while maintaining patient safety?

   • ADisabling the AI alerts and relying solely on clinicians' manual monitoring.
   • BAdjusting the alert threshold to reduce false positives, while maintaining a mandatory human review for all alerts that are acted upon.
   • CSetting the alert threshold to maximize specificity, so that only the most certain cases trigger alerts.
   • DIncreasing the sensitivity further to ensure no deterioration is missed, even if it increases false alarms.
   Reveal Solution Discussion Select an answer first
4. 18expert · hard

   A bank is implementing an AI system for credit scoring. The system uses a complex machine learning model that is difficult to interpret. The bank must conduct a Data Protection Impact Assessment (DPIA) and also comply with regulations that require meaningful explanations for automated decisions. The DPIA reveals that the model may be biased against certain demographic groups. Which combination of actions best addresses both the DPIA findings and the explainability requirement?

   • AUsing an interpretable model (e.g., logistic regression) that provides clear explanations, and conducting bias testing to identify and mitigate disparate impact.
   • BConducting the DPIA but not implementing any changes until regulators require it.
   • CKeeping the complex model but using a separate explanation tool (e.g., LIME) to generate post-hoc explanations.
   • DReplacing the AI system with a rule-based system that is fully transparent, but may be less accurate.
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A social media platform is introducing an AI feature that analyzes user posts to suggest content. The platform wants to respect user data rights, including the right to access, rectification, and erasure. The product team is designing the user interface for consent and data management. Which design choice best supports informed consent and user data rights?

   • AProviding a clear, plain-language consent form that explains what data is collected and how it is used, and offering easy-to-find controls for users to access, correct, or delete their data.
   • BObtaining consent once at account creation and not offering any further data controls.
   • CRequiring users to contact customer support by email to request data access or deletion.
   • DUsing a long, legalistic consent form that covers all possible uses of data.
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   Which safety mechanism is designed to automatically stop an AI system from operating when a critical failure is detected?

   • ARobustness testing
   • BData anonymization
   • CHuman oversight
   • DFail-safe
   Reveal Solution Discussion Select an answer first