Guidelines for Responsible AI — AWS Certified AI Practitioner AIF-C01 practice questions

Which technique supports model explainability (per-prediction feature importance)?

1. SHAP values delivered by SageMaker Clarify
2. SigV4 request signing on the endpoint
3. Multi-AZ deployment for high availability
4. AWS Backup snapshots of the model registry

Answer: A — SHAP values delivered by SageMaker Clarify

A) Correct — SHAP attributes each feature's contribution to a prediction. B) Auth signing. C) Availability, not explainability. D) Backup ≠ explainability.

Bias in ML systems can arise from which source?

1. The choice of programming language
2. Historical bias baked into training data
3. The AWS region hosting the training job
4. The number of parallel training instances

Answer: B — Historical bias baked into training data

A) Language-agnostic. B) Correct — biased labels/sampling propagate to the model. C) Region choice is unrelated. D) Compute count is unrelated.

Which AWS service BEST supports detecting harmful content (hate, insults, violence) in GenAI apps?

1. Amazon Bedrock Guardrails content filters
2. AWS Shield Advanced DDoS mitigation
3. AWS Config compliance rules
4. Amazon EventBridge routing rules

Answer: A — Amazon Bedrock Guardrails content filters

A) Correct — Guardrails category filters (hate, insults, sexual, violence, misconduct) at prompt and response. B) DDoS defense, unrelated. C) Config drift, unrelated. D) Event routing, unrelated.

Which is a 'half-right' responsible-AI answer that misses a key pillar?

1. The model is fair, explainable, transparent, and safely governed
2. The model is fair and accurate, so no oversight is required
3. The team documents fairness, explainability, safety, and governance
4. The model is monitored for drift, bias, and misuse continuously

Answer: B — The model is fair and accurate, so no oversight is required

A) Full coverage. B) Correct — fair + accurate is half-right, but 'no oversight' violates the human-in-the-loop pillar. C) Full documentation. D) Ongoing monitoring in place.

Which is the BEST response to 'we don't need human review because the model is highly accurate'?

1. High accuracy means Guardrails can be safely disabled
2. High accuracy on aggregate can still mask subgroup harms and edge cases
3. High accuracy removes the need for any evaluation dataset
4. High accuracy is the only responsible-AI metric that matters

Answer: B — High accuracy on aggregate can still mask subgroup harms and edge cases

A) Guardrails serve orthogonal safety needs. B) Correct — subgroup performance and edge cases still warrant human oversight. C) Evaluation still matters. D) Ignores fairness / safety / transparency.

Which BEST captures AWS's responsible-AI dimensions?

1. Only fairness and bias, since those are the primary risks
2. Fairness, explainability, privacy/security, safety, transparency, governance
3. Only cost, latency, throughput, and availability metrics
4. Only accuracy on the training set and speed at inference

Answer: B — Fairness, explainability, privacy/security, safety, transparency, governance

A) Too narrow. B) Correct — the AWS responsible-AI pillars. C) Ops metrics, not RAI. D) Ignores fairness/safety.

Which AWS feature helps document a model's intended use, training data, and performance?

1. Amazon Bedrock Guardrails filters only
2. AWS Trusted Advisor cost checks
3. SageMaker Model Cards for governance
4. Amazon CloudWatch Logs Insights queries

Answer: C — SageMaker Model Cards for governance

A) Guardrails are runtime filters, not documentation. B) Cost/perf recs, unrelated. C) Correct — Model Cards capture intended use, training data, and metrics for governance. D) Log analytics.

Which AWS capability detects bias in datasets and ML models?

1. Amazon Macie PII discovery
2. AWS Config compliance rules
3. Amazon CloudWatch metrics
4. SageMaker Clarify bias reports
5. AWS CloudTrail audit trail

Answer: D — SageMaker Clarify bias reports

A) PII discovery in S3. B) Resource-config compliance. C) Metrics/logs monitoring. D) Correct — bias detection + explainability (SHAP) for ML. E) Account-level API auditing.

Why is human oversight important in production AI systems?

1. It shortens the model's training time significantly
2. It removes the need for evaluation datasets entirely
3. It catches errors and provides accountability for decisions
4. It automatically reduces the cost of GPU inference
5. It eliminates the need for any content filtering

Answer: C — It catches errors and provides accountability for decisions

A) Not related to training time. B) Human review supplements, doesn't replace, offline metrics. C) Correct — humans-in-the-loop for high-stakes decisions, appeals, and audit. D) Unrelated. E) Contradicts responsible-AI practice.

Bias detection in ML?

Use SageMaker Clarify. Check across protected attributes (age, gender, race).

Explainability?

Understanding why model made decision. SHAP, LIME, SageMaker Clarify.

SageMaker Clarify?

Detects bias in data and models. Explainability reports.

Why explainability matters?

Trust, debugging, compliance (e.g., GDPR right to explanation), fairness audits.