Human-AI Collaboration Quiz

Answers and explanations appear as you take the quiz.

1. How does the guide suggest dividing work between an AI assistant and a human reviewer?
   • Use the system for suitable repetitive support and people for contextual, accountable judgments
   • Make every decision automatic once the model reaches an average accuracy target
   • Assign people only the work for which the model has no training examples
   • Let the assistant approve its own proposals while people see a later report
2. Why is a required human approval click insufficient as oversight by itself?
   • It always slows the service more than any safety benefit can justify
   • It prevents the assistant from presenting any supporting evidence
   • The reviewer may lack the time, context or authority to evaluate the proposal
   • It automatically transfers responsibility for the model to the reviewer
3. What information belongs in a handoff before a person approves an AI-proposed action?
   • Only the model's final recommendation and a confirmation button
   • Relevant source, uncertainty, consequences and meaningful alternatives
   • Only a long explanation of model architecture, regardless of the task
   • Only the historical average accuracy of the model
4. Why should the interface distinguish an AI recommendation from an action already taken?
   • The two states are interchangeable once a model is trained
   • A recommendation automatically has the same external consequences as execution
   • The distinction matters only for the model developer's logs
   • A reviewer needs to know whether approval can still prevent the external action
5. When should cancellation and escalation be available in an AI-supported workflow?
   • Only in a retrospective incident report after the action completes
   • At the point when a reviewer can still prevent or correct the consequence
   • Only if the model produces no recommendation at all
   • Only to administrators who are not involved in the decision
6. A suggestion model is accurate in isolation, but reviewers miss its rare errors. What should evaluation measure?
   • Combined task quality, reviewer burden and error recovery with realistic users
   • Only the model's offline accuracy on a held-out dataset
   • Only how often reviewers click approve
   • Only the time taken to generate each suggestion
7. An assistant proposes a customer refund. What should its reviewer see before approval?
   • Only the assistant's confidence score and proposed amount
   • Only the final refund button, with source records on another system
   • Only the customer's latest message without the governing policy
   • The request, applicable policy passage, amount and proposed action together
8. What should the organization preserve if automation fails or a task falls outside evaluated conditions?
   • A default approval rule so work never stops
   • The model's prediction without any explanation or owner
   • A usable manual path and an owner who can act on reported errors
   • A hidden queue that is reviewed only at the next model update