Guidelines for Responsible AI
Free AWS Certified AI Practitioner (AIF-C01) practice — 6 questions on Guidelines for Responsible AI, with explanations. No sign-up. Full 12-question mixed test →
A financial services company is deploying an AI system to assess loan applications. During model validation, the team discovers that while overall accuracy is 94%, the approval rate for applicants from historically marginalized communities is 23% lower than other groups with similar credit profiles. The AI Safety team recommends either retraining with balanced data or implementing post-processing adjustments to equalize approval rates. The business team argues this will reduce profitability. Which action BEST aligns with AWS responsible AI guidelines?
AWS responsible AI guidelines prioritize fairness as a core pillar. When disparate impact is detected (23% difference in approval rates for similar profiles), active mitigation is required. Demographic parity constraints during retraining directly address the fairness gap. A slight accuracy trade-off is acceptable when it prevents discriminatory outcomes. This approach is proactive and addresses root cause rather than symptoms.
A healthcare AI startup is building a diagnostic assistant using Amazon Bedrock. The model will suggest potential diagnoses based on patient symptoms. During testing, a physician notices the model occasionally hallucinates rare disease names that don't exist and sometimes contradicts itself when asked to explain reasoning. Regulatory approval requires demonstrating model reliability and appropriate safeguards. Which approach BEST addresses responsible AI requirements for this high-stakes medical use case?
This option addresses both technical mitigation (RAG grounds responses in verified factual sources, directly reducing hallucinations) and process safeguards (human-in-the-loop ensures physician oversight). For high-stakes healthcare AI, responsible AI principles require layered defenses: technical controls to improve reliability AND human oversight to catch remaining errors. RAG anchors the model to authoritative medical knowledge, while mandatory physician verification ensures no AI suggestion reaches patients without expert review. This combination meets regulatory requirements for demonstrable reliability and appropriate risk mitigation in safety-critical applications.
An e-commerce company uses Amazon Personalize to recommend products. A data scientist discovers that the recommendation model creates a feedback loop: it recommends popular items, which get more clicks, which makes them more popular, while niche products rarely appear. The business is satisfied because conversion rates are high. Three months later, the catalog diversity has decreased by 40% as vendors of niche products leave the platform. Which responsible AI principle was violated, and what should have been implemented?
This scenario describes a classic filter bubble and feedback loop problem, which falls under robustness and safety. The model creates a self-reinforcing system that reduces diversity and long-term platform health. Exploration strategies (like epsilon-greedy or Thompson sampling) intentionally inject diversity to prevent feedback loops. This addresses the root technical cause. While monitoring (option C) is important, the question asks for the violated principle AND what should have been implemented as a solution—exploration is the preventive technical control.
A global retailer is implementing an AI-powered dynamic pricing system using Amazon SageMaker. The model adjusts prices based on demand, inventory, competitor pricing, and customer browsing behavior. During a regulatory audit, the team is asked to explain why Customer A was charged $149 for a product while Customer B was charged $127 for the same item at the same time. The data science team cannot provide a clear explanation because the model uses a complex ensemble of neural networks. Which responsible AI requirement is at risk, and what is the appropriate remediation?
Transparency requires providing meaningful information about AI decisions, especially when they affect individuals differently. SageMaker Clarify provides post-hoc explanations using SHAP values or other techniques, which allow explaining complex models without sacrificing performance. For regulated industries, maintaining explanation logs creates an audit trail. This balances business needs (complex models for better pricing) with responsible AI requirements (explainable decisions).
A social media company uses Amazon Rekognition to detect and remove harmful content. The AI system flags content for review, which is then evaluated by human moderators. An internal audit reveals that the AI system has a 12% false positive rate for content from users in Southeast Asian countries (incorrectly flagging benign content as harmful) compared to 3% for users in North America, due to underrepresentation of Southeast Asian cultural contexts in training data. Affected users experience account restrictions. The company has limited resources to retrain the model immediately. Which approach BEST demonstrates responsible AI while addressing the immediate harm?
This approach addresses immediate harm (excessive false positives causing account restrictions) while maintaining platform safety. Prioritizing human review for affected populations provides a near-term mitigation while model improvements are developed. Cultural context training addresses the root cause (misunderstanding of cultural norms). This balances safety, fairness, and practicality. It doesn't lower safety standards but adds human judgment where the AI is demonstrably less reliable.
A pharmaceutical company is developing an AI system using Amazon Bedrock to accelerate drug discovery by predicting molecular interactions. The model will identify promising compounds for further laboratory testing. During deployment planning, the AI ethics board raises concerns about how to handle situations where the model's predictions might be influenced by biases in historical research data (certain disease categories or patient populations were historically underrepresented in clinical trials). The model will be used by research scientists, not directly by patients. Which responsible AI consideration is MOST critical for this use case?
For an AI system used by expert scientists (not direct patient care), transparency through comprehensive documentation is most critical. Model cards that explicitly document known biases, limitations, and appropriate use cases empower researchers to interpret results correctly and compensate for known limitations. Scientists can then make informed decisions about which predictions to pursue and which require additional validation. This enables responsible use without unnecessarily restricting potentially valuable research directions.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$54.99 $17.99 with code FREETEST33 — valid through August 22.
Get my $17.99 deal →