Transparency and explainability: AIF-C01 study notes and practice questions
What to know about Transparency and explainability for the AIF-C01 exam, from the Guidelines for Responsible AI domain (14% of the exam guide), then 2 free questions to test it. Part of the AIF-C01 study guide.
These questions ask when explanations are needed and how to provide them.
Transparent vs opaque models
Transparent (interpretable)
Opaque (“black box”)
Examples
Linear regression, decision trees
Deep neural networks, large foundation models
Explanations
Built in: you can see which factors mattered
Need extra tools, and are approximations
Performance
Often lower on complex tasks
Often higher
The trade-off: the most powerful models are the hardest to explain. When a regulator or customer needs to know why (such as a declined loan), a more transparent model may be the right choice even if it is less accurate.
Why transparency matters
Trust: users and customers understand decisions.
Accountability and regulation: some decisions must be explainable by law.
Debugging: finding why a model makes mistakes or behaves unfairly.
AWS tools
SageMaker Clarify: feature importance (which inputs drove a prediction) and bias reports.
SageMaker Model Cards: document a model’s intended use, risks, training data and evaluation results, for governance.
AWS AI Service Cards: AWS’s own documentation of intended use, limits and responsible AI design for its AI services.
Open-source models and data transparency: knowing what data a model was trained on helps explain its behavior.
Human-centered design for explainable AI: present explanations people can actually use.
Common traps
Assuming you can fully explain a large language model. Tools give partial insight, not a full account.
Model cards vs Clarify. Model cards document; Clarify analyzes.
Choosing the most accurate model when explanations are mandatory. A slightly less accurate model that can explain each decision may be the only one you are allowed to use.
A bank's model risk team wants one governed record for each ML model.
It must hold the intended uses, a risk rating, the training details and the evaluation results, and the team must be able to export the record as a PDF for reviewers.
Which SageMaker AI capability should the team use?
Choose one answer.
AA. SageMaker Model Cards, which document intended uses, risk rating, training details and evaluation results.
BB. SageMaker Model Registry, which stores model versions and their approval status for deployment.
CC. SageMaker Model Monitor, which watches live endpoints for data drift and changes in model quality.
DD. AWS Config, which records configuration changes to AWS resources and evaluates them against rules.
B. Wrong: The registry tracks model versions and approvals. It does not hold intended uses, risk ratings or documented evaluation results.
C. Wrong: Model Monitor watches a deployed model for drift in live data. It does not document a model's intended uses or training details.
D. Wrong: Config records resource configurations, not documentation about an ML model. It cannot hold intended uses or evaluation results.
Explanation
Model cards record what a model is for, its limits, and how it was trained and evaluated, in a structured form that can be exported for governance. The other tools monitor, version or configure models rather than document them. This question tests choosing the documentation tool over the deployment and monitoring tools.
A lender must explain each rejection with the main factors that drove that decision, and those reasons must match what the model actually used.
A deep learning model is 0.5 percentage points more accurate than a simple, transparent scorecard. The lender processes 20,000 applications a month.
Which approach BEST meets these requirements?
Choose one answer.
AA. Keep the deep model and publish its overall accuracy figure in every rejection letter so applicants see its reliability.
BB. Keep the deep model and have underwriters write each rejection reason by hand after reviewing the full application file.
CC. Deploy the transparent scorecard, whose reasons come straight from its inputs, and accept the small accuracy loss.
DD. Keep the deep model and publish its full network architecture so applicants can see how the model is built internally.
Show the answer and explanation
Answer: C
A. Wrong: Overall accuracy describes the model across all applicants. It does not say why this applicant was refused, so it fails the explanation requirement.
B. Wrong: Hand-written reasons can be accurate, but they cannot scale to 20,000 applications a month. They may also drift from what the model actually used.
C. Right: Correct. A transparent scorecard shows each applicant the factors behind the score, and those reasons match what the model used. The small accuracy loss is a reasonable price for explainable decisions. Learn more: Amazon SageMaker Model Cards (opens in a new tab)
D. Wrong: A network architecture describes how the model is built, not what drove one decision. Applicants would still not learn their main reasons.
Explanation
Explainability needs reasons that trace back to the factors the model actually used. A transparent model provides this directly, whereas a deep model needs extra explanation methods that give approximate reasons. This question tests balancing a small accuracy gain against the requirement to explain every decision at volume.
Want more on Transparency and explainability? The 5 full-length practice sets have 15 more Transparency and explainability questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →