Safety and AI CT-AI Practice Questions (Page 2)

Questions 6–10

1. 6foundation · easy

   What is a potential safety risk when an AI system is deployed in an environment that differs significantly from its training data?

   • AThe system will become more accurate because it encounters new data.
   • BThe system may exhibit unintended behavior due to poor generalization.
   • CThe system will require more computational resources to process the new data.
   • DThe system will automatically shut down to prevent harm.
   Reveal Solution Discussion Select an answer first
2. 7expert · hard

   An AI system is used to control a chemical plant. The system has been trained on historical data and is generally reliable, but there is a concern about rare, unforeseen events. The safety team is considering formal verification. Which statement best describes the role of formal verification in this context?

   • AFormal verification can be used to verify the safety of the system's control logic, but it may not be feasible for the machine learning components.
   • BFormal verification can prove that the system will never cause an accident, so it should be the primary safety assurance technique.
   • CFormal verification is not useful for AI systems because they are inherently unpredictable.
   • DFormal verification should be replaced by extensive testing, which is more practical.
   Reveal Solution Discussion Select an answer first
4. 8expert · hard

   An AI system is used to triage patients in an emergency department. The system is designed to be safe, but there is a trade-off between sensitivity (catching all critical cases) and specificity (avoiding unnecessary alarms). The hospital wants to set safety requirements. Which approach is most appropriate?

   • ASet sensitivity and specificity to 50% each to balance the two.
   • BSet specificity to 100% and sensitivity to 0% to avoid unnecessary alarms.
   • CSet a minimum sensitivity and specificity that are clinically acceptable, and monitor the system's performance in practice.
   • DSet sensitivity to 100% and specificity to 0% to ensure no critical case is missed.
   Reveal Solution Discussion Select an answer first
5. 9application · medium

   An AI system is used to control a power grid. The system must be safe, but it also needs to be available and responsive. The safety team is considering using redundancy. Which approach to redundancy is most appropriate?

   • AUse a single AI model with a manual override switch.
   • BUse two different AI models with different architectures and training data, and cross-check their outputs.
   • CUse two identical AI models running in parallel and compare their outputs.
   • DUse a single AI model with a backup power supply.
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   A self-driving car company is testing its AI system in a simulated environment. The system sometimes fails to stop for pedestrians when the pedestrian is partially obscured by a vehicle. This is an example of which type of safety risk?

   • ALack of robustness to sensor noise.
   • BAdversarial attack.
   • CSystem failure due to hardware malfunction.
   • DUnintended behavior due to limitations in the training data.
   Reveal Solution Discussion Select an answer first