This domain is 14% of the scored exam. The questions are usually plain-language: which feature of responsible AI is described, which tool detects bias or adds human review, why a dataset causes unfair output, or what a model card is for. Learn the vocabulary and the tool-to-job mapping.
Dimensions of responsible AI
AWS describes responsible AI through core dimensions. Each can be asked as a definition.
| Dimension | Meaning |
|---|---|
| Fairness | Outputs don’t unjustly favour or harm groups of people |
| Explainability | You can understand and describe why the system gave an output |
| Privacy and security | Data is protected, used with consent, and not leaked |
| Safety | The system doesn’t cause harm to people or society |
| Controllability | Humans can steer, monitor and override the system |
| Veracity and robustness | Outputs are accurate and truthful (veracity), and the system keeps working under unexpected inputs and errors (robustness) |
| Governance | Policies and processes assign accountability and ensure compliance |
| Transparency | Stakeholders can see how and when AI is used and its limits |
| Inclusivity | Designed for, and works well for, diverse people |
Associated terms: bias (systematic skew against a group), trustworthiness, accountability, sustainability.
Tools for responsible AI
Guardrails for Amazon Bedrock
A configurable safety layer applied to both user input and model output, working across models (also callable on text through the standalone ApplyGuardrail API, so it can protect models outside Bedrock).
| Policy | What it does |
|---|---|
| Content filters | Block harmful categories (hate, insults, sexual, violence, misconduct, prompt attacks such as jailbreaks and injection) with adjustable strength |
| Denied topics | Define topics the app must not discuss (for example investment advice) |
| Word filters | Block specific words or profanity |
| Sensitive information filters | Detect PII and block or mask it; custom regex patterns |
| Contextual grounding check | Detects hallucinations by checking that responses are grounded in the source and relevant to the query (useful for RAG) |
| Automated Reasoning checks | Uses formal logic against your policy documents to verify factual claims and reduce hallucination |
Use Guardrails when the question asks to block topics, filter harmful content, redact PII or reduce hallucinations at the application level without retraining.
Other tools
| Tool | Role |
|---|---|
| Amazon SageMaker Clarify | Detects bias in data and in model predictions (before and after training), and provides explainability (feature importance via SHAP); also FM evaluation |
| SageMaker Model Monitor | Watches production models for data drift, model-quality drift, bias drift and feature-attribution drift |
| Amazon Augmented AI (A2I) | Builds human review workflows for low-confidence or sensitive predictions (works with Textract, Rekognition and custom models) |
| SageMaker Ground Truth | Human labeling, including data to improve quality and reduce bias |
| SageMaker Model Cards | Documentation of intended use, training data, evaluation and risks |
| SageMaker Role Manager, Model Dashboard | Governance and visibility |
| AWS AI Service Cards | AWS’s own transparency documents for its AI services |
| Bedrock model evaluation | Tests models for accuracy, robustness and toxicity |
Selecting a model responsibly
- Environmental considerations: training and running big models uses a lot of energy. Smaller, efficient or reused models (and purpose-built chips such as AWS Trainium and Inferentia, plus Regions with cleaner energy) cut the carbon footprint. Choose the smallest model that meets the need.
- Sustainability also means not retraining from scratch when fine-tuning or prompting would do.
- Check the provider’s transparency (documentation, licence, training data origin), the safety record, and whether the model’s bias has been evaluated.
- Prefer models you can evaluate and constrain with guardrails.
Legal and business risks of generative AI
| Risk | Example |
|---|---|
| Intellectual property infringement | Output resembles copyrighted work, or training data was used without rights |
| Biased outputs | Discriminatory hiring or lending text, leading to legal action |
| Loss of customer trust | A harmful or embarrassing answer goes public |
| End-user risk | People act on wrong medical, legal or financial advice |
| Hallucinations | False statements about people or products (defamation, misleading advice) |
| Privacy breaches | PII in prompts or outputs |
| Regulatory non-compliance | Violating data protection or AI laws |
Characteristics of good datasets
| Characteristic | Why it matters |
|---|---|
| Inclusive and diverse | Covers all groups and scenarios the model will meet |
| Balanced | No class or group is badly under-represented |
| Curated sources | Trusted, high-quality, documented origin |
| Accurate, clean labels | Bad labels teach bad behaviour |
| Representative | Matches the real-world population |
| Current | Stale data causes drift |
Bad data produces sampling bias, historical bias (the past was unfair) and label bias.
Effects of bias and variance
| High bias | High variance | |
|---|---|---|
| Fit | Underfitting (too simple) | Overfitting (memorises training data) |
| Result | Inaccurate on all data | Accurate on training data, inaccurate on new data |
Beyond accuracy, bias leads to unequal performance across demographic groups (a face model that works poorly for darker skin tones, a loan model that rejects one group more often). Detect it with subgroup analysis: measure performance separately by group.
Detecting and monitoring bias, trustworthiness and truthfulness
| Method | Use |
|---|---|
| Analyse label quality | Look for inconsistent or prejudiced labels |
| Human audits | People review samples of outputs |
| Subgroup analysis | Compare metrics across groups |
| SageMaker Clarify | Pre-training and post-training bias metrics; feature importance |
| Model Monitor | Continuous production monitoring |
| Amazon A2I | Human-in-the-loop review |
| Grounding checks and citations | Truthfulness of generated content |
Mitigations: more diverse, balanced data, rebalancing or reweighting, bias-aware evaluation before launch, guardrails, and human oversight.
Transparent and explainable models
| Transparent / explainable | Opaque (“black box”) | |
|---|---|---|
| Examples | Linear and logistic regression, decision trees | Deep neural networks, large foundation models |
| Understanding | You can see how inputs lead to outputs | Hard to trace why a decision was made |
| Performance | Often lower on complex tasks | Often higher |
Tradeoff: more complex models often perform better but are harder to interpret. The more a decision affects someone’s life (credit, hiring, health), the more explainability you need. You can also trade some performance for safety and transparency.
Tools and artefacts
- SageMaker Model Cards: a standard record of a model’s purpose, training data, performance and limitations.
- AWS AI Service Cards: transparency docs for AWS AI services.
- Open-source models: weights, licence and often training details are inspectable.
- Data documentation: lineage, provenance and licensing of datasets.
- SageMaker Clarify explanations for individual predictions.
Human-centred design for explainable AI
Design AI around people’s needs: explanations in plain language that match the audience; show confidence and sources; let users question, correct and override results; keep humans in the loop for consequential decisions; involve diverse users in design and testing; give feedback channels.
Common traps
- “Block a topic or mask PII in the chatbot” is Guardrails, not Clarify.
- “Detect bias in a dataset or explain predictions” is Clarify; “detect drift in production” is Model Monitor; “human review of predictions” is A2I; “document a model” is Model Cards.
- Hallucination is a veracity issue; discrimination is a fairness issue.
- A more accurate model isn’t automatically a better choice if it can’t be explained and the decision is high-stakes.
- Training on a bigger dataset doesn’t fix bias if the dataset is unrepresentative.
Exam checklist
- Define fairness, explainability, privacy, safety, controllability, veracity, robustness, governance, transparency.
- Name the Guardrails policy types and what each blocks.
- Match Clarify, Model Monitor, A2I, Ground Truth and Model Cards to jobs.
- Explain bias versus variance and over/underfitting, and subgroup analysis.
- List legal risks of generative AI.
- Describe the explainability versus performance tradeoff and human-centred design.
- Say how model choice affects sustainability.