SaveMyCert
Log in
5 of 5 free questions left today·for 10 a day
AIF-C01 · Domain 4

Guidelines for Responsible AI practice questions

Guidelines for Responsible AI is worth 14% of the AIF-C01 exam — the 4th-heaviest of the 5 domains. Responsible development of AI systems, and the importance of transparent and explainable models. 6 fully worked examples are further down this page, answers included.

Exam weight
14%
the 4th-heaviest of the 5 domains
Questions
58
across 2 topics
Free, no account
5/day
a free account allows 10 a day
Explanations
Every option
right and wrong

Build a practice session

5 free questions left today.

Domains

How many?

5 free questions left today.

Mode

Ready when you are

10 fresh questions drawn across 1 of 5 domains, in Learn mode.

Review what you missed

Every question you answer incorrectly, and every question you flag while practising, is saved here automatically. Finish a session and you can come back to re-drill just those.

6 sample Guidelines for Responsible AI questions, fully explained

Questions from the AIF-C01 bank mapped to domain 4, with the answer key and the reasoning behind every option. None of them repeat the examples on the main AIF-C01 practice page.

Question 1Guidelines for Responsible AI

In AWS responsible-AI terminology, what is the difference between bias and fairness?

Choose one.

  • a
    Bias is a systematic error shared equally by the groups; fairness is equal accuracy on the training records.

    An error shared equally by the groups is not what bias means here; bias skews results for some groups, and fairness means equitable treatment of people, not equal accuracy on the training records.

  • b
    Bias is a systematic error on noisy or adversarial inputs; fairness is steady output under those inputs.

    Reliability under noisy or adversarial input describes robustness, not bias or fairness.

  • c
    Bias is a systematic error of stating false facts; fairness is keeping outputs grounded in sources.

    Stating falsehoods as fact is a veracity (truthfulness) failure, and grounding outputs is about veracity, not fairness.

  • d
    Bias is a systematic error that skews results for some groups; fairness is equitable treatment of groups. Correct

    Bias is the flaw (a systematic error that unfairly favors or disfavors certain groups) and fairness is the goal of treating individuals and groups equitably.

The concept

Bias is the flaw; fairness is the goal reached by removing it.

Why that’s the answer

AWS describes bias as a systematic error that makes a model favor or disfavor certain groups, usually learned from skewed data, and fairness as equitable treatment of individuals and groups. An error shared equally by the groups is not group bias, and accuracy on the training records says nothing about equitable treatment. The other two definitions describe robustness and veracity, which are separate responsible-AI features.

How to reason it out
  1. Recall which term names a problem and which names an outcome you want.
  2. Bias is an error that skews results for some groups, not all equally; fairness is equitable treatment.
  3. Reject definitions that actually describe robustness (noisy inputs) or veracity (false statements).

Exam tip: Bias is the flaw, fairness is the goal.

Responsible AI: Bias, Fairness, and Amazon Bedrock Guardrails — the lesson that teaches this.

Question 2Guidelines for Responsible AI

Which responsible-AI feature refers to a system's ability to keep performing reliably when its inputs are noisy, unexpected, or adversarial?

Choose one.

  • a
    Inclusivity

    Inclusivity is about making the system work well for a wide range of people, not about handling bad inputs.

  • b
    Veracity

    Veracity is about factual truthfulness of outputs, not resilience to difficult inputs.

  • c
    Safety

    Safety is about preventing harmful or offensive outputs, not about performing reliably under noisy inputs.

  • d
    Robustness Correct

    Robustness is the ability to keep performing reliably under noisy, unexpected, or adversarial inputs.

The concept

Robustness is reliable performance under noisy, unexpected, or adversarial inputs.

Why that’s the answer

The definition of resilience to difficult inputs matches robustness. Inclusivity concerns serving diverse people, veracity concerns truthfulness, and safety concerns preventing harmful output, so none of them fit reliability under bad inputs.

How to reason it out
  1. Read the definition: reliable performance under noisy or adversarial inputs.
  2. Compare each feature to that definition.
  3. Only robustness matches.

Exam tip: Robustness is reliable performance despite noisy, unexpected, or adversarial inputs.

Responsible AI: Bias, Fairness, and Amazon Bedrock Guardrails — the lesson that teaches this.

Question 3Guidelines for Responsible AI

A company's generative AI customer-service assistant on Amazon Bedrock sometimes repeats a customer's account number and email address back in its responses. Which guardrail policy should the company configure to mask these values?

Choose one.

  • a
    Sensitive information filters Correct

    Sensitive information filters detect personally identifiable information such as account numbers and email addresses in prompts and responses and mask or block it.

  • b
    Denied topics

    Denied topics refuse whole subjects described in natural language (for example investment advice); they do not detect and mask specific personal data values.

  • c
    Contextual grounding checks

    Contextual grounding checks flag responses unsupported by the source material or irrelevant to the query; they do not look for personal data.

  • d
    Word filters

    Word filters block an exact list of words or phrases, such as profanity; account numbers and email addresses vary per customer, so a fixed word list cannot catch them.

The concept

Sensitive information filters in Amazon Bedrock Guardrails detect and mask or block PII.

Why that’s the answer

Account numbers and email addresses are personally identifiable information, and the sensitive information filter is the Guardrails policy that detects PII entities (and custom regex patterns) and masks or blocks them. Word filters only match a fixed list of words, denied topics refuse whole subjects, and contextual grounding checks test whether an answer is supported by its sources.

How to reason it out
  1. Classify the leaked values: account numbers and email addresses are PII.
  2. Ask which policy detects PII entities rather than fixed words or whole subjects.
  3. Choose sensitive information filters, which can mask or block the values.

Exam tip: Leaking personal data from a Bedrock app points to the Guardrails sensitive information (PII) filter.

Responsible AI: Bias, Fairness, and Amazon Bedrock Guardrails — the lesson that teaches this.

Question 4Guidelines for Responsible AI

A bank wants its Amazon Bedrock chatbot to refuse to give investment advice under any circumstances. Which guardrail policy type is designed for this?

Choose one.

  • a
    Denied topics Correct

    Denied topics let you define subjects the application must refuse, such as investment advice.

  • b
    Word filters

    Word filters block specific words and phrases, such as profanity, not an entire subject area.

  • c
    Sensitive information filters

    Sensitive information filters redact PII; they do not refuse a topic like investment advice.

  • d
    Content filters

    Content filters block harmful categories such as hate or violence, but investment advice is not inherently a harmful category.

The concept

Denied topics make a Bedrock application refuse specific subjects.

Why that’s the answer

Refusing an entire subject such as investment advice is exactly what denied topics do. Word filters block individual words, sensitive information filters handle PII, and content filters block harmful categories, none of which target a defined subject to refuse.

How to reason it out
  1. Note the requirement: refuse a whole subject (investment advice).
  2. Match it to the Guardrails policy that defines refused subjects.
  3. Select denied topics.

Exam tip: Use denied topics to make a Bedrock app refuse defined subjects.

Responsible AI: Bias, Fairness, and Amazon Bedrock Guardrails — the lesson that teaches this.

Question 5Guidelines for Responsible AI

A company runs several generative AI applications, each on a different foundation model. It wants one set of configurable safety policies that blocks harmful content in both user prompts and model responses for all of them, without writing custom filtering code. Which AWS capability meets this need?

Choose one.

  • a
    Amazon Bedrock Knowledge Bases

    Knowledge Bases retrieve company data for retrieval-augmented generation; they do not enforce safety policies on prompts and responses.

  • b
    Amazon Bedrock model evaluation

    Model evaluation scores models on datasets before you choose one; it does not filter live prompts and responses in the application.

  • c
    Amazon SageMaker Clarify

    SageMaker Clarify detects bias and explains predictions for ML models; it is not a runtime filter on generative AI prompts and responses.

  • d
    Amazon Bedrock Guardrails Correct

    Guardrails applies one configurable set of safety policies (content filters, denied topics, word filters, PII filters, grounding checks) to both prompts and responses, consistently across foundation models.

The concept

Amazon Bedrock Guardrails is the managed safety layer that evaluates prompts and responses against configurable policies.

Why that’s the answer

Guardrails is configured once and enforced on both the input prompt and the model response, and the same guardrail can be used with different foundation models. Knowledge Bases add retrieval, model evaluation compares models before selection, and SageMaker Clarify measures bias and explainability; none filters live generative traffic.

How to reason it out
  1. Note the need: runtime filtering of prompts and responses, no custom code.
  2. Note the scope: the same policies across several foundation models.
  3. Rule out tools that retrieve data, evaluate models offline, or analyse bias.
  4. Choose Amazon Bedrock Guardrails.

Exam tip: Runtime safety policies on generative AI prompts and responses, across models: Amazon Bedrock Guardrails.

Responsible AI: Bias, Fairness, and Amazon Bedrock Guardrails — the lesson that teaches this.

Question 6Guidelines for Responsible AI

A team's compact document classifier already meets its accuracy requirement. A stakeholder proposes replacing it with a much larger foundation model. The company has committed to minimizing the environmental impact of its AI workloads. Which choice is most responsible?

Choose one.

  • a
    Switch to the larger foundation model and cache its frequent responses

    Caching trims some repeat requests, but every uncached request still runs a much larger model, so energy use stays higher than with the small model.

  • b
    Train a new model from scratch sized exactly to the classification task

    Training from scratch consumes a great deal of compute and energy; reusing an existing model that already meets the target avoids that cost.

  • c
    Keep the current classifier on self-managed servers sized for peak load

    Keeping the small model is right, but moving it to idle self-managed servers provisioned for peak load wastes energy; shared managed infrastructure runs inference more efficiently.

  • d
    Keep the current classifier and decline the switch to the larger model Correct

    Right-sizing is the responsible choice: the current compact model already meets the requirement and uses far less energy per request than a large foundation model.

The concept

Responsible model selection right-sizes the model and reuses existing models to limit energy use.

Why that’s the answer

Larger models need more compute, and therefore more energy, for every inference. When a small model already meets the accuracy target, it is the sustainable choice, run on efficient shared infrastructure. Caching a larger model only softens its cost, keeping the small model on servers sized for peak load wastes energy while they sit idle, and training a new model from scratch spends far more energy than reusing one that works.

How to reason it out
  1. Note that the accuracy requirement is already met by the small model.
  2. Apply the constraint: minimize environmental impact.
  3. Recall that bigger models and training from scratch both consume more energy.
  4. Keep the smallest model that meets the requirement.

Exam tip: Sustainable AI: pick the smallest model that meets the requirement and reuse rather than retrain.

Responsible AI: Bias, Fairness, and Amazon Bedrock Guardrails — the lesson that teaches this.

What AIF-C01 domain 4 tests, topic by topic

The official exam guide breaks Guidelines for Responsible AI into 2 topics. The question bank follows the same split, so a weak topic shows up as a cluster of misses you can go back and read.

Published AIF-C01 practice questions per topic in Guidelines for Responsible AI
TopicWhat it coversQuestions
Explain the development of AI systems that are responsibleExam guide task 4.1 (AIF-C01). Features of responsible AI (bias, fairness, inclusivity, robustness, safety, veracity); tools to identify them (Amazon Bedrock Guardrails); responsible practices to select a model (environmental considerations, sustainability); legal risks of working with GenAI (intellectual property infringement claims, biased model outputs, loss of customer trust, end user risk, hallucinations); characteristics of datasets (inclusivity, diversity, curated data sources, balanced datasets); effects of bias and variance (effects on demographic groups, inaccuracy, overfitting, underfitting); tools to detect and monitor bias, trustworthiness, and truthfulness (analyzing label quality, human audits, subgroup analysis).29
Recognize the importance of transparent and explainable modelsExam guide task 4.2 (AIF-C01). Differences between models that are transparent and explainable and models that are not; tools to identify transparent and explainable models (Amazon SageMaker Model Cards, SageMaker Clarify, Amazon Bedrock Model Evaluations, open source models, data, licensing); tradeoffs between model safety and transparency (measuring interpretability and performance); principles of human-centered design for explainable AI (user-feedback mechanisms, AI decision transparency).29
Total58

Revise Guidelines for Responsible AI before you drill it

Other AIF-C01 domains

Guidelines for Responsible AI: your questions

Guidelines for Responsible AI is domain 4 of the AIF-C01 exam guide and carries 14% of the scored content — the 4th-heaviest of the 5 domains. On a 65-question paper that works out to roughly 9 questions, though AWS does not publish an exact per-domain count and individual exam forms vary.

Source

The domain weight and topic list on this page come from the official AIF-C01 exam guide.