AI Failure Modes Quiz

Answers and explanations appear as you take the quiz.

1. A system invents a citation. Why classify this separately from a duplicated payment?
   • All errors should use the same mitigation
   • A citation error is always a network failure
   • The triggers, consequences, and useful controls differ
   • The payment system cannot have model errors
2. A model correctly interprets a request, but a tool acts on the wrong account. What kind of issue may this be?
   • Proof the model’s language output was wrong
   • Evidence the user intended the other account
   • A formatting problem only
   • A system-level failure outside the model’s interpretation
3. Why should a retry use transaction identifiers when an action may already have completed?
   • To help prevent a repeated request from duplicating the external action
   • To make the model’s answer more fluent
   • To eliminate authorization checks
   • To ensure every retry succeeds
4. How should ordinary input variability and deliberate misuse be tested?
   • Only with adversarial prompts
   • As distinct test categories, since they expose different weaknesses
   • Only with clean standard inputs
   • As identical cases requiring one metric
5. A fix passes the exact failed example once. What conclusion is justified?
   • That the entire failure class is eliminated
   • That regression tests are no longer needed
   • That this replay passed; meaningful variations and the underlying cause still need testing
   • That all future inputs will behave the same
6. Which control fits a tool that must not change accounts without authorization?
   • A more polished response template
   • A longer model context window
   • A higher image resolution
   • Permission limits and an explicit authorization check
7. Why retain failed cases in a regression suite?
   • To check whether later changes reintroduce known failures
   • To guarantee the model never changes
   • To replace production monitoring
   • To hide system behavior from operators
8. An endpoint returns HTTP 200 with stale results. What does that show?
   • The request necessarily produced a useful answer
   • Transport success alone does not prove task-level correctness
   • The service has no operational issue
   • The model’s training score must be low