Responsible and Ethical AI

Question 1

A data scientist is training a credit approval model and suspects the historical dataset contains bias against a specific demographic group. Which AWS tool should they use to generate bias metrics (like 'Class Imbalance' or 'Difference in Proportions') prior to training?

Accepted Answer

Amazon SageMaker Clarify

Answer

Amazon SageMaker Model Monitor

Answer

Amazon Augmented AI (A2I)

Answer

AWS Audit Manager

Question 2

A healthcare company wants to use a generative AI model to summarize medical records. To comply with HIPAA, they must ensure no Personally Identifiable Information (PII) is visible in the generated summaries or leaked to the model. Which Bedrock solution should they implement?

Accepted Answer

Guardrails for Amazon Bedrock

Answer

Watermarking

Answer

Knowledge Bases

Answer

Amazon Macie

Question 3

Which 'Model Explainability' technique does SageMaker Clarify use to assign a contribution score to each input feature, helping to understand which factors most influenced a specific prediction?

Accepted Answer

SHAP values (SHapley Additive exPlanations)

Answer

Confusion Matrix

Answer

ROC-AUC Curve

Answer

Gradient Descent

Question 4

A financial institution uses a model to approve loans. Occasionally, the model rejects applications with low confidence (e.g., 55%). To ensure fairness and correct errors, they want these edge cases to be manually reviewed by human experts. Which service facilitates this flow?

Accepted Answer

Amazon Augmented AI (A2I)

Answer

SageMaker Ground Truth

Answer

SageMaker Model Registry

Answer

AWS Trusted Advisor

Question 5

An AI developer is concerned about 'Prompt Injection' attacks where malicious users try to manipulate the LLM to reveal system instructions or generate prohibited content. What is the best defense-in-depth using managed services?

Accepted Answer

Implement Guardrails for Amazon Bedrock with content filters and prompt attack detection

Answer

Increase model temperature to vary responses

Answer

Use only small models to limit their knowledge

Answer

Encrypt the prompt with KMS before sending it

Question 6

To comply with governance regulations, a company needs to maintain an immutable record of who requested a model deployment, its bias test results, and its final approval. Which tool centralizes this information?

Accepted Answer

Amazon SageMaker Model Cards and Model Registry

Answer

Amazon CloudWatch Logs

Answer

AWS CloudTrail

Answer

Amazon Inspector

Question 7

An AI application generates images. To prevent the spread of misinformation, the company wants to add an invisible mark identifying the images as 'AI-generated'. Which feature should they enable?

Accepted Answer

Watermarking in Amazon Bedrock

Answer

Server-side encryption (SSE-S3)

Answer

Digital signatures with AWS Signer

Answer

Amazon Rekognition Content Moderation

Question 8

Which post-training bias metric in SageMaker Clarify indicates if the model rejects loans at a different rate for one demographic group compared to another, despite having similar qualification rates?

Accepted Answer

Disparate Impact (DI)

Answer

Accuracy

Answer

Mean Squared Error (MSE)

Answer

Latency

Question 9

You are designing a RAG (Retrieval Augmented Generation) system. Which 'Responsible AI' risk does using RAG primarily mitigate compared to using an LLM directly?

Accepted Answer

Hallucinations and lack of truthfulness

Answer

High latency

Answer

Inference costs

Answer

Bias in the base model's training data

Question 10

A company needs to continuously audit its AI service usage on AWS to verify GDPR compliance and generate automated reports for auditors. Which service simplifies this evidence collection?

Accepted Answer

AWS Audit Manager

Answer

Amazon GuardDuty

Answer

AWS Artifact

Answer

Amazon Detective