Courseiva

ISACA Advanced in AI Risk (AAIR) (AAIR) (AAIR) — Questions 151–206

206 questions total · 3 pages · All types, answers revealed

Take a mock examExam hub
Page 2

Page 3 of 3

151
MCQmedium

When reporting to the Board, what is the best way to present 'AI Risk'?

A.State that AI is completely safe and requires no oversight.
B.Provide a list of all model parameters.
C.Link AI risks to overall business impact, strategic goals, and current risk appetite.
D.Give a 4-hour technical lecture on deep learning.
AnswerC

Boards govern through financial and strategic consequence, not technical detail. Expressing AI risk as business impact tied to strategic goals and the stated risk appetite lets directors compare it against other enterprise risks and prioritise investment, satisfying the reporting audience constraint.

Why this answer

Board-level reporting requires translating technical AI risk into business language. Option C does this by connecting AI risks to overall business impact, strategic goals, and the organization's risk appetite—the three dimensions boards care about most. This approach enables informed governance decisions, resource allocation, and risk tolerance setting, which are the board's actual responsibilities.

It aligns with AAIR's emphasis on AI governance being integrated into enterprise risk management rather than treated as a purely technical concern.

Exam trap

AAIR often tests the distinction between technical AI details and business-oriented governance communication, trapping candidates who confuse model transparency with board-level risk reporting.

How to eliminate wrong answers

Option A is wrong because claiming AI is completely safe and requires no oversight is factually incorrect and dangerously naive; all AI systems carry risks such as bias, drift, and adversarial vulnerability, and governance frameworks like AAIR explicitly require oversight. Option B is wrong because listing all model parameters is a technical detail that provides no meaningful risk or business context to a board; it confuses model transparency with governance reporting and fails to address impact, likelihood, or mitigation. Option D is wrong because a 4-hour technical lecture on deep learning is inappropriate for a board audience; it focuses on how models work rather than the business risks and strategic implications, and it ignores the board's need for concise, decision-oriented information.

Full explanation →
152
MCQeasy

What is the primary risk objective of 'Model Versioning' in the AI lifecycle?

A.To prevent unauthorized users from viewing the code.
B.To reduce the amount of disk space used.
C.To increase the model's speed in production.
D.To ensure auditability, reproducibility, and the ability to roll back to a known stable state.
AnswerD

Versioning records each model artefact with its training data and parameters, so any output can be traced, regenerated exactly, and reverted when a release degrades. This directly satisfies the stem's auditability, reproducibility, and rollback objective across the AI lifecycle.

Why this answer

Model versioning exists to provide auditability (who changed what and when), reproducibility (recreate any past model and its results), and rollback capability (revert to a known stable version when a new release degrades). These properties are foundational for incident response, regulatory review, and safe continuous delivery of AI systems.

Exam trap

AAIR often tests whether candidates confuse versioning with security or performance; the trap is selecting access control or speed because those sound like lifecycle benefits, when versioning's core objective is auditability, reproducibility, and rollback.

How to eliminate wrong answers

Option A is wrong because access control is handled by IAM, authentication, and authorization — versioning does not prevent unauthorized viewing. Option B is wrong because versioning typically increases storage consumption by retaining multiple artifacts, not reducing disk usage. Option C is wrong because inference speed is determined by model architecture, hardware, and optimization (quantization, compilation), not by version control.

Full explanation →
153
MCQhard

A hospital deploys a sepsis-prediction model. The AI risk manager learns that the vendor retrained the model on new patient data but the hospital's integration still calls the same endpoint, and no notification was issued. Which lifecycle risk control failure does this most directly represent?

A.Inadequate consent management for secondary use of patient data.
B.Failure to perform a fairness assessment on protected patient subgroups.
C.Absence of model change management and version traceability for a third-party AI component.
D.Insufficient compute capacity to run the updated model in the clinical environment.
AnswerC

A silent retrain behind an unchanged endpoint breaks version traceability. The hospital can no longer map clinical decisions to a specific model version, validate the change, or reproduce past outputs during an incident investigation. For third-party AI, change notification and version pinning are core lifecycle controls, and their absence is the direct failure described.

Why this answer

When a vendor retrains a model and serves it through the same endpoint without notification, the deploying organisation loses the ability to identify which model version produced a given clinical output. That undermines incident investigation, validation, and regulatory traceability. Effective AI lifecycle risk management requires contractual change notification, version pinning, and revalidation triggers for third-party model updates.

Exam trap

The trap here is treating any unexpected model behaviour as a fairness or consent problem, when the scenario's defining fact is an unannounced version change behind a stable interface.

Full explanation →
154
MCQhard

What is the 'Accountability' principle in the NIST AI RMF intended to address?

A.Ensuring the AI can explain itself.
B.Ensuring that there is clear ownership and responsibility for AI system performance and safety.
C.Making sure the model is free of bugs.
D.Ensuring that the AI system is fast.
AnswerB

Accountability assigns clear ownership for AI system performance and safety across the lifecycle, ensuring identifiable parties answer for outcomes. This addresses the governance gap where harms occur but no role owns remediation, complementing the RMF's other trustworthiness characteristics.

Why this answer

The Accountability principle in the NIST AI RMF (part of the Govern function) is about ensuring clear ownership and responsibility for AI system performance, safety, and outcomes throughout the lifecycle. It answers 'who is responsible?' so that risks are managed, issues are escalated, and redress is possible. It is not about explainability, bug-freeness, or speed.

Exam trap

The trap is confusing Accountability with Explainability or general quality attributes — candidates often pick 'the AI can explain itself' because both relate to trust, but accountability is about human ownership and responsibility, not model transparency.

How to eliminate wrong answers

Option A is wrong because ensuring the AI can explain itself is the Explainability/Interpretability principle, which is related but distinct from Accountability — a model can be explainable yet have no clear owner. Option C is wrong because making sure the model is free of bugs is a software quality/validation concern, not the accountability principle; no model is ever fully bug-free, and accountability is about ownership of residual risk. Option D is wrong because ensuring the AI system is fast is a performance requirement, entirely unrelated to accountability, which concerns responsibility and governance.

Full explanation →
155
MCQeasy

Which of the following is a key component of 'AI Risk Governance'?

A.The choice of programming language
B.The definition of roles and responsibilities
C.The vendor of the cloud hosting platform
D.The speed of model deployment
AnswerB

Defining roles and responsibilities establishes accountability for AI risk decisions, satisfying governance's requirement for clear ownership across the AI lifecycle. Unlike purely technical controls, this assigns who assesses, approves, and monitors AI risks within Microsoft Entra ID-governed environments, ensuring oversight is enforceable rather than implicit.

Why this answer

A key component of AI Risk Governance is the definition of roles and responsibilities, which establishes accountability for risk management across the organization. This includes specifying who owns AI risks, who approves deployments, and who monitors ongoing performance. Without clear roles, governance breaks down and risks go unmanaged.

Exam trap

The trap is confusing technical implementation choices (programming language, cloud vendor) with governance components, causing candidates to overlook the organizational nature of roles and responsibilities.

How to eliminate wrong answers

Option A is wrong because the choice of programming language is an implementation detail that does not define governance structures or accountability. Option C is wrong because the cloud hosting vendor is a procurement and technical decision, not a core governance component, though it may be subject to governance oversight. Option D is wrong because the speed of model deployment is an operational metric, not a governance component; governance may influence speed through controls, but speed itself is not a component.

Full explanation →
156
MCQmedium

When establishing KPIs for an AI program, what is the risk of focusing solely on 'Model Accuracy'?

A.The model will be too computationally expensive.
B.The data scientists will take too long to develop the model.
C.The model will not be deployable in production.
D.Other critical risks such as bias, security, and fairness may be overlooked.
AnswerD

Optimising only for Model Accuracy ignores other dimensions of responsible AI. Bias, security and fairness can degrade undetected, so the programme may ship a model that performs well statistically yet causes harm, which is the risk the stem asks about.

Why this answer

Focusing solely on model accuracy creates a narrow optimization target that ignores other critical AI risks such as bias, fairness, security, privacy, and robustness. Accuracy alone does not guarantee that a model is ethical, compliant, or safe in production, so other risk dimensions can be neglected.

Exam trap

AAIR often tests the misconception that accuracy is a sufficient KPI, causing candidates to overlook bias, fairness, and security risks.

How to eliminate wrong answers

Option A is wrong because computational expense is a performance/cost concern, not the primary risk of an accuracy-only KPI. Option B is wrong because development time is a schedule concern, not a governance risk arising from accuracy-centric measurement. Option C is wrong because deployability depends on many factors (latency, integration, monitoring), and accuracy alone does not directly cause non-deployability; the core issue is overlooked risk dimensions.

Full explanation →
157
MCQhard

A government agency uses an AI system to triage benefit applications. The risk register lists 'automation bias' as a concern. Which monitoring metric would most directly evidence whether this risk is materialising in practice?

A.The number of benefit applications processed per caseworker per day after the AI system was introduced.
B.The average latency between the AI producing a recommendation and the caseworker recording a decision.
C.The overall accuracy of the AI triage model measured against a gold-standard labelled sample.
D.The rate at which caseworkers accept the AI's triage recommendation without modifying it, broken down by recommendation confidence.
AnswerD

Automation bias appears when humans defer to a system even when it is wrong or uncertain. Measuring unmodified acceptance rates against confidence levels reveals whether reviewers rubber-stamp high-confidence outputs and ignore their own judgment. A high acceptance rate on low-confidence recommendations is a strong signal that the bias is present and material.

Why this answer

Automation bias is a human-factors risk, so the evidence must come from the human-AI interaction. Tracking how often caseworkers accept the triage recommendation without change, segmented by model confidence, shows whether reviewers defer even when the system is uncertain. A high unmodified acceptance rate on low-confidence outputs indicates the bias is materially affecting decisions and warrants mitigation.

Exam trap

The trap here is measuring the model itself, such as accuracy or throughput, when automation bias is a property of how humans respond to the model's recommendations.

Full explanation →
158
MCQmedium

In the context of AI Supply Chain risk, what is the primary governance objective when evaluating a vendor's AI model transparency?

A.To minimize the number of API calls made to the vendor.
B.To verify the vendor's financial stability and market share.
C.To ensure the vendor's training data includes proprietary internal company data.
D.To confirm the vendor's model satisfies the internal 'Model Card' and transparency requirements for risk assessment.
AnswerD

The governance objective is verifying that the vendor's model documentation meets internal Model Card and transparency requirements, enabling informed risk assessment of the AI supply chain. This satisfies the stem's focus on evaluating model transparency as part of third-party risk governance.

Why this answer

The primary governance objective when evaluating a vendor's AI model transparency is to ensure the vendor provides sufficient documentation — such as model cards, datasheets, and transparency reports — to assess risks related to bias, explainability, and compliance. Confirming that the vendor's model satisfies internal Model Card and transparency requirements enables the organization to perform due diligence and integrate the model into its risk management framework.

Exam trap

AAIR often tests the confusion between transparency requirements and unrelated vendor evaluation criteria like cost, financial health, or data inclusion, leading candidates to select operational or commercial metrics instead of governance-focused ones.

How to eliminate wrong answers

Option A is wrong because minimizing API calls is a cost or performance optimization, not a governance objective related to transparency. Option B is wrong because financial stability and market share are vendor viability concerns, not directly related to AI model transparency or risk assessment. Option C is wrong because including proprietary internal company data in the vendor's training set would be a data leakage risk, not a governance objective; transparency requires understanding what data was used, not ensuring internal data is included.

Full explanation →
159
MCQeasy

What is 'Bias' in the context of AI risk?

A.The model is missing documentation
B.Systematic prejudice in model outputs leading to unfair results
C.The model is too expensive to maintain
D.The model is running too fast
AnswerB

Bias denotes systematic, repeatable skew in model outputs that disadvantages particular groups, rather than random error. This satisfies the stem's AI risk context by naming prejudice embedded in outputs, producing unfair results across protected characteristics. It aligns with AAIR's treatment of bias as a measurable fairness harm requiring mitigation.

Why this answer

In AI risk taxonomy, bias refers to systematic and unfair prejudice in model outputs that disadvantages individuals or groups based on protected attributes like race, gender, or age. It arises from skewed training data, flawed feature selection, or proxy variables, and manifests as disparate impact in decisions such as hiring, lending, or healthcare. It is a fairness risk, not a documentation, cost, or performance issue.

Exam trap

AAIR often tests the misconception that bias is a technical performance or cost issue — the trap is failing to recognize bias as a fairness risk rooted in systematic output prejudice against protected groups.

How to eliminate wrong answers

Option A is wrong because missing documentation is a model governance and explainability deficiency, not bias — it is a documentation risk, not a fairness risk. Option C is wrong because high maintenance cost is a financial/operational risk, unrelated to fairness in model outputs. Option D is wrong because running too fast is a performance characteristic; speed has no bearing on whether outputs are systematically unfair to a protected group.

Full explanation →
160
MCQhard

You are auditing the integration of AI risk into the Enterprise Risk Management (ERM) system. Which finding indicates a failure in the governance structure?

A.The risk assessment template is 20 pages long.
B.The AI risk register is updated annually instead of monthly.
C.The AI risk committee meets quarterly rather than monthly.
D.AI models deployed in production are missing from the centralized enterprise risk register.
AnswerD

The enterprise risk register is the central artefact through which governance bodies aggregate, prioritise and monitor risk. Production AI models absent from it cannot be governed, reported or escalated, so this finding directly evidences a breakdown in the ERM governance structure rather than a technical control weakness.

Why this answer

A centralized enterprise risk register is the authoritative inventory of organizational risks; if production AI models are absent from it, the governance structure has failed to integrate AI risk into ERM — meaning those models are unmanaged, unreviewed, and invisible to executive and Board oversight. This is a structural governance failure because it breaks the linkage between AI risk identification and enterprise-level risk aggregation, violating the principle that all material risks must flow into the ERM.

Exam trap

AAIR often tests the distinction between process maturity issues (meeting frequency, template length, update cadence) and genuine structural governance failures (AI models missing from the enterprise risk register).

How to eliminate wrong answers

Option A is wrong because a 20-page risk assessment template is a documentation length issue, not a governance failure — thoroughness in templates is generally desirable and does not indicate broken oversight. Option B is wrong because annual risk register updates may be insufficient for fast-moving AI but the frequency alone is a process maturity concern, not a structural governance failure; many organizations update registers quarterly or on a trigger basis. Option C is wrong because quarterly committee meetings can be adequate depending on the organization's AI risk profile and materiality — meeting frequency is a cadence choice, not evidence of a broken governance structure.

Full explanation →
161
MCQeasy

A board of directors requests a dashboard view of AI model drift. Which metric is most appropriate for a board-level risk report?

A.Feature drift threshold violations
B.Mean Squared Error (MSE)
C.GPU utilization percentage
D.Latency per inference request
AnswerA

Feature drift threshold violations express model degradation as a count of breached statistical limits, giving the board a quantified, governance-level risk indicator. This satisfies the stem's requirement for a board-appropriate metric rather than operational or engineering-level detail.

Why this answer

Feature drift threshold violations are the most appropriate board-level metric because they directly signal when input data distributions have shifted beyond acceptable bounds, which is the leading indicator of model risk. Boards need a governance-level signal of emerging model unreliability, not raw performance numbers. Threshold violations translate technical drift into a risk posture that non-technical directors can act on.

Exam trap

AAIR often tests the distinction between technical ML metrics (MSE, latency) and governance-level risk indicators, causing candidates to pick a familiar performance metric instead of a risk-reporting metric.

How to eliminate wrong answers

Option B is wrong because MSE is a model performance metric that measures prediction error on a dataset, not a governance-level drift indicator, and it requires technical interpretation. Option C is wrong because GPU utilization is an infrastructure capacity metric unrelated to model risk or drift. Option D is wrong because latency per inference is an operational performance metric that reflects serving efficiency, not model behavior degradation.

Full explanation →
162
MCQhard

A healthcare organization is implementing the NIST AI Risk Management Framework (AI RMF) and wants to ensure that its AI governance structure addresses the 'Govern' function's requirement for clear roles and responsibilities. The Chief AI Officer proposes several actions. Which action best satisfies the requirement for accountability and oversight?

A.Creating a volunteer-based AI ethics council without formal authority or reporting lines to the board.
B.Outsourcing AI risk oversight to an external consulting firm to gain objective expertise.
C.Assigning the Chief Data Scientist as the sole owner of all AI risk decisions to ensure technical expertise.
D.Establishing an AI Governance Committee with representation from legal, compliance, business units, and data science, and defining a RACI matrix for AI risk decisions.
AnswerD

Establishing an AI Governance Committee with cross-functional representation and a RACI matrix directly addresses the Govern function's call for clear roles and responsibilities. It ensures diverse perspectives, avoids conflicts of interest, and defines accountability for AI risk decisions. This structure promotes oversight and enables effective challenge. The RACI matrix clarifies who is responsible, accountable, consulted, and informed, which is essential for governance.

Why this answer

The correct action is to establish an AI Governance Committee with cross-functional representation and a RACI matrix. This creates clear roles and responsibilities, ensures diverse oversight, and defines accountability for AI risk decisions. It aligns with the AI RMF Govern function's requirement for clear roles and responsibilities, promoting effective challenge and integration with enterprise governance.

Exam trap

The trap here is equating technical expertise with governance authority, overlooking the need for independent oversight and diverse representation.

Full explanation →
163
MCQmedium

When coordinating between Legal and Data Science teams, what is the best way to handle 'Explainability' requirements?

A.Require Data Science to translate technical model output into plain language documentation
B.Restrict Legal from reviewing AI model outputs
C.Let Legal define the technical parameters of the models
D.Only document model logic when a lawsuit occurs
AnswerA

Explainability requires that non-technical stakeholders understand model reasoning, so having Data Science translate technical outputs into plain-language documentation bridges the Legal and Data Science gap. This satisfies the coordination requirement by making model behaviour auditable and comprehensible to Legal reviewers.

Why this answer

Explainability requirements exist so that non-technical stakeholders — Legal, compliance, regulators, and affected individuals — can understand how a model reaches decisions. The Data Science team owns the technical model and is best positioned to translate its behavior into plain-language documentation that Legal can review and rely on. This preserves the division of labor: Data Science explains, Legal interprets and applies.

Exam trap

AAIR often tests role boundaries in AI governance — candidates may pick 'Legal defines technical parameters' or 'restrict Legal access' by confusing who owns explainability versus who consumes it; the correct answer keeps Data Science as translator and Legal as reviewer.

How to eliminate wrong answers

Option B is wrong because restricting Legal from reviewing AI outputs defeats the purpose of explainability and creates regulatory and legal exposure. Option C is wrong because Legal defining technical model parameters inverts the roles — Legal lacks the technical expertise and should not dictate model architecture or hyperparameters. Option D is wrong because documenting model logic only when a lawsuit occurs is reactive and violates proactive governance and regulatory expectations for high-risk AI.

Full explanation →
164
MCQmedium

A retail bank's credit-scoring AI was trained on three years of historical loan data. During an ISACA AAIR-aligned risk review, the risk practitioner discovers that the model's training dataset included the final repayment outcomes of loans that were still open at the time of data extraction. Which risk is most directly introduced by this dataset construction?

A.Sampling bias, because the three-year window excludes applicants who never received a loan.
B.Target leakage, because repayment outcomes encoded future information that would not be available at inference time.
C.Concept drift, because credit behaviour changes over the three-year training window.
D.Label noise, because repayment outcomes were recorded inconsistently across legacy systems.
AnswerB

Repayment outcomes for loans still open at extraction are future events relative to the scoring decision. Including them as features or labels creates target leakage, letting the model learn signals unavailable at scoring time. This inflates validation metrics and produces a model that silently degrades once deployed, which is exactly the risk an AI lifecycle risk review must flag.

Why this answer

The dataset includes repayment outcomes from loans that were still open when data was extracted, so the model sees information that would not exist at the moment a credit decision is made. That temporal contamination is target leakage, and it inflates validation performance while guaranteeing degradation in production. A lifecycle risk review must detect it during data preparation, before the model is approved for deployment.

Exam trap

The trap here is assuming any dataset defect involving historical records is sampling bias or concept drift, when the defining issue is that future outcome information was embedded in the training data.

Full explanation →
165
MCQeasy

When setting KPIs for an AI risk program, which metric is a leading indicator of potential model bias?

A.Average model inference latency
B.Percentage of customer complaints regarding AI decisions
C.Number of model updates per quarter
D.Demographic parity ratio in training data samples
AnswerD

Demographic parity ratio measures outcome rates across protected groups in sampled training data, so an imbalanced ratio precedes biased predictions. It is a leading indicator because it flags imbalance before training, unlike lagging production fairness metrics.

Why this answer

Demographic parity ratio in training data samples is a leading indicator because it measures the representation of protected groups before the model is deployed. A skewed ratio (e.g., far from 1.0) signals that the model may learn biased patterns, allowing proactive mitigation. Unlike lagging metrics such as complaints, it provides early warning during data collection and preprocessing.

Exam trap

The trap here is confusing leading indicators (proactive, data-centric metrics) with lagging indicators (reactive, outcome-based metrics like complaints), causing candidates to select option B.

How to eliminate wrong answers

Option A is wrong because average model inference latency is a performance metric unrelated to bias; it measures speed, not fairness. Option B is wrong because the percentage of customer complaints regarding AI decisions is a lagging indicator—it reflects bias after harm has occurred, not a leading signal. Option C is wrong because the number of model updates per quarter indicates development velocity or maintenance frequency, not bias risk.

Full explanation →
166
MCQmedium

When performing an AI risk assessment, which question is most appropriate to ask?

A.How many people are on the development team?
B.What is the potential impact if the model produces biased results?
C.What is the model's highest accuracy score?
D.Does the model use open-source libraries?
AnswerB

Bias is a core AI-specific risk dimension, so asking about the impact of biased model outputs addresses fairness and potential harm to affected individuals. This frames the assessment around a concrete consequence of model behaviour, which is the appropriate focus for AI risk analysis.

Why this answer

The most appropriate risk-assessment question probes potential harm and impact, such as the consequences of biased model outputs. This focuses the assessment on likelihood and severity of adverse outcomes, which is the core of risk analysis. Questions about team size, accuracy scores, or library choices are inputs or metrics, not risk-impact questions.

Exam trap

AAIR often tests the misconception that technical metrics like accuracy equal risk assessment, when the correct focus is potential harm and impact.

How to eliminate wrong answers

Option A is wrong because development team size is a resourcing detail, not a measure of risk likelihood or impact. Option C is wrong because a single accuracy score does not capture harm, fairness, safety, or business impact and can mask severe failure modes on underrepresented data. Option D is wrong because using open-source libraries is a supply-chain consideration, not a direct risk-impact question; it may inform a sub-risk but does not assess the consequence of model behavior.

Full explanation →
167
MCQhard

An organization is revising its AI Governance to include 'Human-in-the-loop' (HITL) requirements. Why is this a critical risk mitigation strategy?

A.It removes the requirement for model bias testing.
B.It is required for every AI model regardless of impact.
C.It ensures human oversight for critical decisions, reducing the risk of 'black box' failures.
D.It eliminates the need for any other automated risk controls.
AnswerC

Human-in-the-loop places a person in the decision path for consequential outputs, so opaque model reasoning is reviewed before action. This directly mitigates black-box failure risk, which the stem identifies as the governance concern driving the HITL requirement.

Why this answer

Human-in-the-loop (HITL) is a critical risk mitigation because it places a human reviewer in the decision path for consequential AI outputs, providing a check against opaque 'black box' failures, edge cases, and harmful errors that automated controls may miss. It preserves meaningful human judgment and accountability where the cost of an AI mistake is high. It does not replace other controls or eliminate testing obligations.

Exam trap

The trap is treating HITL as a silver bullet that replaces other controls (bias testing, monitoring) or as universally mandatory — the exam tests whether you understand it is a risk-based, complementary control, not a substitute for the rest of the governance framework.

How to eliminate wrong answers

Option A is wrong because HITL does not remove the need for model bias testing — bias testing is a separate, mandatory control that must continue regardless of human oversight, and humans themselves can introduce bias. Option B is wrong because HITL is not required for every AI model; it is risk-based and typically reserved for high-impact or high-autonomy use cases, while low-risk models may operate without it. Option D is wrong because HITL is one control among many and does not eliminate the need for automated monitoring, logging, drift detection, access controls, or other safeguards — defense in depth still applies.

Full explanation →
168
MCQmedium

During the design phase of a machine learning model, a practitioner identifies that the training dataset has significant underrepresentation of a minority demographic. What is the most effective proactive risk mitigation strategy?

A.Deploy the model with a human-in-the-loop oversight mechanism.
B.Apply synthetic data generation techniques to balance the training distribution.
C.Monitor prediction confidence intervals during the post-deployment phase.
D.Increase the regularization parameters in the model architecture.
AnswerB

Synthetic data generation augments the minority class, rebalancing the training distribution before the model learns. This directly counteracts underrepresentation, reducing bias and improving performance for that demographic — a proactive fix at the data layer rather than a post-hoc correction.

Why this answer

When a training dataset underrepresents a demographic, the most effective proactive mitigation at the design phase is to correct the data distribution itself — synthetic data generation (e.g., SMOTE, CTGAN, or targeted augmentation) can increase representation of the minority group so the model learns its patterns rather than defaulting to the majority. Addressing the root cause in the data is preferable to downstream patches because it improves model behavior for the affected group across all future inferences.

Exam trap

AAIR often tests whether candidates choose a root-cause data remedy over downstream controls; the trap is selecting human-in-the-loop or monitoring because they sound responsible, when the question asks for the most effective proactive design-phase mitigation.

How to eliminate wrong answers

Option A is wrong because human-in-the-loop oversight is a downstream control that catches some errors but does not fix the model's learned bias and is not a design-phase data remedy. Option C is wrong because monitoring confidence intervals post-deployment is detective, not proactive, and confidence is not a reliable proxy for fairness — a biased model can be confidently wrong. Option D is wrong because increasing regularization reduces overfitting/variance but does nothing to correct a skewed class or demographic distribution; it may even suppress minority-group signal further.

Full explanation →
169
MCQmedium

In a retail company, which AI risk is most likely to cause direct financial loss?

A.The model documentation is outdated
B.The font size in the UI is too small
C.The pricing model provides incorrect, discounted rates
D.The model's name is not catchy
AnswerC

Discounted pricing directly reduces revenue per transaction, so every affected sale produces immediate, quantifiable margin loss. Other AI risks, such as bias or poor explainability, cause harm indirectly through legal or reputational channels rather than direct financial impact.

Why this answer

The AI risk most likely to cause direct financial loss in a retail company is a pricing model that provides incorrect, discounted rates. This directly impacts revenue and profitability, as customers may be undercharged or the company may lose sales due to pricing errors. Incorrect pricing can also lead to competitive disadvantages and customer distrust.

Exam trap

AAIR often tests the ability to distinguish between direct and indirect financial risks. Candidates may be distracted by governance or usability issues, but the most direct financial loss comes from core business functions like pricing.

How to eliminate wrong answers

Option A is wrong because outdated model documentation is a governance issue that may lead to inefficiencies but does not directly cause financial loss. Option B is wrong because font size in the UI is a usability concern, not a direct financial risk. Option D is wrong because the model's name not being catchy is a marketing issue with no direct financial impact.

Full explanation →
170
MCQmedium

When an AI model is updated (e.g., retrained on new data), what action is required from an AI Risk Governance perspective?

A.The model should be renamed to show it is a new version.
B.No action is needed as the underlying architecture is unchanged.
C.The model should be re-validated to ensure it still meets risk thresholds.
D.The model should be automatically deployed without review.
AnswerC

Retraining alters a model's weights, outputs and failure modes, so prior validation evidence no longer describes the deployed system. Re-validation against the original risk thresholds confirms the updated model still performs within approved tolerances, satisfying the governance requirement that material model changes trigger fresh assessment before continued use.

Why this answer

Retraining a model on new data changes its behavior, so from a risk governance perspective it must be re-validated against the same risk thresholds (accuracy, fairness, robustness, compliance) before being trusted in production. Renaming or assuming no action is needed ignores the fact that the model's outputs and risk profile may have changed. Automatic deployment without review bypasses required oversight.

Exam trap

AAIR often tests whether candidates recognize that retraining is a material change requiring re-validation — the trap is assuming that unchanged architecture or a simple rename satisfies governance requirements.

How to eliminate wrong answers

Option A is wrong because renaming the model is a cosmetic versioning step and does not satisfy the substantive requirement to re-assess its risk and performance. Option B is wrong because unchanged architecture does not mean unchanged behavior — new training data can alter predictions, bias, and failure modes, so re-validation is required. Option D is wrong because automatically deploying a retrained model without review violates change-management and model-risk governance controls.

Full explanation →
171
Multi-Selecteasy

Which TWO factors should be documented in an AI Model Card?

Select 2 answers
A.The developer's favorite snacks
B.The current stock price of the company
C.The employee handbook for new hires
D.Intended use cases
E.Model limitations and known risks
AnswersD, E

Model Cards record intended use cases, defining the tasks and contexts the model was designed and evaluated for, which guides appropriate deployment. This satisfies the stem's requirement for factors that must be documented in an AI Model Card.

Why this answer

Option D (Intended use cases) is correct because an AI Model Card is designed to document the purpose and appropriate applications of a model, so consumers know what tasks it was built for and where it should be deployed. Option E (Model limitations and known risks) is correct because Model Cards must disclose known biases, failure modes, out-of-scope scenarios, and other risks so users can make informed decisions about deployment. Options A, B, and C are incorrect because favorite snacks, stock prices, and new-hire handbooks are unrelated to model transparency, performance, or governance, and are not part of standard Model Card documentation such as that proposed by Mitchell et al. or required under emerging AI regulations.

Exam trap

AAIR often tests whether candidates confuse Model Card contents with general corporate documents, so distractors like handbooks or stock prices are inserted to see if you know the Model Card is strictly a model-level transparency artifact.

Full explanation →
172
MCQmedium

A board of directors requests a dashboard to monitor AI risk exposure. Which metric provides the most effective board-level oversight regarding AI regulatory compliance?

A.Percentage of critical AI systems that have undergone a formal regulatory compliance impact assessment.
B.Mean Time to Detect (MTTD) for model adversarial attacks.
C.Current GPU utilization rates for production inference clusters.
D.Average latency of API requests for generative AI models.
AnswerA

This metric directly evidences whether each critical AI system has been assessed against applicable regulations, giving the board measurable coverage of compliance obligations rather than operational or performance indicators. It quantifies exposure from unassessed systems, supporting governance decisions on regulatory risk.

Why this answer

The board's primary concern is regulatory compliance and risk exposure, not operational performance. The percentage of critical AI systems that have undergone a formal regulatory compliance impact assessment directly measures the organization's adherence to frameworks like the EU AI Act, NIST AI RMF, or ISO/IEC 42001. This metric provides assurance that high-risk AI systems are being evaluated against legal and ethical requirements, which is a core board-level oversight responsibility.

Exam trap

AAIR often tests the distinction between operational metrics (MTTD, latency, utilization) and governance metrics (compliance assessment coverage), so candidates must recognize that board-level oversight focuses on regulatory adherence, not technical performance.

How to eliminate wrong answers

Option B is wrong because MTTD for adversarial attacks is a security operations metric that measures detection speed, not regulatory compliance posture. Option C is wrong because GPU utilization is an infrastructure efficiency metric, irrelevant to compliance oversight. Option D is wrong because API latency is a performance metric that does not reflect adherence to AI regulations or risk management frameworks.

Full explanation →
173
Multi-Selectmedium

Which THREE areas should be covered in an AI risk management program?

Select 3 answers
A.The office floor plan
B.Model performance and accuracy
C.Third-party/Vendor risk management
D.The color of the company logo
E.Legal and regulatory compliance
AnswersB, C, E

Model performance and accuracy must be covered because degraded, drifting or biased outputs create operational and safety risk. Monitoring these metrics ensures the programme detects failures, satisfying the stem's requirement that risk management addresses the model's ongoing behaviour rather than only deployment-time concerns.

Why this answer

Option B (Model performance and accuracy) is correct because an AI risk management program must address model-specific risks such as drift, bias, overfitting, and degraded accuracy that directly affect decision quality and safety. Option C (Third-party/Vendor risk management) is correct because organizations often rely on external model providers, APIs, and training data suppliers, so vendor due diligence, contractual controls, and supply-chain monitoring are essential parts of AI risk governance. Option E (Legal and regulatory compliance) is correct because AI systems must comply with applicable laws and frameworks such as the EU AI Act, GDPR, and sector-specific regulations, making compliance a core risk domain.

Option A (The office floor plan) is incorrect because physical office layout is a facilities concern, not an AI risk management area. Option D (The color of the company logo) is incorrect because branding aesthetics have no bearing on AI risk management.

Exam trap

AAIR often tests whether candidates can distinguish genuine AI risk domains (performance, vendor, compliance) from obvious distractors (office layout, logo color) that have no risk relevance.

Full explanation →
174
MCQeasy

A retail company's AI governance committee is drafting its charter. The committee wants to define clear decision rights so that disputes over AI risk acceptance do not stall deployment decisions. Which element is most essential to include in the charter?

A.The marketing roadmap for AI-powered customer experiences over the next three years.
B.A detailed technical specification of every AI model architecture the company plans to use.
C.A complete list of every AI vendor contract currently under negotiation.
D.A RACI matrix that assigns accountability for AI risk decisions by risk tier and defines the escalation path to the board.
AnswerD

A RACI matrix tied to risk tiers establishes who is accountable, consulted, and informed for AI risk decisions, and the defined escalation path resolves disagreements at the right level. This directly addresses the committee's goal of preventing disputes from stalling deployments by making decision rights explicit and auditable.

Why this answer

A governance charter exists to make authority explicit. A RACI matrix that maps accountability by risk tier, combined with a defined escalation path to the board, ensures that disagreements about AI risk acceptance are resolved by the right decision maker at the right level. This prevents deployment delays caused by ambiguity over who ultimately owns the risk decision.

Exam trap

The trap here is filling a governance charter with operational or strategic content, such as architecture details or roadmaps, when the charter's core purpose is to assign decision rights and escalation authority.

Full explanation →
175
MCQhard

To mitigate 'Training-Serving Skew' in the AI lifecycle, what is the best practice for data processing pipelines?

A.Use a feature store to ensure consistent feature engineering logic for both stages.
B.Write separate code for training and inference to optimize for speed.
C.Use only raw data for both training and inference.
D.Train the model on the production server directly.
AnswerA

A feature store centralises feature engineering logic so training and serving compute identical transformations from the same definitions, eliminating the divergent preprocessing that causes training-serving skew. Without shared logic, pipelines drift apart and predictions degrade in production.

Why this answer

Training-serving skew occurs when the feature computation logic differs between the training pipeline and the inference pipeline, producing inconsistent feature values. A feature store (A) centralizes feature definitions and serves the same computed values to both training and online inference, eliminating the divergence. This is the industry-standard mitigation because it enforces a single source of truth for feature engineering.

Exam trap

The trap is confusing 'optimize for speed' or 'use raw data' with consistency; the exam wants you to recognize that skew is caused by divergent transformation logic, and only a shared feature definition (feature store) fixes it.

How to eliminate wrong answers

Option B is wrong because writing separate code for training and inference is the direct cause of training-serving skew, not a mitigation — divergent implementations inevitably produce different feature values. Option C is wrong because using only raw data for both stages ignores the need for consistent transformations; raw data still requires identical preprocessing at both ends, and skipping feature engineering is not a viable mitigation. Option D is wrong because training on the production server does not align feature logic and introduces operational risk; it conflates environment with pipeline consistency.

Full explanation →
176
MCQeasy

What is the primary role of the AI 'Data Steward' within the AI Governance framework?

A.To oversee the lifecycle of the data, ensuring its quality and compliance.
B.To manage the budget for AI hardware.
C.To build the model architecture.
D.To handle marketing and customer relations.
AnswerA

The Data Steward owns data quality, lineage and compliance across the AI lifecycle, ensuring training and inference data meets regulatory and organisational standards. This satisfies the governance framework's need for accountability over data assets, distinct from model validation or ethical review roles.

Why this answer

The AI Data Steward owns the data lifecycle within the AI governance framework — ensuring datasets are accurate, complete, appropriately sourced, and compliant with privacy and licensing requirements throughout collection, use, retention, and disposal. This role is the data-side counterpart to model governance, ensuring that what feeds the model is trustworthy and lawful. Budget, architecture, and marketing responsibilities belong to entirely different roles.

Exam trap

The trap is conflating the Data Steward with the ML Engineer or Data Scientist — AAIR tests whether candidates recognize that the steward governs data quality and compliance across the lifecycle, not model construction or infrastructure.

How to eliminate wrong answers

Option B is wrong because AI hardware budgeting is a financial/IT procurement function, not a data governance responsibility — the steward does not control capital expenditure. Option C is wrong because building model architecture is the job of data scientists and ML engineers, not the data steward, whose focus is the data feeding the model rather than the model itself. Option D is wrong because marketing and customer relations are business functions unrelated to data lifecycle governance, quality assurance, or regulatory compliance.

Full explanation →
177
MCQmedium

Your organization is integrating AI risks into the existing COSO Enterprise Risk Management (ERM) framework. Which action best ensures that AI-specific model drift risk is formally addressed within the 'Review and Revision' component?

A.Integrating model performance metrics into the quarterly enterprise risk assessment and management review processes.
B.Establishing automated triggers within the CI/CD pipeline to halt deployment if drift exceeds threshold.
C.Mapping AI model inventory to the IT asset register in the CMDB.
D.Assigning a Data Scientist to the Board Risk Committee.
AnswerA

Embedding model performance metrics into quarterly enterprise risk assessments directly satisfies the 'Review and Revision' component, which requires organisations to identify and assess changes that substantially affect the ERM framework. Continuous monitoring of drift indicators through existing governance reviews ensures AI-specific risks are formally revisited rather than managed in isolation.

Why this answer

COSO ERM's 'Review and Revision' component requires that risk responses be reassessed and adjusted as conditions change. Integrating model performance metrics (including drift indicators) into the quarterly enterprise risk assessment and management review processes ensures drift risk is formally reviewed, escalated, and acted upon at the governance level. This embeds AI risk into existing ERM cadence rather than treating it as a purely technical pipeline concern.

Exam trap

AAIR often tests the confusion between technical monitoring controls (CI/CD triggers, automated drift detection) and governance-level review processes — candidates pick the technical control when the question asks about a COSO ERM component.

How to eliminate wrong answers

Option B is wrong because automated CI/CD drift triggers are a technical control within the 'Implement' or operational layer, not a governance review mechanism under COSO's 'Review and Revision' component. Option C is wrong because mapping models to the CMDB is an asset inventory activity supporting 'Identify' and 'Assess,' not periodic review and revision. Option D is wrong because assigning a Data Scientist to the Board Risk Committee is a resourcing/structure action, not a process for reviewing and revising AI risk responses.

Full explanation →
178
Multi-Selectmedium

Which THREE types of 'Bias' need to be managed in an AI project?

Select 3 answers
A.Measurement bias (data collection)
B.The bias towards free coffee
C.The bias of the local weather
D.Sampling bias (data selection)
E.Algorithmic bias (model logic)
AnswersA, D, E

Measurement bias arises when data collection instruments or sampling methods systematically distort values, so the training set misrepresents the population the model must serve. Recognising it as a distinct bias category satisfies the stem's requirement to identify three manageable bias types, since it can be mitigated through validated collection tools and representative sampling.

Why this answer

Measurement bias (A) must be managed because it arises during data collection when the instruments, sensors, or labeling procedures systematically distort the values recorded, so the training data does not faithfully represent the real-world phenomenon. Sampling bias (D) must be managed because it occurs during data selection when the dataset is drawn from a non-representative subset of the population, causing the model to learn patterns that do not generalize to the intended deployment context. Algorithmic bias (E) must be managed because it emerges from the model's logic, objective function, or optimization process, which can amplify or encode unfair correlations even when the input data is otherwise sound.

The remaining options do not belong: the bias towards free coffee (B) is a humorous personal preference with no bearing on AI data or model behavior, and the bias of the local weather (C) is an environmental condition rather than a systematic distortion in an AI pipeline.

Exam trap

The trap here is that candidates may confuse general 'bias' in everyday language with the specific technical categories of bias tested in AI risk management exams.

Full explanation →
179
Multi-Selecthard

A cross-functional committee is evaluating the deployment of a high-impact AI model. Which TWO factors are critical to include in the initial AI risk assessment to ensure comprehensive coverage?

Select 2 answers
A.The frequency of software patching for the underlying cloud infrastructure.
B.The socio-technical implications and potential impact on protected groups.
C.The underlying training data lineage, including provenance and potential bias sources.
D.The name of the vendor providing the cloud compute resources.
E.The specific programming language used for the model development.
AnswersB, C

High-impact AI can entrench discrimination, so the assessment must examine socio-technical implications and effects on protected groups. This satisfies the stem's demand for comprehensive coverage by addressing societal harm and fairness, not merely technical performance or narrow accuracy metrics.

Why this answer

Option B is correct because a comprehensive AI risk assessment must evaluate socio-technical implications, including how the model's deployment could disproportionately affect protected groups (e.g., by race, gender, age, or disability), which is central to fairness, equity, and regulatory compliance concerns such as those in the EU AI Act and NIST AI RMF. Option C is correct because understanding training data lineage—provenance, collection methods, and embedded bias sources—is essential to identifying data-driven risks like historical bias, sampling bias, and label bias that propagate into model outputs. Options A, D, and E are not critical initial risk-assessment factors: patching frequency (A) is a general infrastructure hygiene concern, the vendor name (D) is a procurement detail rather than a risk dimension, and the programming language (E) is an implementation choice that does not materially affect AI risk exposure.

Exam trap

AAIR often tests whether candidates confuse IT operational hygiene (patching, vendor names, languages) with genuine AI risk dimensions (data lineage, bias, socio-technical impact) — the distractor options sound plausible but are not AI-specific risks.

Full explanation →
180
MCQeasy

Which of the following is a key objective of an AI Risk Management Program?

A.To eliminate all AI risks.
B.To maximize the budget spent on AI tools.
C.To enable the ethical and safe adoption of AI technologies.
D.To replace human employees with AI.
AnswerC

The programme governs AI through risk identification, assessment and mitigation across its lifecycle, which directly enables ethical and safe adoption of AI technologies. Other objectives, such as compliance or cost control, are subordinate outcomes rather than the central purpose.

Why this answer

The key objective of an AI Risk Management Program is to enable the ethical and safe adoption of AI technologies by identifying, assessing, mitigating, and monitoring AI-specific risks such as bias, opacity, privacy violations, and safety failures. It supports responsible innovation rather than blocking AI use. This aligns with frameworks like NIST AI RMF and ISO/IEC 42001.

Exam trap

AAIR often tests whether candidates recognize that AI risk management is about enabling safe, ethical adoption rather than eliminating risk or maximizing spend, tempting them to choose absolutist or budget-focused answers.

How to eliminate wrong answers

Option A is wrong because eliminating all AI risks is neither feasible nor desirable — risk management aims to manage risk to acceptable levels, not eliminate it entirely. Option B is wrong because maximizing budget spend on AI tools is a spending goal, not a risk management objective, and could even increase risk. Option D is wrong because replacing human employees with AI is a business strategy decision, not the objective of an AI risk management program.

Full explanation →
181
MCQhard

A global bank's AI risk committee is defining key risk indicators (KRIs) for its portfolio of customer-facing AI systems. The committee wants KRIs that provide early warning of emerging risk rather than lagging confirmation of past failures. Which KRI design best meets this objective?

A.Percentage of AI models whose input data drift exceeds a defined threshold over a rolling 30-day window.
B.Aggregate spend on AI risk remediation activities compared with the annual budget.
C.Number of confirmed regulatory findings related to AI systems in the past quarter.
D.Total count of AI-related customer complaints resolved by the service desk each month.
AnswerA

Data drift beyond a threshold is a leading indicator: it signals that model inputs are diverging from training conditions before performance degradation or biased outcomes become visible to customers or regulators. Tracking the percentage of affected models over a rolling window gives the committee an early, actionable signal to trigger retraining or enhanced monitoring.

Why this answer

Leading KRIs measure conditions that precede adverse outcomes, giving the committee time to intervene. Input data drift beyond a defined threshold is detectable before model outputs degrade, before customers are harmed, and before regulators take notice. By tracking the proportion of models breaching the drift threshold over a rolling window, the committee gains a forward-looking signal that can trigger retraining, rollback, or heightened oversight.

Exam trap

The trap here is treating any frequently reported AI metric, such as complaint counts or remediation spend, as a leading indicator when those measures only confirm harm or cost that has already been incurred.

Full explanation →
182
MCQeasy

A financial institution is integrating AI risk into its ERM framework. Which action most effectively aligns AI risk appetite with enterprise risk appetite?

A.Defining quantitative risk thresholds in the AI governance policy
B.Purchasing cybersecurity insurance for AI systems
C.Assigning AI risk ownership to the IT department
D.Implementing a standalone AI audit program
AnswerA

Defining quantitative thresholds in the AI governance policy directly satisfies the stem's alignment constraint: it translates the enterprise risk appetite statement into measurable AI tolerances, so AI initiatives are approved or escalated against the same monetary and likelihood limits used across other risk domains.

Why this answer

Defining quantitative risk thresholds in the AI governance policy is correct because it translates the enterprise's stated risk appetite into measurable, enforceable limits that AI initiatives must respect. Risk appetite only becomes operational when expressed as concrete metrics (e.g., maximum acceptable bias score, error rate, or financial exposure), allowing AI risk to be aggregated and compared alongside other enterprise risks in the ERM framework. This creates direct traceability between board-level tolerance and day-to-day AI decisions.

Exam trap

AAIR often tests the confusion between risk appetite alignment (defining measurable thresholds) and risk mitigation or transfer activities (insurance, audits, ownership assignment), tempting candidates to pick controls that address symptoms rather than governance alignment.

How to eliminate wrong answers

Option B is wrong because cybersecurity insurance is a risk transfer mechanism that addresses residual financial impact after an incident, not a means of aligning AI risk appetite with enterprise risk appetite. Option C is wrong because assigning AI risk ownership solely to IT creates a siloed, technology-centric view that fails to integrate AI risk into enterprise-wide governance and can conflict with the Three Lines of Defense model. Option D is wrong because a standalone AI audit program provides independent assurance after the fact but does not itself define or align risk appetite thresholds with the ERM framework.

Full explanation →
183
MCQmedium

A company is deploying an AI system that interacts with human users. According to the AI RMF, what should be the focus of the 'Map' function in the risk governance process?

A.Choosing the most efficient algorithm.
B.Identifying stakeholders, intended use, and potential human impacts.
C.Selecting the cloud service provider.
D.Calculating the total budget for the project.
AnswerB

The Map function establishes context by documenting intended purpose, stakeholders, affected populations and potential human impacts, forming the baseline for subsequent risk measurement and treatment. This satisfies the AI RMF requirement to frame risks before assessing them.

Why this answer

In the NIST AI RMF, the Map function focuses on understanding the context of the AI system: identifying stakeholders, intended use, affected populations, and potential human impacts. It establishes the foundation for risk assessment by framing what the system does, who it affects, and what could go wrong. This contextual mapping is essential before measuring or managing risks.

Exam trap

AAIR often tests the specific purpose of each AI RMF function, so the trap is confusing Map (context and stakeholder identification) with technical, financial, or procurement decisions that belong to other phases.

How to eliminate wrong answers

Option A is wrong because choosing the most efficient algorithm is a technical design decision, not the purpose of the Map function, which is about context and risk framing. Option C is wrong because selecting a cloud service provider is an infrastructure/procurement decision, not risk mapping. Option D is wrong because calculating the project budget is a financial planning activity, not part of the AI RMF Map function.

Full explanation →
184
MCQmedium

When a 'Data Drift' alert is triggered in a production environment, what is the first step in the risk management lifecycle?

A.Immediately retrain the model.
B.Perform a root cause analysis to verify if the shift is statistical noise or a genuine change in input patterns.
C.Delete all current production logs.
D.Roll back to the previous model version.
AnswerB

Root cause analysis establishes whether the drift reflects genuine input change or random variation before any mitigation is chosen. This satisfies the lifecycle's first step: investigating and validating the alert, so subsequent risk decisions rest on evidence rather than an unverified signal.

Why this answer

The first step in the risk management lifecycle when a data drift alert triggers is to perform a root cause analysis to determine whether the drift is due to statistical noise or a genuine change in input patterns. This verification step is crucial before taking any corrective action like retraining or rolling back, as it prevents unnecessary interventions. It aligns with the risk management process of identifying and assessing the risk before mitigating it.

Exam trap

AAIR often tests the order of operations in risk management, and candidates may jump to mitigation (retraining) without first verifying the drift's validity.

How to eliminate wrong answers

Option A is wrong because immediately retraining the model without understanding the cause could waste resources and may not address the underlying issue if the drift is just noise. Option C is wrong because deleting production logs would destroy evidence needed for analysis and is not a risk management step. Option D is wrong because rolling back to a previous model version is a mitigation action that should only be taken after confirming that the drift is harmful and the current model is underperforming.

Full explanation →
185
Multi-Selectmedium

Which TWO types of documentation are critical for AI Transparency requirements?

Select 2 answers
A.Data Sheets for Datasets explaining the provenance and makeup of training data.
B.Model Cards detailing intended use, limitations, and performance.
C.A weekly calendar of team meetings.
D.The employee handbook.
E.A list of all employees who have access to the office.
AnswersA, B

Data Sheets for Datasets document a training dataset's provenance, composition, collection methods, and known limitations. This satisfies AI transparency by disclosing what data shaped the model, enabling stakeholders to assess suitability, bias exposure, and fitness for purpose rather than treating the model as an opaque artefact.

Why this answer

Data Sheets for Datasets (A) are critical because AI transparency requires documenting the provenance, composition, collection methods, and preprocessing of training data so stakeholders can assess bias, consent, and suitability. Model Cards (B) are equally essential since they disclose a model's intended use, limitations, performance metrics across relevant subgroups, and ethical considerations, directly supporting transparency obligations. Together these artifacts give the technical and contextual information regulators and users need to understand how an AI system was built and how it should be used.

The remaining options do not belong: a weekly meeting calendar (C) and an employee handbook (D) are general administrative documents unrelated to AI system transparency, and an office access list (E) concerns physical security, not AI documentation.

Exam trap

AAIR often tests whether candidates confuse general corporate documentation (handbooks, calendars, access lists) with AI-specific transparency artifacts, so the trap is selecting any document that sounds 'official' rather than one that specifically describes datasets or models.

Full explanation →
186
Multi-Selectmedium

Which TWO metrics are most appropriate for measuring the effectiveness of an AI Risk Program?

Select 2 answers
A.The average salary of data scientists.
B.Number of coffee machines in the building.
C.Number of social media followers of the company.
D.Total number of AI-related incidents reported and remediated.
E.Percentage of AI models that have undergone a formal risk assessment.
AnswersD, E

Counting reported and remediated AI incidents gives a direct, quantifiable measure of the risk programme's operational effectiveness, satisfying the stem's demand for metrics rather than qualitative assessments. Remediation volume evidences that identified risks are actually being treated, not merely logged.

Why this answer

Option D is correct because tracking the total number of AI-related incidents reported and remediated directly measures how well the AI Risk Program detects, responds to, and resolves real risk events, which is a core effectiveness indicator. Option E is correct because the percentage of AI models that have undergone a formal risk assessment measures the program's coverage and consistency in applying risk controls across the AI model inventory, reflecting proactive risk management maturity. Options A, B, and C are not appropriate because average data scientist salary, coffee machine count, and social media followers are unrelated to AI risk identification, assessment, mitigation, or governance outcomes and do not indicate program effectiveness.

Exam trap

AAIR often tests the distinction between vanity/business metrics (salaries, followers, facilities) and genuine risk program effectiveness metrics, so candidates must recognize that only metrics tied to risk identification, assessment, or remediation are valid.

Full explanation →
187
MCQmedium

A financial institution uses an AI model to approve loans. After deployment, they notice the model is rejecting loan applications from a specific region at a rate 30% higher than historical human benchmarks. What is the most appropriate next step in the risk lifecycle?

A.Ignore the shift as it may be a temporary market trend.
B.Conduct a bias investigation by comparing input distribution shifts between training data and current inference requests.
C.Immediately shut down the system.
D.Train a secondary model to override the primary one.
AnswerB

Comparing training versus inference input distributions isolates whether the elevated regional rejection rate stems from data drift rather than genuine risk factors, directly addressing the 30% disparity against human benchmarks before any retraining or deployment change.

Why this answer

A 30% higher rejection rate for a specific region versus historical human benchmarks is a classic bias/drift signal, so the correct risk-lifecycle step is to investigate by comparing the input distribution at inference time against the training distribution. This determines whether the disparity stems from data drift, a biased feature, or a genuine change in the population, and it produces evidence before any remediation decision. Ignoring, shutting down, or building an override model are premature or inappropriate responses.

Exam trap

The trap is choosing an extreme action (shut down the system) or a dismissive one (ignore the trend) — the risk lifecycle consistently rewards investigation and evidence-gathering as the first step before remediation.

How to eliminate wrong answers

Option A is wrong because dismissing the shift as a 'temporary market trend' ignores a statistically significant disparity that could constitute illegal discrimination and violates the monitoring obligations of an AI risk framework. Option C is wrong because immediately shutting down the system is a disproportionate response that halts business operations before root cause is known — risk frameworks call for investigation and proportionate mitigation, not reflexive shutdown. Option D is wrong because training a secondary model to override the primary one adds complexity and a second unvalidated model without diagnosing the original bias, which compounds rather than resolves the risk.

Full explanation →
188
MCQmedium

A retail bank's credit-scoring AI was validated on 2023 applicant data and deployed in early 2024. By Q4 2024, portfolio default rates among approved loans have crept up while the model's input feature distributions remain statistically unchanged. The risk team wants to isolate the most probable lifecycle cause before retraining. Which risk management activity best addresses this situation?

A.Recalibrate the model's probability outputs using isotonic regression on the original 2023 validation set.
B.Increase the decision threshold for approval to reduce the volume of higher-risk applicants entering the portfolio.
C.Run a Kolmogorov-Smirnov test on each input feature to detect data drift in the production population.
D.Recompute the model's discrimination metrics against a fresh outcome-labeled holdout and compare them to the validation baseline.
AnswerD

Here the inputs are stable but realized defaults are worsening, which points to concept drift: the relationship between features and the default outcome has changed. Recomputing performance metrics against newly labeled outcomes directly tests that hypothesis and produces evidence for the model risk committee. This is the correct diagnostic step before any retraining, recalibration, or threshold change is authorized.

Why this answer

When input distributions are stable but realized outcomes worsen, the lifecycle risk is concept drift rather than data drift. The only control listed that measures the changed feature-to-outcome relationship is re-evaluating discrimination and calibration against newly labeled outcomes and comparing to the validation baseline. That evidence determines whether retraining, recalibration, or redesign is warranted.

Exam trap

The trap here is assuming that rising defaults must mean the input data drifted, when stable feature distributions with worsening outcomes indicate the feature-to-label relationship itself has shifted.

Full explanation →
189
MCQeasy

Which of the following is an example of 'Data Quality Risk' in the AI lifecycle?

A.Setting a high learning rate.
B.Using Python instead of R.
C.Using a GPU that is two generations old.
D.High rates of mislabeled training data.
AnswerD

Mislabeled training data corrupts the ground truth the model learns from, degrading accuracy and fairness. This is a data quality risk because the defect originates in the dataset itself, not in model architecture, deployment, or inference infrastructure.

Why this answer

High rates of mislabeled training data are a classic Data Quality Risk because labels directly determine what the model learns; systematic mislabeling introduces bias and degrades accuracy. Data quality risks encompass issues with accuracy, completeness, consistency, and provenance of training data. Mislabeling is a data integrity problem that propagates into model behavior and is hard to detect post-training.

Exam trap

AAIR often tests whether candidates can separate data quality risks (labels, bias, completeness) from model configuration risks (hyperparameters) and infrastructure risks (hardware, tooling).

How to eliminate wrong answers

Option A is wrong because setting a high learning rate is a hyperparameter tuning issue, which is a model development/configuration risk, not a data quality risk. Option B is wrong because choosing Python over R is a tooling preference with no inherent data quality implication. Option C is wrong because using an older GPU is an infrastructure/cost concern, not a data quality issue.

Full explanation →
190
Multi-Selecthard

Which THREE strategies are effective for mitigating 'Adversarial Risk' in an AI lifecycle?

Select 3 answers
A.Hiring more documentation staff.
B.Rate limiting on API calls to prevent brute-force querying.
C.Using a more expensive cloud provider.
D.Adversarial training to harden the model against perturbed inputs.
E.Input sanitization and validation on all model endpoints.
AnswersB, D, E

Rate limiting constrains the volume and frequency of queries an attacker can submit, directly blunting brute-force and model-extraction attempts that rely on mass probing. It satisfies the adversarial-risk constraint by capping automated abuse without blocking legitimate traffic.

Why this answer

Option B is correct because rate limiting on API calls throttles the query volume an attacker can send, which directly counters brute-force and model-extraction attacks that rely on high-volume probing of the inference endpoint. Option D is correct because adversarial training explicitly augments the training set with perturbed examples, hardening the model's decision boundaries against evasion and adversarial-input attacks. Option E is correct because input sanitization and validation on all model endpoints rejects malformed, out-of-distribution, or injection-style payloads before they reach the model, reducing the attack surface for adversarial manipulation.

Option A does not belong because adding documentation staff addresses process and knowledge management, not adversarial manipulation of the model. Option C does not belong because a more expensive cloud provider changes cost, not the security posture against adversarial inputs.

Exam trap

AAIR often tests the misconception that generic IT measures (like more staff or expensive providers) mitigate AI-specific adversarial risks, when only targeted technical controls like rate limiting, adversarial training, and input validation are effective.

Full explanation →
191
MCQmedium

You are mapping existing ISO 31000 risk management processes to an AI-specific application. Which step is most crucial to ensure 'Risk Treatment' is appropriately scaled for AI?

A.Automating all risk logging in the SIEM.
B.Discontinuing the use of third-party AI APIs to avoid vendor risk.
C.Standardizing all AI models under one security policy regardless of function.
D.Implementing a tiered control framework where mitigations scale based on the system's impact assessment.
AnswerD

A tiered control framework directly satisfies the scaling constraint by matching mitigation strength to each system's assessed impact level, rather than applying uniform controls across all AI systems. This ensures high-impact deployments receive proportionately stronger treatment, aligning ISO 31000 risk treatment with AI-specific risk magnitudes.

Why this answer

A tiered control framework is the correct answer because ISO 31000's Risk Treatment step requires selecting controls proportionate to the assessed risk level. For AI, this means mitigations (e.g., human-in-the-loop review, explainability requirements, monitoring intensity) must scale with the system's impact classification — a low-risk internal chatbot needs lighter controls than a high-risk hiring or medical diagnostic model. This impact-based tiering directly operationalizes ISO 31000's principle that treatment options be evaluated against risk criteria and organizational context.

Exam trap

AAIR often tests the misconception that any single control (logging, policy standardization, or vendor elimination) satisfies Risk Treatment, when ISO 31000 actually demands proportionate, impact-scaled mitigation selection.

How to eliminate wrong answers

Option A is wrong because automating risk logging in a SIEM addresses detection and monitoring, not Risk Treatment design — SIEM logging is a detective control, not a scaled mitigation framework. Option B is wrong because blanket discontinuation of third-party AI APIs is a risk avoidance strategy applied indiscriminately, which violates ISO 31000's requirement that treatment be proportionate and context-driven; it also ignores that vendor risk can be managed through contractual and technical controls. Option C is wrong because standardizing all AI models under one policy regardless of function ignores the impact assessment entirely, defeating the purpose of scaling treatment to risk level.

Full explanation →
192
MCQeasy

Which of the following is a classic example of 'AI Hallucination' that needs monitoring?

A.The system is running slowly
B.The system requires too much electricity
C.The system password has expired
D.The system generates confident but factually incorrect information
AnswerD

Generating fluent, confident output that is factually wrong is the defining hallucination failure mode, so monitoring must detect unsupported claims rather than merely flagging uncertain or hedged responses. This satisfies the stem's requirement for a classic example needing oversight.

Why this answer

AI hallucination refers to a model generating confident but factually incorrect or nonsensical information, which is a key risk that requires monitoring. This can lead to misinformation, user harm, and reputational damage. The other options describe operational or infrastructure issues, not hallucination.

Exam trap

AAIR often tests the confusion between operational issues (performance, cost) and AI-specific risks like hallucination, leading candidates to pick irrelevant technical problems.

How to eliminate wrong answers

Option A is wrong because slow system performance is a latency or resource issue, not a hallucination. Option B is wrong because high electricity consumption is an efficiency concern, not a factual accuracy problem. Option C is wrong because an expired password is a security/authentication issue, unrelated to model output correctness.

Full explanation →
193
MCQmedium

Why is it critical to have 'Human-in-the-Loop' (HITL) for high-risk AI decisions?

A.To make the computer faster
B.To replace the need for model testing
C.To provide accountability and a safeguard against errors
D.To save money on hardware
AnswerC

Human-in-the-Loop inserts a person who reviews and can override high-risk AI outputs, creating a traceable decision owner. This delivers accountability and acts as a safeguard against model errors, satisfying the stem's requirement for high-risk decisions.

Why this answer

Human-in-the-Loop (HITL) ensures that a qualified human reviews and can override high-risk AI decisions, providing both accountability (a responsible person is answerable for the outcome) and a safeguard against model errors, bias, or edge cases the model was not trained for. This is essential for high-stakes domains like healthcare, finance, and criminal justice.

Exam trap

The trap is thinking HITL is about performance or cost optimization — it is fundamentally about accountability and error mitigation in high-risk decisions.

How to eliminate wrong answers

Option A is wrong because HITL adds human review time and does not make the computer faster — it may slow the process. Option B is wrong because HITL complements, not replaces, model testing; testing validates the model before deployment, while HITL catches issues in production. Option D is wrong because HITL typically increases operational cost (human reviewers), not reduce hardware costs.

Full explanation →
194
Multi-Selecteasy

Which TWO teams should be involved in the initial risk review for an AI chatbot?

Select 2 answers
A.Data Science/ML Engineering
B.Legal Department
C.The company mailroom staff
D.The local fitness coach
E.The custodial cleaning staff
AnswersA, B

Data Science/ML Engineering must join the initial risk review because they understand the chatbot's model, training data and known limitations. This satisfies the stem's constraint by supplying the technical evidence needed to assess model-specific risks before deployment.

Why this answer

Option A (Data Science/ML Engineering) is correct because this team owns the chatbot's model design, training data, and evaluation metrics, so they are best positioned to identify technical risks such as bias, hallucination, prompt injection, and data leakage during an initial risk review. Option B (Legal Department) is correct because legal must assess regulatory compliance (e.g., GDPR, CCPA), intellectual property, liability, and contractual exposure arising from the chatbot's outputs and data handling. The remaining options do not belong: the mailroom staff (C), local fitness coach (D), and custodial cleaning staff (E) have no role in AI system risk assessment, data governance, or regulatory compliance for a chatbot.

Exam trap

The trap is selecting irrelevant roles (mailroom, fitness coach, custodial staff) as distractors — exam questions about AI risk review expect you to identify the technical and legal functions that actually own AI risk.

Full explanation →
195
MCQhard

In the context of the 'AI Lifecycle', what is the risk of 'Model Decay'?

A.The model physically degrades over time.
B.The model's code becomes deprecated due to language updates.
C.The model's predictions become less accurate because the underlying environment or user behavior changes.
D.The training hardware becomes too slow to run.
AnswerC

Model decay occurs when the statistical relationship learned during training no longer matches live data, so accuracy degrades as user behaviour or the operating environment shifts. Retraining or monitoring is needed to restore predictive reliability.

Why this answer

Model decay (also called model drift or concept drift) is the degradation of a model's predictive accuracy over time because the real-world environment, data distributions, or user behavior change after training. The model itself is unchanged, but its assumptions no longer match reality. This is why ongoing monitoring and retraining are essential in the AI lifecycle.

Exam trap

AAIR often tests whether candidates confuse model decay (statistical/environmental drift) with software deprecation or hardware obsolescence, luring them toward literal interpretations of 'decay.'

How to eliminate wrong answers

Option A is wrong because models are software artifacts and do not physically degrade; decay is a statistical/behavioral phenomenon, not hardware wear. Option B is wrong because code deprecation due to language updates is a software maintenance issue, not model decay. Option D is wrong because slow training hardware is an infrastructure performance concern, not a degradation of model accuracy.

Full explanation →
196
MCQmedium

A global bank has adopted the NIST AI Risk Management Framework (AI RMF) and wants to ensure that its AI risk governance is integrated with its existing enterprise risk management (ERM) processes. The Chief Risk Officer asks the AI governance lead to identify the primary mechanism for achieving this integration. Which action best accomplishes this?

A.Mapping the AI RMF core functions to the bank's existing ERM risk categories and reporting lines.
B.Delegating all AI risk decisions to the data science team to ensure technical accuracy.
C.Adopting a separate AI risk framework that replaces the bank's existing ERM framework.
D.Creating a standalone AI risk register that is maintained separately from the enterprise risk register.
AnswerA

Mapping AI RMF functions to existing ERM categories and reporting lines directly integrates AI risk into established governance structures, ensuring consistent risk assessment, reporting, and oversight. This avoids creating parallel, disconnected processes and leverages existing ERM frameworks for AI-specific risks, which is a key goal of the AI RMF Govern function. It enables the bank to treat AI risk as a subset of enterprise risk, facilitating holistic management.

Why this answer

The correct action is to map AI RMF core functions to the bank's existing ERM risk categories and reporting lines. This integrates AI risk into established governance structures, ensuring consistent oversight and reporting. It leverages existing ERM processes rather than creating parallel systems, which aligns with the AI RMF's emphasis on integration.

This approach facilitates holistic risk management and avoids silos.

Exam trap

The trap here is assuming that AI risk requires a completely separate framework, rather than integrating it into existing ERM processes.

Full explanation →
197
MCQmedium

When integrating AI risk into the corporate GRC (Governance, Risk, and Compliance) platform, what is the best practice for handling 'AI Model Drift' alerts?

A.Send all alerts to the help desk.
B.Ignore drift alerts until a manual audit happens annually.
C.Integrate drift alerts as risk indicators in the GRC platform for automated tracking.
D.Disable the alerts to prevent 'alert fatigue'.
AnswerC

Feeding drift alerts into the GRC platform as risk indicators lets thresholds, ownership and remediation workflows trigger automatically, satisfying the requirement to integrate AI risk into existing governance processes rather than handling drift in isolated ML tooling.

Why this answer

Best practice is to integrate model drift alerts directly into the GRC platform as risk indicators, so they are tracked, scored, and routed through the same governance workflows as other enterprise risks. This enables automated tracking, trend analysis, and timely escalation. The other options either misroute alerts, ignore them, or suppress them, all of which defeat the purpose of drift monitoring.

Exam trap

AAIR often tests whether candidates treat drift alerts as noise to be suppressed or as governance signals to be integrated — the trap is choosing 'disable alerts' or 'ignore until audit' as a shortcut for alert fatigue.

How to eliminate wrong answers

Option A is wrong because sending all drift alerts to the help desk misroutes a model-risk signal to a team without the authority or context to assess AI risk. Option B is wrong because ignoring drift alerts until an annual audit leaves the model unmonitored for up to a year, allowing silent degradation. Option D is wrong because disabling alerts to avoid fatigue removes the very signal needed to detect model degradation — the correct response is tuning and prioritization, not suppression.

Full explanation →
198
Multi-Selectmedium

When reporting AI risks to the Board of Directors, which THREE elements should be included to ensure informed decision-making?

Select 3 answers
A.List of all developers who worked on the model.
B.Granular technical logs of model inference errors.
C.Qualitative overview of ethical AI alignment with company values.
D.Summary of current AI risk appetite and threshold breaches.
E.Strategic assessment of regulatory compliance and potential legal exposure.
AnswersC, D, E

A qualitative overview of ethical AI alignment maps model behaviour and governance choices to stated company values, giving the Board the value-based context needed to judge reputational and ethical exposure. This satisfies the requirement for informed decision-making beyond purely quantitative risk metrics.

Why this answer

Option C is correct because a qualitative overview of ethical AI alignment with company values gives the Board the governance-level context needed to judge whether AI use is consistent with the organization's stated principles and risk culture. Option D is correct because reporting the current AI risk appetite together with any threshold breaches tells the Board whether actual risk exposure remains within the limits they approved, which is essential for informed oversight and escalation. Option E is correct because a strategic assessment of regulatory compliance and potential legal exposure highlights material obligations, enforcement trends, and liabilities that could affect strategy, reputation, and financial standing.

Options A and B are not appropriate for Board-level reporting: a list of developers is an operational staffing detail, and granular technical logs of model inference errors are too low-level and tactical, belonging instead to engineering or model-risk teams rather than the Board.

Full explanation →
199
Multi-Selectmedium

Which THREE factors should be evaluated when reviewing an AI model for 'Fairness'?

Select 3 answers
A.Performance disparity across different demographic cohorts.
B.The number of hours the server was powered on.
C.Defined metrics for success that explicitly account for equity.
D.Training data representation across protected groups.
E.The model's file size in kilobytes.
AnswersA, C, D

Fairness requires measuring whether model outcomes differ across protected groups, so comparing error rates, selection rates and accuracy between demographic cohorts exposes disparate impact. This directly satisfies the stem's demand for a fairness factor, since parity across cohorts is the defining test of equitable model behaviour.

Why this answer

Option A is correct because fairness review requires measuring performance disparity across different demographic cohorts — comparing metrics such as accuracy, precision, recall, or false positive/negative rates between protected groups reveals disparate impact. Option C is correct because fairness cannot be assessed without defined metrics for success that explicitly account for equity, such as equalized odds, demographic parity, or equal opportunity, which operationalize what 'fair' means for the use case. Option D is correct because training data representation across protected groups directly affects model behavior; underrepresentation or historical bias in the data propagates into biased predictions, so representation must be audited.

Option B is not relevant because server uptime is an infrastructure/availability metric unrelated to model fairness. Option E is not relevant because model file size in kilobytes is a storage or deployment concern, not a fairness indicator.

Exam trap

AAIR often tests the distinction between fairness factors and irrelevant operational metrics, and candidates may mistakenly include infrastructure or storage metrics as fairness considerations.

Full explanation →
200
MCQmedium

A financial firm is integrating AI risk into its ERM. The Chief Risk Officer asks for a mapping between the NIST AI Risk Management Framework (RMF) and the existing COSO ERM framework. Which action most effectively facilitates this alignment?

A.Replace the COSO risk taxonomy with the NIST AI RMF taxonomy entirely.
B.Automate all AI model monitoring using only the NIST RMF 'Measure' function.
C.Assign AI model owners to report directly to the NIST oversight board.
D.Map NIST AI RMF 'Govern' and 'Map' functions to COSO 'Governance and Culture' components.
AnswerD

The Govern and Map functions address governance structures, roles and context setting, mirroring COSO's Governance and Culture components. This mapping lets the firm slot AI-specific controls into existing ERM governance, satisfying the CRO's requirement for a coherent cross-framework alignment rather than parallel, disconnected risk registers.

Why this answer

Mapping the NIST AI RMF 'Govern' and 'Map' functions to COSO's 'Governance and Culture' component is the most effective alignment action because both address oversight, accountability, and risk context. COSO ERM's Governance and Culture component parallels the AI RMF's emphasis on governance structures and risk mapping, enabling the firm to integrate AI risk into existing ERM processes without replacing them. This mapping preserves the firm's investment in COSO while satisfying NIST AI RMF expectations.

Exam trap

The trap is choosing an action that replaces or bypasses the existing framework rather than mapping to it; candidates must recognize that alignment means integration, not substitution, and that NIST is not an oversight authority.

How to eliminate wrong answers

Option A is wrong because replacing the COSO taxonomy entirely would discard the firm's established ERM framework, create disruption, and is not a mapping action; alignment means integrating, not replacing. Option B is wrong because automating monitoring using only the 'Measure' function ignores the other AI RMF functions and does not address the CRO's request for a mapping between frameworks. Option C is wrong because assigning AI model owners to report directly to a NIST oversight board is an organizational change that does not facilitate framework alignment and misrepresents NIST as an oversight body rather than a standards publisher.

Full explanation →
201
Multi-Selecthard

A financial services firm is integrating the NIST AI Risk Management Framework (AI RMF) into its existing governance structure. The AI governance lead is tasked with ensuring that the 'Govern' function is properly implemented. Which two actions are essential for establishing effective AI risk governance? (Choose two.)

Select 2 answers
A.Ensuring that all AI models are developed using a single, standardized programming language.
B.Requiring that all AI risk decisions be made by the Chief Technology Officer.
C.Establishing a process for regular review and update of AI risk policies and procedures.
D.Defining and documenting roles and responsibilities for AI risk management across the organization.
E.Implementing a policy that prohibits the use of AI in any customer-facing applications.
AnswersC, D

Establishing a process for regular review and update of AI risk policies ensures that governance remains relevant as AI technologies and risks evolve. The Govern function emphasizes the need for policies and procedures to be maintained and updated. This process helps the organization adapt to new regulatory requirements, changes in risk appetite, and lessons learned. It is a key component of a living governance framework, preventing stagnation and ensuring ongoing alignment with organizational objectives.

Why this answer

The essential actions for effective AI risk governance under the Govern function are defining and documenting roles and responsibilities, and establishing a process for regular review and update of AI risk policies. These ensure accountability, clarity, and adaptability. The other options are either technical details, centralize authority inappropriately, or represent excessive risk avoidance, none of which are required by the AI RMF Govern function.

Exam trap

The trap here is confusing technical standardization or risk avoidance with governance requirements, which focus on organizational structures and processes.

Full explanation →
202
MCQmedium

A company is integrating an AI model that uses 'Online Inference'. What is the most significant risk regarding model security?

A.The risk of the training data being corrupted by the user.
B.The risk of high database storage costs.
C.The risk of the Python interpreter crashing.
D.The risk of model inversion attacks on the inference endpoint.
AnswerD

Online inference exposes a live queryable endpoint, letting attackers probe outputs to reconstruct training data or extract sensitive attributes through model inversion. Batch or offline inference lacks this continuously exposed attack surface, making endpoint hardening essential.

Why this answer

Online inference endpoints are exposed to the public/network, creating surface area for adversarial attacks.

Full explanation →
203
MCQeasy

Which of the following is a primary risk during the 'Deployment' stage of the AI lifecycle?

A.Unexpected performance degradation due to infrastructure differences between development and production.
B.Cleaning the initial dataset.
C.Writing the initial research paper.
D.Choosing the wrong programming language.
AnswerA

Infrastructure differences between development and production cause distribution shift, so model accuracy degrades once live traffic arrives. This satisfies the Deployment-stage constraint: risks emerge when the model meets real-world conditions, not during training or design. Performance monitoring and rollback plans therefore become essential at this stage.

Why this answer

During deployment, the model moves from a controlled development environment into production, where infrastructure, data distributions, and traffic patterns differ. This shift commonly causes unexpected performance degradation — for example, different hardware, library versions, or input distributions than those used during training and validation. The other options describe activities from earlier lifecycle stages (data preparation, research) or are not AI lifecycle risks at all.

Exam trap

AAIR often tests lifecycle-stage mapping — the trap is choosing an answer that is a real AI concern but belongs to a different stage (e.g., data cleaning) rather than the Deployment stage named in the question.

How to eliminate wrong answers

Option B is wrong because cleaning the initial dataset is a Data Preparation / data engineering activity that occurs before training, not a Deployment-stage risk. Option C is wrong because writing the initial research paper belongs to the Research or Ideation phase, well before deployment. Option D is wrong because choosing the wrong programming language is a design/development decision, not a risk specific to the Deployment stage.

Full explanation →
204
MCQhard

When an AI model fails in production, what is the 'Post-Mortem' goal?

A.To identify root causes and update risk management practices
B.To hide the failure from management
C.To fire the developer responsible
D.To delete all logs associated with the model
AnswerA

A post-mortem examines the production failure to establish why the model failed, then feeds those findings back into risk controls, monitoring and governance. This closes the loop, reducing recurrence and improving the organisation's AI risk management practices.

Why this answer

A post-mortem's goal is to identify root causes of the failure and feed those lessons back into risk management practices, controls, and monitoring. It is a blameless, learning-oriented exercise that improves future resilience rather than assigning fault. This aligns with AAIR's emphasis on continuous improvement of the AI risk management lifecycle.

Exam trap

AAIR often tests the misconception that post-mortems are about assigning blame or hiding incidents, when the correct answer is blameless root-cause learning that improves risk practices.

How to eliminate wrong answers

Option B is wrong because hiding failures from management violates governance, transparency, and incident-reporting obligations and prevents corrective action. Option C is wrong because firing the developer is a punitive, blame-oriented response that suppresses reporting and does not address systemic root causes. Option D is wrong because deleting logs destroys the forensic evidence needed for root-cause analysis and may violate regulatory retention requirements.

Full explanation →
205
MCQeasy

Which team is usually responsible for the 'Technical Implementation' of risk controls in an AI model?

A.Internal Audit
B.Human Resources
C.Legal Department
D.Data Science/Machine Learning Team
AnswerD

The Data Science/Machine Learning team builds and operates the model, so it implements the technical controls — such as input validation, bias testing and monitoring hooks — within the pipeline. Business, legal and audit functions define requirements and oversee assurance rather than performing the engineering work.

Why this answer

The Data Science/Machine Learning Team is responsible for the technical implementation of risk controls in an AI model because they build, train, and deploy the model. They implement controls such as bias mitigation, model monitoring, access controls, and performance safeguards directly in the code and pipeline. Other teams provide oversight, policy, or compliance guidance but do not typically implement technical controls.

Exam trap

The trap here is that candidates may associate risk controls with audit or legal teams, but the question specifically asks about technical implementation, which falls to the ML team.

How to eliminate wrong answers

Option A is wrong because Internal Audit assesses and verifies controls but does not implement them. Option B is wrong because Human Resources manages personnel and policy, not technical AI controls. Option C is wrong because the Legal Department provides legal guidance and compliance oversight, not technical implementation.

Full explanation →
206
MCQhard

In an AI risk program, how should 'Residual Risk' be treated?

A.It must be formally documented and periodically reviewed against risk appetite
B.It should be ignored if below a specific performance threshold
C.It should always be zero for production systems
D.It should be formally accepted by the risk owner if it exceeds appetite
AnswerA

Residual risk is the risk remaining after controls are applied, so it must be recorded in the risk register and re-evaluated at intervals to confirm it stays within the organisation's stated risk appetite and tolerance thresholds.

Why this answer

Residual risk is the risk remaining after controls are applied, and it must be formally documented and periodically reviewed against the organization's risk appetite to confirm it stays within tolerance. This creates an auditable trail and ensures changes in the threat landscape or model behavior trigger reassessment. Treating residual risk as a one-time calculation rather than a living value is a common program failure.

Exam trap

The trap is the reversed condition in option D — candidates who know risk owners accept risk may pick it without noticing it applies only when residual risk EXCEEDS appetite, not within it.

How to eliminate wrong answers

Option B is wrong because ignoring residual risk below a performance threshold conflates model performance with risk — a model can perform well yet still create compliance, bias, or privacy exposure that must be tracked. Option C is wrong because zero residual risk is unattainable in practice; claiming it indicates either a misunderstanding of risk or a failure to identify threats. Option D is wrong because formal acceptance by the risk owner is required when residual risk EXCEEDS appetite, not when it is within appetite — the option reverses the condition and misstates the governance trigger.

Full explanation →
Page 2

Page 3 of 3

All pages

123

Practice AAIR by domain

Target a specific domain to shore up weak areas.

AI Risk Program ManagementAI Risk Governance And Framework IntegrationAI Lifecycle Risk Management
See all domains with question counts →