Task 4.1: Explain the development of AI systems that are responsible.

• AAmazon Bedrock Guardrails

  Correct

  Guardrails can block denied topics such as investment advice in prompts and responses, and their sensitive information filters can block or mask PII.

• BAmazon Macie

  Incorrect

  Macie discovers sensitive data such as PII in Amazon S3. It does not sit between users and a model to filter conversations.

• CAmazon Inspector

  Incorrect

  Inspector scans workloads for software vulnerabilities and unintended network exposure. It does not filter model prompts or responses.

• DAmazon SageMaker Model Registry

  Incorrect

  Model Registry catalogs model versions and their approval status. It does not inspect conversations.

Denied topics and PII masking are both Guardrails safeguards, applied to user inputs and model responses alike.

• AFairness

  Incorrect

  Fairness is about considering the impacts on different groups of stakeholders, not about resisting unusual inputs.

• BVeracity and robustness

  Correct

  AWS defines veracity and robustness as achieving correct system outputs, even with unexpected or adversarial inputs.

• CTransparency

  Incorrect

  Transparency is about enabling stakeholders to make informed choices about their engagement with an AI system.

• DControllability

  Incorrect

  Controllability means having mechanisms to monitor and steer AI system behavior.

AWS lists eight dimensions: fairness, explainability, privacy and security, safety, controllability, veracity and robustness, governance and transparency.

• AIt is overfit, a high-variance model that does not generalize

  Correct

  Overfit models perform well on the training set but not on the test set because they fit the training data too closely; AWS describes this as high variance.

• BIt is underfit, a high-bias model that is too simple for the data

  Incorrect

  Underfit models produce inaccurate results on both the training data and the test set. This model is very accurate on its training data.

• CIt is well fitted, because training accuracy is what matters

  Incorrect

  A well-fitted model captures the dominant trend for both seen and unseen data. Accuracy on new data is what shows whether a model generalizes.

• DIt is hallucinating, because only generative models make this error

  Incorrect

  Hallucination describes incorrect output from a generative model. The gap between training and test accuracy is the classic sign of overfitting in any ML model.

Great on training data and poor on new data means high variance (overfitting); poor on both means high bias (underfitting).

• ARetrain the model on more data so that the overall accuracy rises

  Incorrect

  A higher overall number can still hide a group the model serves badly; the question is about differences between groups.

• BMeasure the model's performance separately for each applicant group

  Correct

  Bias includes disparities in a model's performance across different groups, and cohort-level analysis examines how the model behaves for a particular group.

• CRaise the approval threshold for every applicant by the same amount

  Incorrect

  Changing every decision equally does not reveal whether any group is treated differently, and it is not a measurement.

• DRemove the reviewer's concern by publishing the overall accuracy

  Incorrect

  Publishing an aggregate figure does not test for bias; disparities across groups have to be measured.

Fairness problems hide inside averages. Subgroup analysis, comparing metrics group by group, is how they are found.

• AClaims that generated content infringes someone's copyright

  Correct

  Copyright claims arising from generative output are a recognized risk; AWS even offers an IP indemnity for copyright claims on output from certain of its generative AI services.

• BFalse statements about products that mislead customers and erode trust

  Correct

  Hallucinations are incorrect or misleading outputs, and they reduce user trust; published as product claims, they create risk for the business.

• CThe model's context window being too small for long prompts

  Incorrect

  A context window limit is a technical design constraint, not a legal or reputational risk.

• DHigher latency during peak shopping hours

  Incorrect

  Latency is a performance concern for model selection, not a legal risk.

• EPaying per token instead of per hour for inference

  Incorrect

  Token-based pricing is a cost model. It does not create legal exposure.

Generative output carries intellectual property, accuracy and bias risks that a human review and clear policies must address before publication.

• AThe model is overfit and has high variance

  Incorrect

  Overfit models perform well on the training set but not on the test set; this model is poor on both.

• BThe model is well fitted and ready to deploy

  Incorrect

  A well-fitted model captures the dominant trend for seen and unseen data, which this model does not.

• CThe model is underfit and has high bias

  Correct

  Underfit models produce inaccurate results on both training and test data, for example because the model is too simple or was not trained long enough; AWS describes this as high bias.

• DThe model has data drift from production traffic

  Incorrect

  The model already fails on its own training data, so the problem is the fit, not a change in production data.

Poor on both sets means high bias (underfitting); good on training but poor on new data means high variance (overfitting).

• AContent filters

  Correct

  Content filters detect and filter harmful content in prompts and responses across categories such as hate, insults, sexual, violence and misconduct, with a configurable strength per category.

• BSensitive information filters

  Incorrect

  Sensitive information filters detect and block or mask PII. They do not rate hate or violence.

• CContextual grounding check

  Incorrect

  The contextual grounding check detects hallucinations against a reference source, not harmful content.

• DDenied topics

  Incorrect

  Denied topics block specific subjects that are undesirable for the application, such as investment advice. Category-based harm filtering with adjustable strength is the content filter.

Guardrails safeguards each target a different risk: content filters for harm categories, denied topics for subjects, sensitive information filters for PII, grounding checks for hallucinations.

• ADuplicate the existing records to make the dataset larger

  Incorrect

  Copying the same records adds volume but not diversity; the groups that are missing stay missing.

• BKeep only the most recent month of data to stay current

  Incorrect

  Cutting the dataset to one month shrinks it further and does nothing to fix which groups it represents.

• CAdd data so that all relevant applicant groups are represented

  Correct

  Pretraining bias occurs when training data does not fairly represent the real-world distribution, for example with too little data from specific age groups.

• DTrain longer on the same data until accuracy stops improving

  Incorrect

  More training on unrepresentative data can fit it more closely, but it cannot represent groups the data does not contain.

• EBalance the dataset so that no group dominates training

  Correct

  Imbalanced data can bias training, so the model performs well on the majority and poorly on the rest; a broad, diverse and unbiased dataset makes accurate responses more likely.

Inclusive, diverse and balanced datasets are the first defense against biased models; detect remaining bias by measuring results per group.

• AAlways choose the largest available model to reduce retries

  Incorrect

  The lens recommends the opposite direction: smaller, optimized models that meet the performance goal.

• BKeep a fixed fleet of large GPU instances running around the clock, even when idle

  Incorrect

  The lens recommends auto scaling and serverless architectures to optimize resource utilization, not idle fixed capacity.

• CFine-tune a new copy of the model for every customer request

  Incorrect

  Repeated training adds compute and energy use; the lens recommends efficient customization and smaller models.

• DUse smaller models with techniques such as quantization and distillation

  Correct

  The lens recommends smaller models and optimization techniques such as quantization, pruning and distillation to reduce resource use and promote sustainability while meeting performance goals.

Responsible model selection includes environmental cost: pick the smallest model that meets the requirement, and optimize how it runs.

• AContext coverage

  Incorrect

  Context coverage measures how much of the ground truth the retrieved texts cover.

• BStereotyping

  Correct

  Stereotyping measures generalized statements about individuals or groups of people in responses.

• CCitation precision

  Incorrect

  Citation precision measures how many cited passages were correctly cited.

• DCompleteness

  Incorrect

  Completeness measures how well responses resolve all aspects of the questions.

Built-in responsible AI metrics such as stereotyping and harmfulness let teams monitor bias alongside quality metrics.