AI Governance and Responsible AI

Question 45

A bank uses Amazon SageMaker AI to classify credit applications for additional review. The model is tracked in SageMaker Model Registry, and training data is stored in Amazon S3. Overall accuracy is 92%, but SageMaker Clarify reports a large Difference in Positive Proportions between demographic groups. Which action should the governance team take before expanding deployment?

A. Approve expansion because 92% overall accuracy demonstrates that the model is reliable.

B. Pause expansion, investigate training and label data, evaluate mitigation options with SageMaker Clarify, and document the decision before approving a new model version.

C. Encrypt the training data with AWS KMS and consider the fairness issue resolved.

D. Deploy broadly because the model only recommends additional review and does not make the final credit decision.

Question 46

An insurer uses Amazon SageMaker AI to prioritize claims for adjuster review. Regulators and customers require explanations for the prioritization process. The insurer stores model artifacts in Amazon S3 and monitors services with Amazon CloudWatch. Which actions should the governance team select? Select three options.

A. Publish only the aggregate model-accuracy score to customers and regulators.

B. Use SageMaker Clarify feature-attribution analysis to support explanations of individual prioritization outcomes.

C. Maintain tamper-evident audit and decision logs using application logging and Amazon CloudWatch.

D. Create SageMaker Model Cards describing intended use, limitations, training-data context, and known risks.

E. Use AWS IAM policies as a substitute for model explainability.

Question 47

A hospital uses Amazon Bedrock to draft appointment instructions from patient records. Occasional hallucinations could cause patient harm. The architecture uses AWS Lambda for workflow orchestration and Amazon CloudWatch for operational logging. Which design is most appropriate?

A. Add an AI-generated disclaimer to every response and allow the assistant to deliver all instructions automatically.

B. Deliver all responses immediately and have clinical staff review a weekly sample.

C. Require manual approval for every response, including low-risk routine instructions, without using risk-based routing.

D. Apply Guardrails for Amazon Bedrock, define confidence and policy-sensitive escalation criteria, route qualifying outputs to qualified staff, and record escalation outcomes in Amazon CloudWatch.

Question 48

A retailer is planning a customer-support assistant with Amazon Bedrock. The design will use approved content in Amazon S3, access controls in AWS IAM, and evaluation support from Amazon SageMaker AI. The product team wants governance by design rather than controls added after launch. Which activities should be completed during planning? Select four options.

A. Wait for post-launch complaints and conduct the first governance review after production deployment.

B. Identify personal data and establish handling, retention, and access requirements.

C. Create representative ...