Explain the development of responsible AI sys..

Production monitoring (Option B) should compare relevant data and model metrics against established baselines and generate alerts when meaningful deviations occur. This can help identify changes in data quality, model performance, or fairness-related metrics so the team can investigate and respond appropriately. Ignoring metrics (Options A and D) or blindly retraining (Option C) does not proactively detect these shifts.

Subgroup analysis (Option C) helps identify disparities in model performance across patient demographic groups. Human-in-the-loop review (Option D) provides human oversight for uncertain or high-impact predictions and can support safer use of the system. Restricting the dataset to a single hospital (Option A) can reduce representativeness and introduce sampling bias. Removing human oversight (Option B) eliminates an important mechanism for reviewing appropriate cases. L2 regularization (Option E) is primarily used to reduce overfitting and is not a dedicated method for detecting or preventing demographic bias.

Option B is correct because generating unauthorized content that reproduces copyrighted material can create intellectual property and copyright risks. Option A describes a bias and fairness concern. Option C describes hallucination and veracity concerns. Option D describes a safety and end-user harm concern. These are important risks, but they are distinct from IP and copyright risk.

Considering environmental impact and resource consumption (Option A) incorporates sustainability into model selection. Choosing a smaller or right-sized model when it satisfies the required performance (Option C) directly reduces computational and energy requirements. Comparing the computational and energy requirements of candidate architectures (Option E) helps balance model effectiveness with environmental impact. Training a model from scratch (Option B) is highly energy-intensive and less sustainable than leveraging pre-trained models. Deploying across all regions (Option D) multiplies the infrastructure footprint and wastes energy, violating AWS sustainability best practices.

Inclusivity ensures that AI systems are designed to be accessible, usable, and effective for a diverse range of users, actively accommodating differences in language, physical ability, or cultural background.

SageMaker Clarify (Option B) provides capabilities for analyzing bias in data and model predictions and for explaining model predictions using feature-attribution techniques such as SHAP. IAM Access Analyzer (Option A) focuses on resource-access permissions, Amazon Macie (Option C) helps discover and protect sensitive data in Amazon S3, and AWS CloudTrail (Option D) records API activity. Note: SageMaker Clarify is no longer open to new customers, so this question should be retained only if the exam bank intentionally tests legacy AWS service knowledge.

Amazon Bedrock Guardrails (Option C) provides configurable safeguards that can detect and filter harmful content in model inputs and outputs. Content filters include categories such as Hate, Insults, Sexual, Violence, and Misconduct, with configurable filter strength. Macie (Option B) focuses on sensitive data discovery in S3, IAM Access Analyzer (Option A) focuses on permissions, and CloudTrail (Option D) records API activity.

Amazon SageMaker Clarify is the correct choice because its primary purpose is to help developers detect potential bias in pre-training data and post-training models, as well as to provide explainability for how models make predictions. Amazon Bedrock Guardrails (Option A) is incorrect because it applies safety filters and topic restrictions to live generative AI text outputs, but it doesn't analyze tabular training data for historical bias. Amazon Rekognition (Option C) is a computer vision service for analyzing images and video, which doesn't fit this tabular data scenario. AWS IAM Access Analyzer (Option D) helps you identify unintended access to your AWS resources and IAM policies, which has nothing to do with machine learning bias or fairness.