Guidelines for Responsible AI practice questions

Bias is the systematic error; fairness is the equitable outcome you reach by removing it.

AWS frames bias as a systematic error that skews results for or against groups, usually from skewed data, while fairness is the goal of equitable treatment. The other options reverse the pair, equate them, or invent scope limits that do not exist.

1. Define bias: a systematic, unfair error favoring or disfavoring groups.
2. Define fairness: the goal of treating individuals and groups equitably.
3. Relate them: remove bias to achieve fairness.

Exam tip: Bias is the flaw; fairness is the goal reached by removing it.

Robustness is reliable performance under noisy, unexpected, or adversarial inputs.

The definition of resilience to difficult inputs matches robustness. Inclusivity concerns serving diverse people, veracity concerns truthfulness, and safety concerns preventing harmful output, so none of them fit reliability under bad inputs.

1. Read the definition: reliable performance under noisy or adversarial inputs.
2. Compare each feature to that definition.
3. Only robustness matches.

Exam tip: Robustness is reliable performance despite noisy, unexpected, or adversarial inputs.

Amazon Bedrock Guardrails sensitive information filters redact or block PII in responses.

Account numbers and emails are PII, and the sensitive information filter is the guardrail policy that detects and redacts PII. Denied topics refuse subjects, contextual grounding checks target hallucinations, and content filters target harmful categories, so none of them stop personal-data leakage.

1. Identify the leaked data as PII (account number, email).
2. Recall the Guardrails policy that handles PII detection and redaction.
3. Select sensitive information filters.

Exam tip: To stop a generative AI app from leaking PII, use Bedrock Guardrails sensitive information filters.

Denied topics make a Bedrock application refuse specific subjects.

Refusing an entire subject such as investment advice is exactly what denied topics do. Word filters block individual words, sensitive information filters handle PII, and content filters block harmful categories, none of which target a defined subject to refuse.

1. Note the requirement: refuse a whole subject (investment advice).
2. Match it to the Guardrails policy that defines refused subjects.
3. Select denied topics.

Exam tip: Use denied topics to make a Bedrock app refuse defined subjects.

Amazon Bedrock Guardrails is a managed safety layer for generative AI on Bedrock.

Guardrails lets you configure policies that Bedrock enforces on both prompts and responses, applied consistently across foundation models. It does not train models, document them, or track billing, which rules out the other options.

1. Recall that Guardrails sits between the user and the model.
2. It enforces configured policies on prompts and responses.
3. That is a managed safety layer, matching option D.

Exam tip: Bedrock Guardrails is a managed safety layer enforcing policies on prompts and responses.

Responsible model selection includes sustainability: right-size the model because bigger models use more energy.

Since compute drives energy use and larger models use more compute, choosing the smallest model that meets accuracy and latency needs is the sustainable choice. Always picking the biggest model, denying the size-energy link, or claiming larger models are more efficient all contradict this.

1. Recognize both models meet the accuracy requirement.
2. Recall that larger models consume more energy to train and run.
3. Right-size to the smaller model for sustainability.

Exam tip: Right-size the model: bigger models cost more energy, so pick the smallest that meets the need.