AI Capabilities, Data Dependencies, and Failure Modes CAIPM Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A company wants to use AI to transcribe customer service calls and analyze sentiment. The call recordings are of varying audio quality, and some contain background noise. Which data-related factor is most likely to affect the AI system's performance?

   • AThe length of the calls.
   • BThe quality and clarity of the audio recordings.
   • CThe sentiment analysis algorithm's complexity.
   • DThe number of languages spoken in the calls.
   Reveal Solution Discussion Select an answer first
2. 22foundation · easy

   Which practice is most effective for detecting data drift in a deployed AI model?

   • AContinuous monitoring of model performance and input data
   • BOne-time validation before deployment
   • CUsing more training data from the original period
   • DIncreasing the model's complexity
   Reveal Solution Discussion Select an answer first
4. 23expert · hard

   A government agency uses an AI model to screen visa applications. The model was trained on historical decisions that were later found to contain biases against certain nationalities. The agency must continue using the model but wants to reduce bias without losing too much accuracy. Which approach best balances these competing goals?

   • AApply fairness constraints during training, such as equalized odds, and evaluate the trade-off between bias and accuracy.
   • BRemove all nationality-related features from the model.
   • CSet a lower decision threshold for applicants from historically disadvantaged nationalities.
   • DRetrain the model on the same biased data but with a different algorithm.
   Reveal Solution Discussion Select an answer first
5. 24expert · hard

   A bank's loan approval model is being audited for compliance. The model uses many features, including some that are correlated with protected attributes. The bank wants to reduce bias without significantly reducing overall accuracy. Which approach best balances these goals?

   • AApply a fairness constraint during model training that limits disparate impact while optimizing accuracy.
   • BRemove all features correlated with protected attributes.
   • CUse a post-processing threshold adjustment for different groups.
   • DAdd more data from underrepresented groups to the training set.
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   A credit approval AI model consistently denies loans to applicants from a specific ethnic group, even when their financial profiles are similar to approved applicants. Which failure mode does this exemplify?

   • ABias
   • BModel degradation
   • CData drift
   • DOverfitting
   Reveal Solution Discussion Select an answer first