Task 4.2: Recognize the importance of transparent and explainable models.

• AThe neural network is the safer choice, because a more accurate model is always easier to explain

  Incorrect

  ML models such as deep neural networks are similar to black boxes whose inner workings are hidden; accuracy does not make them easier to explain.

• BBoth choices are equally transparent, because any model can be explained exactly after the decision is made

  Incorrect

  Post hoc explanations of black-box models can lack robustness, can be fooled by adversarial attacks, and can be ambiguous when several explanations fit.

• CThe interpretable model lets reviewers trace each decision, at the cost of some accuracy

  Correct

  Interpretable models can be understood in the way they work, while explanations added after the fact to a black box can be ambiguous. The bank trades a little accuracy for decisions it can justify.

• DNeither choice matters, because regulators only review a model's overall accuracy

  Incorrect

  Financial providers may have to demonstrate explainability in the AI systems they deploy, and individual loan decisions are what local explanations examine.

Transparency versus performance is a real tradeoff. When each decision must be justified, an interpretable model is often worth a small loss of accuracy.

• AAWS AI Service Cards

  Incorrect

  AI Service Cards are published by AWS about AWS's own AI services and models. They do not document the company's models.

• BAWS Artifact

  Incorrect

  AWS Artifact gives access to AWS security and compliance reports and agreements, not documentation of customer models.

• CAmazon CloudWatch

  Incorrect

  CloudWatch collects metrics and logs to monitor resources and applications. It is not a model documentation tool.

• DAmazon SageMaker Model Cards

  Correct

  Model Cards document critical details about ML models in one place, such as intended use, risk rating, training details and metrics, and evaluation results, for governance and reporting.

SageMaker Model Cards document your models; AWS AI Service Cards document AWS's AI services. Both support transparency.

• AAWS AI Service Cards

  Correct

  AI Service Cards give a single place to find the intended use cases and limitations, responsible AI design choices and performance best practices for AWS AI services and models.

• BAmazon SageMaker Model Cards

  Incorrect

  Model Cards are where a customer documents its own ML models. They are not AWS's documentation of its services.

• CAWS Trusted Advisor

  Incorrect

  Trusted Advisor inspects your AWS environment and recommends ways to save money, improve availability and performance, or close security gaps.

• DAWS CloudTrail

  Incorrect

  CloudTrail records the actions taken in your AWS account for auditing. It holds no information about a service's design choices.

Transparency runs both ways: AWS documents its services in AI Service Cards, and you document your models in SageMaker Model Cards.

• AHide that the ranking comes from AI so adjusters trust it more

  Incorrect

  Transparency means enabling stakeholders to make informed choices about their engagement with an AI system, which requires telling them it is there.

• BRemove the adjuster from the loop so decisions are made faster

  Incorrect

  Explainable AI principles aim to augment human capabilities rather than replace people.

• CShow the adjuster the main factors behind each recommendation

  Correct

  Explainability is about understanding and evaluating system outputs; with explanations, the intended users are more likely to understand and trust the results.

• DLet adjusters flag a wrong recommendation and feed that back for review

  Correct

  Controllability calls for mechanisms to monitor and steer AI behavior, and collecting user feedback on model output is a documented way to improve a model for real people.

• EDisplay the model's raw parameter values as the explanation

  Incorrect

  Raw parameters do not help a user understand a decision; explanations are meant to make results understandable to their intended users.

Human-centered explainable AI keeps people informed (why this result) and in control (a way to push back), so trust is earned rather than assumed.

• ALocal explainability

  Correct

  Local explainability investigates individual decisions the model makes during actual usage, such as one rejected loan application.

• BGlobal explainability

  Incorrect

  Global explainability explains how the model works overall and which features most influence its output, not one decision.

• CCohort explainability

  Incorrect

  Cohort explainability investigates the model's behavior for a group of data, not for a single applicant.

• DPretraining bias analysis

  Incorrect

  Pretraining bias analysis checks whether training data fairly represents the real world; it does not explain one prediction.

Global explains the whole model, cohort explains a group, local explains one decision.

• AThe model's intended use and risk rating

  Correct

  Model cards catalog details such as the intended use and risk rating of a model.

• BEvaluation results and observations

  Correct

  Model cards capture training details and metrics as well as evaluation results and observations.

• CThe live text of every prompt sent to the model

  Incorrect

  Model cards document a model for governance; recording every request is a logging job.

• DThe database passwords the application uses

  Incorrect

  Credentials belong in a secrets store such as AWS Secrets Manager, not in model documentation.

• EAWS's compliance reports for SageMaker AI

  Incorrect

  AWS compliance reports are downloaded from AWS Artifact; model cards document your own models.

Model cards make a model transparent to reviewers: what it is for, how risky it is, how it was trained and how it performed.

• AAmazon Bedrock Prompt Management

  Incorrect

  Prompt Management stores and versions prompts; it does not measure model behavior.

• BAWS Trusted Advisor

  Incorrect

  Trusted Advisor checks an AWS environment for cost, performance and security improvements, not model behavior.

• CAmazon Bedrock Model Evaluation

  Correct

  Model Evaluation on Amazon Bedrock helps you evaluate, compare and select FMs based on metrics such as accuracy, robustness and toxicity.

• DAmazon Bedrock batch inference

  Incorrect

  Batch inference runs many prompts asynchronously; it produces outputs but no evaluation metrics or comparison.

Transparent model choice rests on published evidence: evaluation results that others can inspect and repeat.

• AMake the assistant sound fully human and never mention that it is an AI system

  Incorrect

  Hiding the AI takes away the customer's ability to make an informed choice about engaging with it.

• BDisable all logging so customers' questions are never reviewed

  Incorrect

  This removes oversight; it does nothing to inform customers about the system they are using.

• CGive answers without sources so that they look more confident

  Incorrect

  Hiding where answers come from makes outputs harder for people to understand and evaluate.

• DTell customers they are using AI and offer a way to reach a person

  Correct

  AWS defines transparency as enabling stakeholders to make informed choices about their engagement with an AI system, and controllability as having mechanisms to steer AI behavior.

Human-centered design keeps people informed (this is AI, here is why) and in control (here is how to reach a person or push back).

• AContent filters

  Incorrect

  Content filters detect harmful content categories such as hate or violence; they do not verify answers against policy rules.

• BAutomated Reasoning checks

  Correct

  Automated Reasoning checks use formal logic to validate content against policies you define and provide structured feedback on why a response is correct or incorrect.

• CSensitive information filters

  Incorrect

  Sensitive information filters detect and mask PII; they do not explain whether an answer follows policy.

• DWord filters

  Incorrect

  Word filters block listed words and phrases by matching; they provide no logical explanation of correctness.

Automated Reasoning checks add explainable, auditable verification, which regulated industries value alongside accuracy.