Task Statement 4.1: Explain the Development of AI Systems That Are Responsible. AIF-C01 Practice Questions (Page 4)

Questions 16–20

1. 16application · medium

   A data science team has trained a model to predict customer churn. The model achieves 99% accuracy on the training data but only 82% accuracy on the validation data. The team suspects the model has memorized noise in the training set. Which pair of concepts best describes this situation, and which tool should the team use to investigate it?

   • AUnderfitting and bias; use Amazon SageMaker Model Monitor to track drift in the input data distribution
   • BUnderfitting and variance; use Amazon SageMaker Clarify to generate a bias report for the training data
   • COverfitting and variance; use Amazon SageMaker Clarify to analyze feature importance and detect data bias
   • DOverfitting and bias; use Amazon A2I to have humans review the model's predictions on the training data
   Reveal Solution Discussion Select an answer first
2. 17foundation · easy

   A company is selecting a large language model for a new application. As part of responsible model selection, which factor should they consider regarding environmental impact?

   • AThe model's accuracy on a benchmark dataset
   • BThe model's license type
   • CThe model's carbon footprint during training and inference
   • DThe model's inference latency
   Reveal Solution Discussion Select an answer first
4. 18foundation · easy

   A developer wants to use Amazon Bedrock Guardrails to prevent a generative AI application from producing toxic or harmful content. Which type of guardrail policy should they configure?

   • AContextual grounding checks
   • BSensitive information filters
   • CWord filters
   • DContent filters
   Reveal Solution Discussion Select an answer first
5. 19foundation · easy

   A company is deploying a generative AI system that creates marketing copy. Which legal risk is most directly associated with the model generating text that closely resembles a copyrighted book?

   • AHallucinations
   • BLoss of customer trust
   • CEnd-user risk
   • DIntellectual property infringement
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A pharmaceutical company is using a generative AI model to assist researchers in summarizing clinical trial results. The company is concerned about two risks: (1) the model may generate plausible but incorrect summaries of trial data (hallucination), and (2) the model may reproduce proprietary information from competitor trials that were included in the training data (IP infringement). The company wants to deploy the model with mitigations for both risks. Which combination of mitigations is most appropriate?

   • AImplement a human review workflow with Amazon A2I for all generated summaries, and configure Amazon Bedrock Guardrails to filter content that resembles known proprietary text
   • BRetrain the model on a curated dataset that excludes competitor trial data, and rely on the model's training to avoid hallucinations
   • CUse Amazon SageMaker Clarify to detect bias in the training data and Amazon SageMaker Model Monitor to track the model's performance in production
   • DConfigure Amazon Bedrock Guardrails with a high confidence threshold for content filters to block all potentially incorrect or proprietary content
   Reveal Solution Discussion Select an answer first