AIF-C01 Guidelines for Responsible AI — 20 Questions

Question 1

A company wants to ensure accountability for its machine learning models by tracking all changes to the model and its training data. Which AWS feature should they use?

Accepted Answer

AWS CloudTrail. AWS CloudTrail is the correct choice because it records API activity across AWS services, enabling governance, compliance, and operational auditing. For machine learning models, CloudTrail logs all changes to model artifacts, training data, and SageMaker resources by capturing calls to the SageMaker API, such as CreateTrainingJob, UpdateEndpoint, or PutModelPackageGroupPolicy. This provides a complete audit trail for accountability, as required by responsible AI guidelines.

Answer

Amazon CloudWatch

Answer

Amazon SageMaker Experiments

Answer

AWS CodeCommit

Question 2

A security team is concerned about adversarial attacks on their image classification model deployed on Amazon SageMaker. They want to test robustness against carefully crafted inputs that cause misclassification. What approach should they use?

Accepted Answer

Generating adversarial examples using SageMaker Clarify. SageMaker Clarify includes built-in capabilities for generating adversarial examples to evaluate model robustness against inputs designed to cause misclassification. This directly addresses the security team's concern by simulating adversarial attacks, allowing them to measure and improve the model's resilience before deployment.

Answer

Data augmentation on the training set

Answer

A/B testing between two similar models

Answer

SageMaker Model Monitor with adversarial drift

Question 3

Refer to the exhibit. A team is creating an IAM policy for a SageMaker notebook user. The user needs to access training data in an S3 bucket and create models. Which responsible AI concern is most relevant to this policy?

Accepted Answer

The policy grants overly broad permissions, violating the principle of least privilege.. The IAM policy grants the SageMaker notebook user overly broad permissions (e.g., s3:* or sagemaker:*), which violates the principle of least privilege. This is a core responsible AI concern as it increases the risk of unauthorized access or accidental modification of training data and models. The policy should restrict actions to only those necessary for the user's specific role, such as s3:GetObject for specific buckets and sagemaker:CreateModel with resource constraints.

Answer

The policy does not enforce encryption for the notebook.

Answer

The policy does not restrict which S3 buckets the user can read.

Answer

The policy does not include a condition for model explainability.

Question 4

A financial services company deploys a generative AI chatbot for customer support. They want to prevent the chatbot from generating harmful or misleading information. Which AWS service can help monitor and filter responses?

Accepted Answer

Amazon Augmented AI (A2I) with human review. Amazon Augmented AI (A2I) enables human review of machine learning predictions, making it suitable for monitoring and filtering chatbot responses for harmful or misleading content. Other options: Amazon GuardDuty is a threat detection service, not for content filtering. AWS WAF protects web applications from common web exploits. Amazon Comprehend is a natural language processing service for insights, but not designed for real-time filtering with human review.

Answer

Amazon GuardDuty

Answer

AWS WAF

Answer

Amazon Comprehend

Question 5

An e-commerce company uses a recommendation system built with Amazon Personalize. They want to explain to customers why certain products are recommended. Which AWS service can provide model explanations?

Accepted Answer

Amazon SageMaker Clarify. Amazon SageMaker Clarify is the correct choice because it provides model explainability features, including feature importance and SHAP-based explanations, which can be used to interpret why Amazon Personalize recommends specific products. This aligns with the requirement to explain recommendations to customers, as SageMaker Clarify integrates with Personalize to generate human-readable explanations for model predictions.

Answer

AWS Config

Answer

AWS CloudTrail

Answer

Amazon Detective

Question 6

A government agency uses Amazon Rekognition for identity verification. They want to ensure the model is robust against adversarial attacks. What should they do?

Accepted Answer

Train the model with adversarial examples. Training the model with adversarial examples (a technique known as adversarial training) directly improves the model's robustness against adversarial attacks by exposing it to perturbed inputs during training, forcing the model to learn decision boundaries that are less sensitive to small, malicious perturbations. This is a fundamental approach in building resilient AI systems, especially for sensitive tasks like identity verification.

Answer

Use SageMaker Model Monitor to detect adversarial inputs

Answer

Implement input validation and use Amazon SageMaker Shadow Testing

Answer

Apply AWS Shield Advanced

Question 7

Which THREE practices support transparency in AI systems? (Select three.)

Accepted Answer

Provide documentation of model limitations. Option C is correct because documenting a model's limitations—such as known failure modes, bias risks, and out-of-scope use cases—gives stakeholders the information needed to understand what the system can and cannot reliably do, which is a core element of transparency. Option D is correct because publishing model performance metrics (e.g., accuracy, precision/recall, fairness measures across subgroups) lets users and auditors verify how the system behaves rather than taking claims on trust. Option E is correct because allowing users to opt out of data collection gives them visibility and control over how their data feeds the AI system, supporting transparent data-handling practices. Option A is not among the marked correct answers here: while SHAP is a legitimate explainability technique, the question's intended transparency practices are documentation, metric publication, and user opt-out, so it does not belong to this answer set. Option B is also not marked correct because using open-source models exclusively is a licensing/procurement choice, not a transparency practice—proprietary models can still be transparent if their limitations and metrics are disclosed.

Answer

Implement explainability methods like SHAP

Answer

Use open-source models exclusively

Question 8

Refer to the exhibit. An AWS administrator sets up a SageMaker Model Monitor schedule for bias detection. What is the primary purpose of this configuration?

Accepted Answer

To generate weekly bias reports for the deployed endpoint. SageMaker Model Monitor's bias detection feature is specifically designed to detect and report bias in model predictions over time. By configuring a weekly schedule, the administrator ensures that the deployed endpoint's predictions are regularly analyzed for bias, and a weekly bias report is generated. This aligns with the responsible AI goal of monitoring for fairness, not retraining or replacing the model.

Answer

To retrain the model weekly

Answer

To replace the model with a new version

Answer

To monitor data drift in training data

Question 9

A large e-commerce company uses a recommendation system to suggest products to customers. Recently, a data scientist noticed that the model's recommendations for high-value luxury items are predominantly shown to users in affluent zip codes, while users in less affluent areas rarely see these items, even if they have searched for them. The company is concerned about fairness and wants to ensure all customers have equal access to recommendations regardless of location. The current model uses collaborative filtering on historical purchase data. The team needs to modify the system without sacrificing overall recommendation accuracy. Which action best addresses the fairness concern while maintaining performance?

Accepted Answer

Remove zip code and any income-correlated features from the training data. The most effective approach is to ensure the model does not use zip code or any feature correlated with income as a direct or indirect input. This removes the proxy for socioeconomic status. Simply equalizing recommendation frequency artificially may hurt relevance. Personalizing based on search history is already being done but zip code bias remains. Adding more training data may not help if the bias is in the features.

Answer

Randomly show luxury recommendations to a subset of users regardless of their behavior

Answer

Add more training data from less affluent areas to balance the dataset

Answer

Implement a separate recommendation pipeline for luxury items based only on search history

Question 10

A government agency is deploying an AI system to detect fraudulent benefit claims. The system uses a neural network trained on historical claims data, which includes a disproportionate number of false positives from a particular ethnic group due to historical over-policing. The agency must ensure the system does not perpetuate discrimination. They have a rigorous testing procedure but limited budget. The project lead wants to balance fairness with detection performance. Which combination of steps should they prioritize?

Accepted Answer

Rebalance the training data to have equal representation across groups and evaluate using a fairness metric like equal opportunity. Rebalancing the training data to ensure equal representation across groups directly addresses the root cause of the bias: historical over-policing led to disproportionate false positives for a particular ethnic group. Evaluating with a fairness metric like equal opportunity ensures that the true positive rate is similar across groups, which is critical for a fraud detection system where both fairness and detection performance matter. This approach balances fairness and performance without discarding the neural network's capability.

Answer

Remove the race feature from the model and rely on performance metrics alone

Answer

Replace the neural network with a logistic regression model retrained on the same data

Answer

Use a post-hoc explainability tool to identify biased predictions and manually override them

Question 11

A financial services company uses a machine learning model to automatically reject credit card transactions suspected of fraud. The model was trained on transaction data from the past two years. Over the last three months, the model's false positive rate has increased significantly, causing legitimate transactions to be declined and leading to customer complaints. The company needs to restore the model's accuracy quickly. Initial analysis shows that the distribution of transaction amounts and locations has shifted compared to the training period. The data science team is under pressure to deploy an update within a week. Which approach should they take to most effectively address the issue while adhering to responsible AI guidelines?

Accepted Answer

Retrain the model using only the most recent three months of transaction data and evaluate on current distribution. Retraining the model using only the most recent three months of transaction data and evaluating on the current distribution directly addresses the data drift identified in the scenario. This approach restores accuracy by aligning the training data with the current distribution, and evaluating on current data ensures the model performs well on the live population, adhering to responsible AI principles of transparency and fairness.

Answer

Deploy a rule-based system with fixed rules for fraud detection

Answer

Adjust the decision threshold to reduce false positives without retraining

Answer

Build an ensemble model that combines predictions from the old model and a new model trained on recent data

Question 12

Refer to the exhibit. A data scientist used SageMaker Clarify to evaluate bias in a binary classification model predicting loan approval. The exhibit shows bias metrics for the female facet. What does the analysis indicate about the model's impact on the female group?

Accepted Answer

The model has a post-training accuracy difference indicating a negative bias against the female group.. The exhibit shows a post-training accuracy difference metric (e.g., difference in positive predicted values or accuracy) that is negative for the female facet, indicating the model's predictions are less accurate for the female group compared to the overall or male group. This negative difference signifies a bias against the female group, as the model performs worse for them after training. SageMaker Clarify computes such metrics to detect disparate impact, and a negative value here directly points to adverse treatment.

Answer

The metrics are within acceptable thresholds, so no action is needed.

Answer

The model shows a high positive bias toward the female group.

Answer

The model exhibits a pre-training class imbalance but no post-training bias.

Question 13

A healthcare company uses Amazon SageMaker to train a model that predicts patient readmission risk based on electronic health records (EHRs) stored in Amazon HealthLake. The training dataset contains 2 million records from the past three years, with a significant gender imbalance: 70% male and 30% female. The model achieved high overall accuracy, but further analysis using SageMaker Clarify revealed that the precision for female patients is 0.65 while for male patients it is 0.88. Additionally, the model's false positive rate for female patients is significantly higher. The company must comply with healthcare regulations that require fairness and non-discrimination. The data science team has already used SageMaker Data Wrangler for initial preprocessing and SageMaker Clarify for bias detection. They need to take immediate action to mitigate the bias before deploying to production. Which course of action should the team take?

Accepted Answer

Use SageMaker Clarify's bias mitigation feature to apply reweighing techniques and retrain the model with adjusted sample weights.. To use SageMaker Clarify's built-in bias mitigation technique (reweighing) as it directly addresses the disparity in precision and false positive rates between male and female patients by adjusting sample weights during training. SageMaker Clarify's bias mitigation feature is designed for this purpose. Option B: Generating SHAP values explains predictions but does not mitigate bias. Option C: SMOTE addresses class imbalance but not fairness in terms of group accuracy disparity; it may even worsen bias by oversampling without considering fairness. Option D: SageMaker Model Monitor is for detecting drift in production data, not for mitigating bias during training. Therefore, Option A is the best choice.

Answer

Use SageMaker Clarify to generate SHAP values and adjust the model's feature importance by removing biased features.

Answer

Use SMOTE (Synthetic Minority Oversampling Technique) to balance the training dataset before retraining.

Answer

Use SageMaker Model Monitor to detect feature drift and automatically retrain the model with updated data.

Question 14

A media company deploys an Amazon Bedrock text-generation assistant that writes marketing copy. Legal requires the assistant to refuse requests for defamatory content and to avoid generating copyrighted passages, and it must record who submitted each prompt for audit. Which TWO AWS capabilities should the team implement to meet these responsible AI requirements? (Choose two.)

Accepted Answer

Enable AWS CloudTrail to log Bedrock API calls and use the logs to capture the identity of each prompt submitter. The requirements combine content-policy enforcement and auditability. Bedrock Guardrails applies denied-topic and content policies at inference time so the assistant refuses prohibited requests, while CloudTrail captures the calling identity for every Bedrock invocation, providing the audit record. Capacity, metrics, and sampling settings do not enforce content rules or attribute prompts to submitters.

Answer

Create an Amazon CloudWatch dashboard that charts Bedrock invocation counts and latency percentiles

Answer

Enable provisioned throughput for the foundation model to guarantee consistent capacity during peak hours

Answer

Increase the model's temperature setting so the assistant produces more varied and creative marketing copy

Question 15

A financial services company in Europe is deploying a generative AI assistant that explains loan denials to applicants. Regulatory counsel requires that every applicant receive a plain-language reason for the decision, and that the company be able to demonstrate which model version produced each reason. The team already uses Amazon Bedrock with a foundation model and Amazon CloudWatch for monitoring. Which combination of AWS capabilities best satisfies these two requirements?

Accepted Answer

Use Amazon Bedrock model invocation logging for traceability, and design the assistant to generate explanations grounded in the input features using retrieval augmented generation from a curated decision-reason knowledge base.. The two regulatory needs are per-decision explainability and model-version traceability. Invocation logging in Amazon Bedrock records the model ID and the exact prompt and response for each call, which satisfies the audit trail. Retrieval augmented generation from a curated reason knowledge base forces the explanation to be grounded in the applicant's actual decision factors, producing plain-language justifications rather than free-form model output.

Answer

Use AWS CloudTrail data events on the Amazon Bedrock endpoint and Amazon Comprehend sentiment analysis on the generated text to confirm explanations are appropriately worded.

Answer

Enable model invocation logging in Amazon Bedrock to capture requests and responses, and use Amazon Bedrock Guardrails to enforce a denied-topics policy on the assistant.

Answer

Deploy the foundation model on Amazon SageMaker and use SageMaker Model Monitor to detect explanation drift, then publish a model card describing the assistant's intended use.

Question 16

A media company uses Amazon Bedrock to power a customer-facing content summarization tool. Legal has flagged that the tool must not reproduce copyrighted song lyrics or generate defamatory statements about public figures, and the security team wants an auditable record whenever the tool blocks a request. The team is configuring Amazon Bedrock Guardrails. Which TWO configurations should they apply to meet these requirements? (Choose two.)

Accepted Answer

Create a denied topics policy that blocks requests seeking to reproduce copyrighted lyrics, and enable Guardrails logging so blocked invocations are recorded for audit.. The requirements are content prohibition for copyrighted lyrics, suppression of unsupported defamatory claims, and an auditable record of blocks. A denied topics policy blocks lyric-reproduction requests, and contextual grounding checks suppress responses not supported by source content, which is the practical control against fabricated statements. Guardrail logging captures each intervention for the audit trail, and SNS notification operationalizes awareness of blocks.

Answer

Apply a sensitive information filter configured for PERSONALLY IDENTIFIABLE INFORMATION and set the guardrail to redact, then store redacted outputs in Amazon S3 for compliance review.

Answer

Configure automated reasoning checks with a formal policy that encodes defamation law, and set the guardrail action to NONE so responses are annotated rather than blocked.

Answer

Add a word filter with a custom list of public figures' names, and set the profanity filter strength to HIGH to suppress defamatory language.

Question 17

A financial services company uses Amazon Rekognition to verify customer identities. To ensure responsible AI practices, which measure should the company prioritize?

Accepted Answer

Regularly audit the model for demographic bias. Regularly auditing the model for demographic bias is a core responsible AI practice, especially for identity verification systems where biased outcomes could lead to unfair treatment of certain customer groups. Amazon Rekognition's facial analysis and comparison features must be tested across diverse demographics to ensure equitable performance, as bias can arise from imbalanced training data or algorithmic artifacts.

Answer

Use only black-box models to protect intellectual property

Answer

Increase model complexity to improve accuracy

Answer

Minimize the amount of training data collected

Question 18

A healthcare startup deploys a model to predict patient readmission risk using Amazon SageMaker. After deployment, the model shows higher false-positive rates for a specific age group. What is the most responsible first step?

Accepted Answer

Use Amazon SageMaker Clarify to detect bias in predictions. Amazon SageMaker Clarify is purpose-built for detecting bias in ML models and data. It provides bias metrics (e.g., Difference in Positive Proportions in Predicted Labels, Disparate Impact) that can quantify whether the model's predictions are systematically skewed against a specific age group. This is the most responsible first step because it objectively measures the bias before any corrective action is taken.

Answer

Increase the prediction threshold for the affected group

Answer

Retrain the model with more data from the affected group

Answer

Immediately retire the model to prevent harm

Question 19

A company uses an AI system to automate loan approvals. The model uses demographic features and achieves high accuracy, but the company wants to ensure compliance with responsible AI guidelines. Which practice best balances performance and fairness?

Accepted Answer

Remove sensitive attributes and monitor for proxy bias. Removing sensitive attributes (e.g., race, gender) from the training data directly addresses fairness by preventing the model from explicitly using these features. However, simply removing them is insufficient; monitoring for proxy bias (e.g., zip code or income correlating with race) is critical to ensure the model does not inadvertently learn discriminatory patterns through correlated features. This approach balances performance by retaining predictive power from non-sensitive features while actively auditing for fairness violations.

Answer

Use demographic features but with minimal monitoring

Answer

Use a complex black-box model and rely on post-hoc explanations

Answer

Optimize the model solely for accuracy on historical data

Question 20

A retail company uses a recommendation system that occasionally suggests inappropriate products to minors. Which responsible AI practice should be applied?

Accepted Answer

Implement human review of flagged recommendations. The correct practice is to implement human review of flagged recommendations. This aligns with the responsible AI principle of accountability, where automated systems must have oversight mechanisms to catch and correct inappropriate outputs, especially when minors are involved. Human-in-the-loop (HITL) validation ensures that edge cases or subtle context (e.g., age-inappropriate product suggestions) are caught before they reach end users, rather than relying solely on automated filters or feedback loops.

Answer

Rely solely on user feedback to improve

Answer

Disable the recommendation system entirely

Answer

Increase the volume of training data