AWS AI Practitioner Guidelines for Responsible AI Practice Questions (Domain 4, 14%)

Domain 4 is 14% of scored content. It covers responsible AI features, Amazon Bedrock Guardrails, bias, variance and dataset quality, legal risks of generative AI, transparent and explainable models, SageMaker Model Cards, and the tradeoff between safety and transparency. Answer these 10 free questions with no signup, written to AWS's AIF-C01 exam guide v1.2.

Last updated November 11, 2026.

10 Free AIF-C01 Guidelines for Responsible AI Questions with Answers

Question 1: Guidelines for Responsible AI

A health ministry is building a model to triage symptom reports from across the country. Patients vary by region, age and language, and some conditions are far rarer than others. Which combination of dataset characteristics matters most?

  1. The largest possible volume, and removal of every record for rare conditions
  2. Recency of every record, and one single data source to keep all formats consistent
  3. Diversity across regions, ages and languages, and balance that represents rare conditions
  4. Records from the capital city only, and labels assigned automatically by a model

Correct answer: C. C is correct because diverse data lets the model serve every patient group, and balancing rare conditions stops the model from defaulting to common ones. A is incorrect because volume does not fix gaps in coverage, and removing rare conditions makes the model unable to triage them. B is incorrect because a single source is less likely to cover all groups, and recency alone does not ensure groups or conditions are represented. D is incorrect because one region cannot represent national patients, and unchecked automatic labels can introduce errors and bias. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.1.

Question 2: Guidelines for Responsible AI

Match each finding to its most likely cause. 1. A model scores poorly on both its training data and new data. 2. A model scores very well on its training data but poorly on new data from the same population. 3. A model scores well overall but much worse for a group that makes up a small share of its training data.

  1. 1 underfitting, 2 overfitting, 3 underrepresentation of the group in the data
  2. 1 overfitting, 2 underfitting, 3 underrepresentation of the group in the data
  3. 1 underfitting, 2 underrepresentation of the group in the data, 3 overfitting
  4. 1 underrepresentation of the group in the data, 2 underfitting, 3 overfitting

Correct answer: A. A is correct because poor results everywhere point to underfitting (high bias), a train-test gap points to overfitting (high variance), and a gap for one small group points to that group being underrepresented. B is incorrect because overfitting gives strong training results, so it cannot explain poor training scores, and underfitting cannot explain very good ones. C is incorrect because a train-test gap on the same population is overfitting, while a gap for one small group is a representation problem. D is incorrect because underrepresentation affects one group rather than all data, underfitting cannot explain very good training scores, and overfitting shows a train-test gap rather than a gap for one group. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.1.

Question 3: Guidelines for Responsible AI

A company plans to fine-tune a customer-service model on text scraped from public forums. A sample shows insults, unverified medical claims and posts copied from other sites. Which dataset practice addresses these problems most directly?

  1. Use all of the forum text, since a larger dataset always improves quality
  2. Balance the dataset so that each forum contributes the same number of posts
  3. Add a Guardrails content filter later and keep the dataset unchanged
  4. Use curated, vetted sources and filter out toxic or unverified content

Correct answer: D. D is correct because curated data sources are chosen and checked for quality, rights and appropriateness, which removes toxic, false and copied content before the model learns from it. A is incorrect because more data containing insults and false claims teaches the model those behaviors; size does not fix quality. B is incorrect because equal contributions from each forum still include the toxic, unverified and copied posts. C is incorrect because a filter on responses can block some insults, but the model would still learn false claims and copied material during fine-tuning. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.1.

Question 4: Guidelines for Responsible AI

A real estate firm's price model has a high error rate on its training data and a similarly high error rate on new listings. Which description fits the model?

  1. Good generalization, because the two error rates are close to each other
  2. Underfitting, because the model is too simple to capture the patterns
  3. Data drift, because the new listings differ from the training listings
  4. Overfitting, because the model memorized the training listings too well

Correct answer: B. B is correct because poor performance on both the training data and new data shows the model has not learned the underlying patterns, which is underfitting, a high-bias problem. A is incorrect because similar error rates only matter if they are low; here both are high, so the model is not useful. C is incorrect because drift would show good training results with worse results on new data, but here training results are poor as well. D is incorrect because an overfit model scores well on its training data, and this model performs poorly on it. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.1.

Question 5: Guidelines for Responsible AI

An agency deploys a generative AI tool that writes short assessments of job applicants from their resumes, which recruiters use to choose who to interview. A review finds the assessments are noticeably more negative for applicants over 50. Which risk is most serious, and what should the agency do first?

  1. Loss of customer trust; add a disclaimer that the assessments are AI-written
  2. Discrimination claims from biased outputs; stop and test outcomes by age group
  3. Intellectual property risk; confirm the model's training data was licensed
  4. Hallucination risk; add a contextual grounding check against each resume

Correct answer: B. B is correct because more negative assessments for older applicants can lead to unequal hiring outcomes and age-discrimination claims, so the agency should pause the tool and measure outcomes by age before relying on it. A is incorrect because a disclaimer does not stop biased assessments from shaping who gets interviewed, which is the source of legal exposure. C is incorrect because licensing concerns ownership of training content, while the finding here is unequal treatment of an age group. D is incorrect because grounding checks catch unsupported statements, but an assessment can be fully grounded in the resume and still be systematically harsher for older applicants. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.1.

Question 6: Guidelines for Responsible AI

A school district is choosing between three foundation models on Amazon Bedrock for a tutoring assistant. It wants evidence of how each model performs on accuracy and toxicity using prompts from its own curriculum before deciding. Which approach fits best?

  1. Choose the model with the most parameters, since larger models are less toxic
  2. Add Amazon Bedrock Guardrails to all three models and pick the cheapest one
  3. Run Amazon Bedrock Model Evaluation jobs on each model with the district's prompts
  4. Compute bias metrics on each provider's training data before choosing

Correct answer: C. C is correct because Bedrock Model Evaluation can score candidate models with automatic metrics such as accuracy and toxicity on a custom prompt dataset, giving comparable evidence before a model is chosen. A is incorrect because size does not guarantee lower toxicity or better accuracy for this curriculum, so the claim needs evidence. B is incorrect because Guardrails filter content at run time, but they do not measure how accurate each model is on the curriculum. D is incorrect because the district has no access to the providers' training data, and bias metrics on it would not measure accuracy or toxicity on its prompts. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.1.

Question 7: Guidelines for Responsible AI

A benefits agency compares two eligibility models on the same holdout data. A logistic regression model is 86% accurate and its full decision logic can be inspected. A deep learning model is 88% accurate. The governing rule says every decision's logic must be fully inspectable, and 85% is the minimum accuracy. Which model should the agency use?

  1. The deep learning model, since higher accuracy always outweighs interpretability
  2. Neither model, since no model can be both accurate and fully interpretable
  3. The deep learning model, explained with SHAP values to satisfy the inspection rule
  4. The logistic regression model, which meets the accuracy minimum and is inspectable

Correct answer: D. D is correct because the rule requires fully inspectable logic, and the interpretable model also clears the 85% minimum, so the 2-point gain does not justify losing transparency. A is incorrect because the governing rule makes inspectable logic a requirement, so higher accuracy cannot override it. B is incorrect because the logistic regression model already meets the 85% minimum while remaining fully interpretable. C is incorrect because SHAP values approximate a model's behavior after the fact, which does not make its full decision logic inspectable. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.2.

Question 8: Guidelines for Responsible AI

A city council will only use an AI model if the provider publishes documentation of the model's training data sources, intended uses, evaluation results and known limitations. Which responsible AI goal does this requirement most directly support?

  1. Transparency, so stakeholders can understand how the model was built
  2. Robustness, so the model handles noisy inputs without failing
  3. Sustainability, so the model uses less energy for each request
  4. Controllability, so administrators can tune the model's behavior

Correct answer: A. A is correct because published documentation about data, purpose, evaluation and limits lets others understand and scrutinize a model, which is the core of transparency. B is incorrect because documentation describes the model; it does not change how the model handles noisy input. C is incorrect because documentation does not reduce the model's compute or energy use. D is incorrect because reading about a model does not give the council any means to adjust how it behaves. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.2.

Question 9: Guidelines for Responsible AI

A regulator requires that a lender's loan officers can see exactly how each applicant input affects the approval decision, by reading the model itself rather than relying on an approximation. Which kind of model fits this requirement?

  1. A large language model asked to describe the reasons for its decision
  2. An inherently interpretable model, such as a shallow decision tree
  3. A deep neural network with many hidden layers and millions of weights
  4. An ensemble of hundreds of models that vote together on each loan decision

Correct answer: B. B is correct because a shallow decision tree is transparent by design: its splits can be read directly, so officers can trace exactly how each input led to the decision. A is incorrect because an LLM's self-described reasons are generated text, not a reliable view of how the decision was actually made. C is incorrect because a deep network's weights cannot be read as decision logic, so it would need post-hoc explanations, which are approximations. D is incorrect because combining hundreds of models makes the decision logic harder, not easier, to read directly. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.2.

Question 10: Guidelines for Responsible AI

A hospital network wants to be transparent about the data used to train its readmission model, but the training records contain patient health information. Which approach best serves transparency without creating a privacy risk?

  1. Publish nothing at all, since health data can never be described publicly
  2. Publish the records after removing only the patients' names from each one
  3. Publish data sources, collection dates and group coverage, not the raw records
  4. Publish the raw training records so anyone can verify the model directly

Correct answer: C. C is correct because describing where the data came from, when it was collected and which patient groups it covers explains the data to stakeholders without exposing any individual's health information. A is incorrect because a summary of sources and coverage contains no individual's data and still gives stakeholders useful transparency. B is incorrect because records without names can still identify patients through other details, such as dates and rare conditions. D is incorrect because releasing patient records would expose health information, which is a serious privacy and legal risk. Exam guide: Domain 4 (Guidelines for Responsible AI), Task Statement 4.2.

Guidelines for Responsible AI Task Statements in the Exam Guide

TaskExam guide task statementDomainQuestions
4.1Explain the development of AI systems that are responsibleGuidelines for Responsible AI33
4.2Recognize the importance of transparent and explainable modelsGuidelines for Responsible AI22

More AWS AI Practitioner Practice

FlashGenius is not affiliated with or endorsed by Amazon Web Services. AWS and AWS Certified AI Practitioner are trademarks of Amazon.com, Inc. or its affiliates. These are original practice questions, not questions from the real exam. Your score here is practice accuracy, not an AWS score or a pass prediction.