GARP RAI: Responsible and Ethical AI

Use this page to isolate Responsible and Ethical AI before returning to mixed GARP RAI practice.

Topic snapshot

Sample questions

These questions are original Finance Prep practice items aligned to this topic area. They are designed for self-assessment and are not official exam questions.

Question 1

Topic: Responsible and Ethical AI

A bank’s responsible AI policy states that each AI use case must have an assigned business owner who approves the use, ensures controls operate, tracks remediation of issues, and remains answerable for customer and risk outcomes even when outputs are generated by a model. Which responsible AI concept is this describing?

• A. Transparency
• B. Human oversight
• C. Fairness
• D. Accountability

Best answer: D

What this tests: Responsible and Ethical AI

Explanation: The description maps to accountability. In responsible AI, accountability ensures that a defined person, role, or governance body remains responsible for how an AI system is approved, controlled, monitored, and remediated. The key point is that responsibility does not shift to the model or vendor simply because the system produces automated recommendations or outputs. The assigned owner must ensure appropriate controls exist, issues are escalated and resolved, and business, customer, and risk outcomes are managed within policy and risk appetite. Human oversight may support accountability, but oversight alone is not the same as clear ownership for decisions and outcomes.

• Transparency concerns making AI use, data, logic, or limitations understandable, not assigning ownership.
• Fairness concerns avoiding unjustified bias or disparate impact, not responsibility for controls and issue remediation.
• Human oversight involves review or intervention by people, but the stem emphasizes answerability and ownership.

Accountability means clear ownership is retained for AI decisions, controls, issues, and outcomes.

Question 2

Topic: Responsible and Ethical AI

A bank uses an AI system to auto-approve low-risk loan increases and refer marginal cases to underwriters. A daily monitor shows a sudden data-feed outage for one credit bureau attribute, and model confidence falls below the bank’s approved threshold for many applicants. The business unit wants to keep auto-approvals running to avoid service delays while data engineering investigates. What is the BEST action?

• A. Lower the model confidence threshold temporarily so the system can maintain normal approval volumes during the outage.
• B. Continue auto-approvals but add a note that the missing attribute will be restored after the data team completes its investigation.
• C. Wait until the next scheduled model validation cycle before changing the approval workflow.
• D. Activate the approved fallback by suspending affected auto-approvals, routing cases to manual or rules-based review, and logging the issue for escalation.

Best answer: D

What this tests: Responsible and Ethical AI

Explanation: Fallback procedures are needed because an AI system may become unreliable when inputs fail, model performance degrades, or outputs fall below an approved confidence level. In this scenario, the missing credit bureau attribute directly affects the model’s input quality, and monitoring already shows low confidence. Continuing automated approvals would create avoidable credit, fairness, compliance, and customer-impact risks. The best action is to use a preapproved fallback such as manual review or a simpler rules-based process for affected cases, while logging and escalating the issue. This maintains service continuity without relying on outputs that no longer meet safe-use criteria.

• Continuing auto-approvals ignores the evidence that the system is operating outside approved conditions.
• Lowering the confidence threshold masks the control breach and increases reliance on degraded outputs.
• Waiting for scheduled validation is too slow because the issue is active and affects current decisions.

A fallback preserves safe decisioning and accountability when the AI system is degraded or producing low-confidence outputs.

Question 3

Topic: Responsible and Ethical AI

A bank uses an AI model to prioritize small-business hardship assistance requests. Quarterly testing shows overall accuracy of 92%, within the approved risk appetite, but a complaint review finds that applicants from one minority-language community are being incorrectly classified as low hardship at twice the rate of other applicants because translated income documents are often misread. Several affected applicants received delayed assistance. What is the BEST action?

• A. Escalate a fairness issue, review affected cases for remediation, and correct the data and process controls causing the disparity.
• B. Continue using the model because overall accuracy remains within the approved risk appetite.
• C. Retrain the model at the next scheduled cycle without reopening prior decisions.
• D. Update the model documentation to disclose the limitation and continue routine monitoring.

Best answer: A

What this tests: Responsible and Ethical AI

Explanation: Responsible AI governance must consider stakeholder impact, not only aggregate performance metrics. A model can meet overall accuracy targets while still causing disproportionate harm to a subgroup, especially when the affected group is small relative to the full population. Here, there is concrete evidence of harm: a specific community is misclassified at a higher rate because translated documents are misread, and assistance was delayed. The best action is to treat this as a fairness and stakeholder-impact issue, review affected decisions, remediate harmed applicants where appropriate, and fix the underlying data or process control failure. Routine monitoring or future retraining alone is insufficient because it does not address past harm or the active disparity.

• Relying on overall accuracy ignores subgroup harm and fails the fairness objective.
• Future retraining may help model performance but does not remediate applicants already harmed.
• Disclosure and monitoring are useful governance tools, but they are inadequate when a known disparity has already caused adverse impact.

Documented stakeholder harm and a subgroup disparity require targeted remediation even when aggregate model performance is acceptable.

Question 4

Topic: Responsible and Ethical AI

A financial institution’s AI policy requires each high-impact AI system to have a named business owner who approves its intended use, tracks issues, and is answerable to a governance committee for decisions made with the system. Which responsible AI principle is most directly reflected by this requirement?

• A. Fairness
• B. Transparency
• C. Reliability
• D. Accountability

Best answer: D

What this tests: Responsible and Ethical AI

Explanation: Accountability means that responsibility for an AI system’s use, controls, decisions, and impacts is clearly assigned. In the scenario, the key feature is not merely that the model is documented or technically robust, but that a specific business owner approves use, tracks issues, and reports to governance. This creates a line of responsibility for managing the AI system within the organization’s risk appetite. Other responsible AI principles may also be relevant to high-impact AI systems, but the described requirement most directly addresses who is answerable for the system and its consequences.

• Transparency focuses on explainability, disclosure, and understandable information about the AI system, not primarily on ownership.
• Fairness concerns avoiding unjustified bias or discriminatory impacts across groups.
• Reliability concerns consistent, accurate, and dependable performance under intended conditions.

Assigning a named owner who is answerable for AI use and outcomes directly supports accountability.

Question 5

Topic: Responsible and Ethical AI

A bank is preparing to launch an AI tool that recommends credit-line increases. Independent validation found that the tool is less reliable for customers with limited credit history, but the business sponsor asks the team to remove this limitation from the approval package to meet a launch date. What is the best escalation action?

• A. Add a generic user disclaimer stating that AI outputs may be imperfect and proceed with approval.
• B. Document the limitation and the sponsor’s request, refuse to omit the issue, and escalate through the AI governance or model-risk process before launch.
• C. Remove the limitation from the approval package but keep informal notes in the project team’s files.
• D. Launch on schedule and track complaints after deployment because the business sponsor accepts the risk.

Best answer: B

What this tests: Responsible and Ethical AI

Explanation: When a sponsor pressures a team to ignore a known AI limitation, the best action is to protect the integrity of the governance process. The limitation is material because it affects reliability for a customer segment and may create fairness, conduct, and credit-decision risks. The team should document both the validation finding and the pressure to suppress it, refuse to alter the approval record, and escalate through the organization’s AI governance, model-risk, compliance, or ethics channel as appropriate. Business ownership does not permit hiding material risk information from reviewers or approvers.

• Launching because the sponsor accepts the risk fails because accountability still requires transparent review and appropriate controls before use.
• Keeping informal notes while removing the limitation undermines auditability and prevents informed approval.
• A generic disclaimer does not remediate or govern a known limitation affecting decision quality for a customer segment.

Ethical escalation requires preserving material risk information and using the approved governance channel when pressured to ignore an AI limitation.

Question 6

Topic: Responsible and Ethical AI

A retail bank plans to use an AI model to prioritize small-business loan renewals. Pilot results show faster processing and higher retention, but fairness testing indicates that businesses in historically underserved neighborhoods are disproportionately routed to lower-priority queues because location-linked features proxy for limited credit history. Which action best balances the business value with fairness and stakeholder-impact considerations?

• A. Cancel the AI initiative because any disparity in pilot outcomes makes the model unsuitable for business use.
• B. Deploy the model unchanged because the pilot improves retention and does not use explicit protected-class fields.
• C. Remove all location and credit-history features and use a uniform queueing rule for every renewal.
• D. Run a documented impact review, test mitigation options, and deploy only with controls such as human review and segment-level monitoring.

Best answer: D

What this tests: Responsible and Ethical AI

Explanation: A responsible AI action should balance expected business value with the risk of unfair or harmful impacts on stakeholders. The pilot has a clear operational benefit, but it also reveals a disproportionate adverse effect linked to proxy variables. The best response is not to ignore the disparity or abandon the use case automatically. Instead, the bank should document the stakeholder impact, test mitigations, assess whether performance remains acceptable, add human oversight where needed, and monitor outcomes after deployment. This approach supports informed decision-making and accountability while recognizing that fairness concerns often require controls, redesign, and ongoing review rather than a simple yes-or-no decision.

• Improving retention does not justify deployment when fairness testing shows a material disparate impact.
• Canceling the initiative is too absolute; responsible AI can allow controlled use after mitigation and review.
• Removing broad feature categories without testing may reduce model usefulness and does not ensure fairer outcomes.

This preserves the potential business benefit while requiring fairness analysis, mitigation, oversight, and monitoring for affected stakeholders.

Question 7

Topic: Responsible and Ethical AI

A bank operations team wants to use a publicly available generative AI chatbot to summarize customer call transcripts. The transcripts include names, account numbers, and hardship details, and the chatbot terms allow submitted prompts to be retained and used to improve the service. The use case is not yet in the bank’s AI inventory or approved-data list. What is the best action?

• A. Proceed with a small sample because the transcripts are already accessible to operations staff.
• B. Pause the pilot until privacy and AI governance review approve a controlled approach using minimized or masked data and contractual retention and no-training safeguards.
• C. Use the chatbot and monitor the summaries for hallucinations before sharing them internally.
• D. Allow the pilot if employees remove customer names before submitting transcripts.

Best answer: B

What this tests: Responsible and Ethical AI

Explanation: The main concern is not only output accuracy; it is unauthorized disclosure or secondary use of personal, sensitive, and restricted customer data. Removing names alone may not de-identify records when account numbers, hardship details, or other identifiers remain. Because the tool’s terms allow prompt retention and model improvement, submitting transcripts could expose confidential information outside approved controls. The best action is to pause and route the use case through privacy and AI governance, with data minimization, masking or de-identification, approved tooling, access controls, retention limits, and contractual restrictions on vendor use of submitted data.

• Removing only names is insufficient because account numbers and hardship details can still be sensitive or identifying.
• Internal access by operations staff does not authorize sending restricted data to an external AI service.
• Hallucination monitoring addresses output reliability, not the confidentiality risk created by submitting the data.

This directly addresses the personal, sensitive, and restricted-data exposure before any data is submitted to the external AI service.

Question 8

Topic: Responsible and Ethical AI

A bank uses an AI model to rank small-business loan applications for underwriter review. Its responsible AI policy names fairness as a principle, but a recent review found no repeatable step that checks whether rankings differ materially across relevant demographic or proxy segments before deployment or during monitoring. The product owner proposes adding the sentence “The system must be fair” to the model card. What is the best action for the governance team?

• A. Replace the fairness principle with a general accuracy objective for all applications.
• B. Approve the model card update because stating the fairness principle is sufficient governance evidence.
• C. Require a documented fairness control with defined metrics, test timing, accountable owner, escalation criteria, and retained evidence.
• D. Delegate the issue to the vendor because the model’s technical implementation is externally supplied.

Best answer: C

What this tests: Responsible and Ethical AI

Explanation: Responsible AI principles are high-level commitments, such as fairness, transparency, or accountability. They guide expected behavior but do not by themselves prove that the organization has controlled the risk. A control is a specific mechanism that makes the principle actionable and testable: it has an owner, timing, method, evidence, and escalation path. In this scenario, merely adding “the system must be fair” to documentation restates the principle without checking whether the model’s rankings create unfair disparities. The best action is to require a repeatable fairness assessment and monitoring control tied to governance evidence and escalation.

• Stating fairness in a model card documents intent, but it does not create a repeatable or testable control.
• General accuracy does not substitute for fairness because strong average performance can still mask segment-level disparities.
• Vendor involvement does not remove the bank’s accountability for governing how the AI system is used.

A responsible AI principle is operationalized through a specific, repeatable, owned, and evidenced control.

Question 9

Topic: Responsible and Ethical AI

A retail bank is piloting an AI system to prioritize small-business loan applications. Validation shows acceptable overall accuracy, but false-denial rates are materially higher for applicants from a historically underserved region because a location variable acts as a proxy for limited prior credit access. The product team wants to launch because the model meets the profitability target. Which responsible AI principle is most directly implicated?

• A. Privacy protection for personal data used in training
• B. Security of the model deployment environment
• C. Transparency of the model architecture to developers
• D. Fairness and non-discrimination in model outcomes

Best answer: D

What this tests: Responsible and Ethical AI

Explanation: Responsible AI principles often overlap, but the most direct principle depends on the risk evidence in the scenario. Here, the decisive facts are not a data breach, unauthorized access, or inability of developers to inspect the system. The validation result shows materially higher false-denial rates for applicants from a historically underserved region, and the location variable is acting as a proxy for prior credit access limitations. That points to a fairness and non-discrimination concern: the model may produce systematically adverse outcomes for a group even while overall accuracy and profitability appear acceptable. The best responsible AI response would be to assess and mitigate the disparate impact before launch.

• Privacy is not the main issue because the scenario does not describe improper collection, use, sharing, or exposure of personal data.
• Security is not the main issue because there is no indication of unauthorized access, adversarial attack, or deployment vulnerability.
• Transparency may support review, but the decisive concern is the unequal outcome pattern caused by a proxy variable.

The key issue is disproportionate adverse impact on a group due to a proxy variable, so fairness is the principle most directly implicated.

Question 10

Topic: Responsible and Ethical AI

A bank plans to use an AI tool to recommend whether a customer receives a credit limit increase. Call-center agents usually follow the recommendation, and customers who are declined receive only the message, “Request not approved.” Validation also shows weaker performance for customers with limited credit history. What is the best action before deployment to address transparency and disclosure needs?

• A. Publish the model’s source code and training data so customers can independently inspect the system.
• B. Limit disclosure to internal model documentation because the tool is only a recommendation to agents.
• C. Use the existing decline message because detailed explanations could reduce model effectiveness.
• D. Provide plain-language disclosure that AI supports the decision, explain the main factors and known limitations, and offer a path for human review.

Best answer: D

What this tests: Responsible and Ethical AI

Explanation: Transparency in responsible AI requires more than internal documentation when AI materially influences outcomes for people. Here, agents usually follow the model’s recommendation, declined customers are affected, and a known limitation exists for customers with limited credit history. The best action is to disclose AI involvement in understandable language, provide meaningful rationale such as key decision factors, describe relevant limitations, and make escalation or human review available. Transparency should be fit for the audience; it does not require exposing proprietary code or raw training data, but it should enable users and affected parties to understand how AI is involved and what its limits are.

• Treating the tool as “only a recommendation” ignores that agents usually follow it, so the AI materially affects customers.
• Publishing source code or training data is not a practical or privacy-safe substitute for understandable disclosure.
• Keeping a generic decline message fails to explain AI involvement, rationale, or known limitations.

This makes AI involvement, decision rationale, and relevant limitations understandable to users and affected customers.

Continue with full practice

Use the GARP RAI Practice Test page for the full Finance Prep practice bank, mixed-topic practice, timed mock exams, and explanations.

Related focused pages

• Free GARP RAI Full-Length Practice Exam
• GARP RAI: History and Overview of AI Concepts
• GARP RAI: AI Tools and Techniques
• GARP RAI: Risks and Risk Factors
• GARP RAI: Data and AI Model Governance

Free review resource

Use the full Finance Prep practice page above for the latest review links and practice page.