Responsible AI and guardrails: AIF-C01 study notes and practice questions
What to know about Responsible AI and guardrails for the AIF-C01 exam, from the Guidelines for Responsible AI domain (14% of the exam guide), then 5 free questions to test it. Part of the AIF-C01 study guide.
B. Wrong: Word filters match exact words, so a customer who asks for advice in different words would still get a reply. They do not understand the topic.
C. Wrong: Sensitive information filters protect personal data. They do not stop the chatbot from giving investment advice.
D. Wrong: A grounding check looks at whether an answer is supported by source material. It does not block a whole subject like investment advice.
Explanation
Amazon Bedrock Guardrails has several policy types, and each handles a different risk. Denied topics lets a bank define a subject the assistant must not discuss, which is the requirement here. This question tests matching a requirement to the right guardrail policy.
A utility company is about to launch a billing assistant.
Customers will type quickly, misspell words and sometimes try to trick the assistant into revealing other customers' data or ignoring its rules. The team must show that the assistant holds up under these inputs before launch.
Which testing approach fits BEST?
Choose one answer.
AA. Adversarial testing with difficult prompts, including misspellings and manipulation attempts, logging each failure.
BB. Accuracy testing on the standard FAQ answers, using the same clean, well-formed wording the assistant was built and tuned with.
CC. A load test sending thousands of simple, well-formed billing questions per minute to measure response times and errors.
DD. A customer satisfaction survey run after launch, with changes to the assistant's wording made from the feedback customers return.
B. Wrong: Clean wording only tests the easy cases. The assistant could still fail when customers misspell words or try to manipulate it.
C. Wrong: Load testing checks capacity and speed. It says nothing about how the assistant handles tricky or misleading inputs.
D. Wrong: A survey only collects feedback after launch. It cannot show that the assistant holds up before customers use it, so it does not meet a pre-launch requirement.
Explanation
Responsible AI practice includes testing a model against unexpected and adversarial inputs, because real users will not always write clean questions. Robustness testing before launch finds weak points while they are still cheap to fix. This question tests choosing a testing method that matches the risk.
A company with a net zero target runs a high-volume document classification workload on AWS.
It uses the largest available model on an endpoint that runs all day, although traffic is heavy only during office hours.
Which TWO changes BEST reduce energy use while keeping the required accuracy? (Choose two.)
Choose two answers.
AA. Move to a smaller model that meets the agreed accuracy target on a test set of the company's own documents.
BB. Keep the largest model, because a bigger model reduces the need to check results later.
CC. Stop or scale down the endpoint outside office hours, so that capacity does not sit idle overnight.
DD. Duplicate the always-on endpoint in a second Region, so that classification continues if one Region fails.
EE. Store every version of every scanned document in S3 Standard indefinitely, so that no file is ever lost.
Show the answer and explanation
Answers: A, C
A. Right: Correct. Right-sizing uses the least compute that meets the requirement, and a smaller model that passes the accuracy test uses less energy for each document.
B. Wrong: A larger model uses more compute for every document. It does not reduce energy use and works against the net zero goal.
D. Wrong: A second always-on endpoint doubles idle capacity and adds energy use. It improves resilience but does not reduce the workload's footprint.
E. Wrong: Keeping every version forever raises storage energy and cost with no need from the workload. Retention should follow the business rule, not an open-ended default.
Explanation
The Sustainability Pillar of the Well-Architected Framework asks teams to right-size workloads and to eliminate idle resources, since both reduce the energy needed. A smaller model that still meets accuracy and reduced capacity outside office hours apply those two principles to an AI workload. This question tests applying them to a model choice and an endpoint schedule.
A toy company uses an image model to design cereal-box packaging.
One generated picture closely resembles a cartoon character owned by another company, and the company plans to print it.
Which legal risk is MOST directly raised by printing this image?
Choose one answer.
AA. Trademark or copyright infringement claims from the character's owner over the close resemblance.
BB. A personal data breach, because the generated image may contain faces of real customers.
CC. Discrimination claims from customers who feel the packaging design excludes people in their group.
DD. Defamation claims from the character's creator, because the picture makes false statements about them.
Show the answer and explanation
Answer: A
A. Right: Correct. Printing an image that copies a protected character can infringe the owner's trademark or copyright, so the owner could bring a claim against the toy company.
B. Wrong: A personal data breach needs personal information to be exposed or misused. A generated cartoon does not raise that risk, since the question concerns the picture's resemblance to a character.
C. Wrong: Discrimination claims depend on unfair treatment of people in a protected group. A packaging design that resembles a character does not raise this risk.
D. Wrong: Defamation requires a false statement that harms a person's reputation. A picture that resembles a character does not make any such statement.
Explanation
Generated images can resemble existing protected works, which creates intellectual property risk for any business that publishes them. Reviewing outputs for resemblance to known marks and getting legal advice before printing are common controls. This question tests recognizing which risk a generated image creates.
Many residents are elderly or not confident online, and some need help in a second language. Benefit decisions can affect rent and food for a household.
Which design choice BEST reflects human-centered design for this service?
Choose one answer.
AA. Plain-language answers, with a clear and always-visible option to reach a human advisor by phone or in person.
BB. Detailed legal answers with citations to every regulation, so that residents can check the rules themselves.
CC. Hide the human-advisor option until a resident has tried the chatbot three times, to reduce call volume.
DD. A fast chatbot that answers in English only, with identical wording for every resident, to keep running costs low.
Show the answer and explanation
Answer: A
A. Right: Correct. Plain language and an easy route to a person suit residents who may struggle online or in a second language. A human route matters when decisions affect rent and food.
B. Wrong: Citations help professionals, but dense legal text is hard for many residents to read. It can overwhelm people who need help, and they may give up.
C. Wrong: Hiding human help to cut calls blocks the residents who most need it. It treats the service as a cost to control rather than a need to meet.
D. Wrong: English-only answers leave out residents who need another language. Low cost does not justify excluding them from a public service.
Explanation
Human-centered design starts from the needs of the people who use a service, including those with limited digital skills and language needs. It also gives clear routes to a person when the stakes are high. This question tests applying that principle to a public-facing AI service.
Want more on Responsible AI and guardrails? The 5 full-length practice sets have 21 more Responsible AI and guardrails questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →