Guidelines for Responsible AI

A healthcare company deploys a customer-facing chatbot on Amazon Bedrock. Compliance requires that the assistant never provide medical diagnoses, refuse to discuss competitor products, and automatically redact any patient email addresses or phone numbers that appear in user input before they reach the model. The team wants a configurable, model-agnostic way to enforce all of these controls without writing custom filtering code for each rule. Which approach best meets these requirements?

A bank uses SageMaker to train a credit approval model. After training, the compliance team wants to verify whether the model approves applications at different rates across gender groups, so they can document and address any unfair outcomes before deployment. Which approach BEST supports this goal?

A bank uses a machine learning model to approve or deny personal loan applications. Regulators require the bank to provide each rejected applicant with a clear reason showing which specific input factors (such as income, debt ratio, or credit history) most influenced the individual decision. Which AWS capability best addresses this requirement?

A bank deploys a machine learning model that approves or denies loan applications. During an internal review, a compliance officer asks the data science team to provide two separate deliverables: (1) documentation describing how the model was built, its intended use, training data sources, and known limitations, and (2) the ability to explain, for any individual denied applicant, which input features most influenced that specific decision. Which pair of responsible-AI concepts do these two requests correspond to, respectively?

A company builds an ML model that screens job applicants and recommends candidates for interviews. The compliance team wants to verify that the model recommends candidates at similar rates across gender groups, regardless of any differences in the applicant pool. Which fairness concept should the team measure to confirm this specific requirement?

A bank uses a machine learning model to approve loan applications. During review, the responsible-AI team finds the model achieves very high overall accuracy but approves applicants from one demographic group at a substantially higher rate than equally qualified applicants from another group. Leadership wants the model to treat qualified applicants equitably, even if it means the top-line accuracy number decreases slightly. Which responsible-AI dimension and trade-off does this situation primarily illustrate?

A financial services company deploys a customer-facing chatbot on Amazon Bedrock that answers questions using a company knowledge base. Compliance requires that the chatbot never fabricate information and only respond based on retrieved source documents. The team wants a managed capability that automatically detects and filters responses that are not factually grounded in the provided source material. Which approach best meets this requirement?

A retail bank deploys a customer-facing chatbot on Amazon Bedrock to answer questions about savings accounts and card services. Compliance requires that the bot never engage in discussions about investment advice, stock picks, or cryptocurrency, even when users ask about them indirectly. Which Guardrails for Amazon Bedrock capability most directly enforces this requirement?

A healthcare company deploys a Bedrock-based patient-facing assistant. Leadership wants to add aggressive content filters that block any response touching on medication dosages, symptoms, or diagnoses to eliminate liability risk. The product team warns that this will make the assistant refuse most legitimate questions and frustrate users. Which statement best describes the responsible-AI trade-off the team is facing?

A company deploys a customer-facing chatbot on Amazon Bedrock. Leadership requires that the assistant never recommends or discusses specific named competitor products, and that any user attempt to make it do so is blocked with a standard refusal message. Which capability of Guardrails for Amazon Bedrock most directly enforces this requirement?

A bank deploys a generative AI assistant that drafts loan approval recommendations for officers to review. To align with responsible-AI principles for a high-stakes decision, which practice should the bank prioritize?

A bank is building a loan-approval model. Regulators require that the bank be able to explain to each rejected applicant exactly which factors drove the decision. The data science team is choosing between a highly accurate deep neural network and a simpler, inherently interpretable model. Which consideration should most influence their choice for this use case?

A bank deployed a loan-approval ML model six months ago. The team validated fairness across demographic groups before launch. Recently, customer complaints suggest approval rates have shifted unfavorably for certain groups, even though the model code has not changed. Which responsible-AI practice most directly addresses this situation?

A bank deployed a loan-approval ML model six months ago. Regulators now require the bank to demonstrate that the model's approval outcomes do not systematically disadvantage applicants based on protected attributes like gender, using live production data. Which approach best satisfies this requirement?

A financial services company deploys a foundation-model chatbot for customer inquiries. During pre-launch testing, the security team discovers that carefully crafted, malicious prompts can manipulate the model into ignoring its instructions and revealing internal system details. The company wants to strengthen the responsible-AI dimension that ensures the model performs reliably and safely even when facing unexpected or adversarial inputs. Which responsible-AI dimension is the team primarily addressing?

A bank deploys a foundation-model-based loan-inquiry assistant. During testing, the team notices that slightly rephrasing a question, adding extra whitespace, or including minor typos sometimes causes the model to produce wildly different or inconsistent answers. The team wants to ensure the assistant reliably handles these varied and imperfect inputs without degrading its performance. Which responsible-AI dimension is the team primarily addressing?

A company launches a customer-facing chatbot built on Amazon Bedrock. During testing, the team notices the model occasionally produces hateful and profane language when users provoke it. The team wants a managed way to detect and block toxic and offensive content in both user inputs and model outputs without retraining the model. Which approach best addresses this responsible-AI safety concern?

A financial services company is deploying a machine learning model that approves or denies small-business loan applications. Regulators and internal auditors have requested standardized documentation describing the model's intended use, training data characteristics, performance across different applicant groups, and known limitations. Which AWS capability best addresses this transparency requirement?

A healthcare company deploys a generative AI assistant that answers patient questions about medications. During testing, the team notices the model occasionally states confident but factually incorrect drug interaction information. Which responsible-AI dimension does this problem most directly relate to, and what is an appropriate mitigation?

A bank has deployed a machine learning model that approves or denies personal loan applications. Regulators require the bank to explain, for any individual denied applicant, which input features most influenced that specific decision. Which AWS capability best satisfies this requirement?