AAIR Practice Exam 4 — 20 Questions

Question 1

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

Accepted Answer

The pricing model provides incorrect, discounted rates. 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.

Answer

The model documentation is outdated

Answer

The font size in the UI is too small

Answer

The model's name is not catchy

Question 2

You are managing AI supply chain risk. Which control is most critical for third-party AI models?

Accepted Answer

Requiring a documented 'AI Bill of Materials' to understand data lineage and model provenance.. The most critical control for third-party AI models is requiring a documented 'AI Bill of Materials' (AI BOM) to understand data lineage and model provenance. This provides transparency into the components, data sources, and training methods used in the model, enabling effective risk assessment and mitigation. Without this visibility, organizations cannot verify compliance, detect biases, or manage supply chain risks.

Answer

Conducting a manual audit of the vendor's source code.

Answer

Requiring the vendor to submit a monthly performance report.

Answer

Setting a limit on the number of third-party vendors allowed.

Question 3

Which AI lifecycle stage is most susceptible to the risk of 'data leakage' where training data inadvertently contains information from the future/target?

Accepted Answer

Data preparation and training phase.. Data leakage occurs when information that should not be available at inference time (e.g., the target label, future data, or holdout data) is inadvertently included in the training set. This happens during data preparation and training, where feature engineering, joins, and train/test splits are performed. If a feature is computed using the target or future timestamps, the model learns a shortcut that inflates training metrics but fails in production.

Answer

Model monitoring phase.

Answer

Model decommission phase.

Answer

Deployment phase.

Question 4

An organization is integrating AI risk into its existing ISO 31000 framework. How should the 'Risk Assessment' process be modified to account for AI-specific 'black box' issues?

Accepted Answer

Add a model explainability and interpretability assessment step to the process.. Inclusion of model explainability assessments is necessary to address the opacity inherent in deep learning models.

Answer

Remove the risk identification step, as AI risks are too unpredictable.

Answer

Replace the qualitative assessment with a purely quantitative financial impact model.

Answer

Delegate all risk assessments to the external software vendor.

Question 5

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

Accepted Answer

It ensures human oversight for critical decisions, reducing the risk of 'black box' failures.. 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.

Answer

It removes the requirement for model bias testing.

Answer

It is required for every AI model regardless of impact.

Answer

It eliminates the need for any other automated risk controls.

Question 6

When coordinating AI risk across a global organization, why is it critical to include local legal counsel in the AI steering committee?

Accepted Answer

To interpret the varied regulatory landscapes in different operational regions.. Local legal counsel is essential on a global AI steering committee because AI regulations vary significantly by jurisdiction — the EU AI Act, US state laws, China's algorithm rules, Brazil's LGPD, etc. — and local counsel interprets how these apply to the organization's operations in each region. This ensures the committee's risk decisions are legally grounded and regionally compliant rather than assuming a single global standard.

Answer

To audit the training datasets for copyright infringement.

Answer

To handle administrative tasks for the committee.

Answer

To approve the technical architecture of the AI models.

Question 7

Which TWO strategies are recommended for 'Mitigating' AI model bias?

Accepted Answer

Applying fairness constraints during model training. Option C is correct because applying fairness constraints during model training directly mitigates bias by incorporating fairness metrics (such as demographic parity or equalized odds) into the objective function or as regularization terms, forcing the model to reduce disparate impact across protected groups. Option E is correct because training on more diverse and representative datasets addresses bias at its source—the data—by ensuring the model learns patterns from a broader population, reducing sampling bias and underrepresentation that lead to skewed predictions. Options A, B, and D are not recommended mitigation strategies: deleting all data destroys the training set and makes learning impossible, asking users to manually fix bias shifts responsibility away from systematic model design and is not scalable or reliable, and ignoring bias allows it to persist and potentially worsen in production.

Answer

Removing all data from the database

Answer

Asking the user to manually fix the bias

Answer

Ignoring the bias and hoping it goes away

Question 8

Which THREE factors are essential to consider when determining the 'AI Risk Appetite' for an enterprise?

Accepted Answer

The organization's tolerance for legal and reputational damage.. Option B is correct because AI risk appetite must be calibrated to how much legal exposure and reputational harm the organization is willing to absorb, since these are the primary consequence categories of AI failures and directly bound the acceptable risk threshold. Option C is correct because the complexity and degree of autonomy of the deployed AI models determine the severity and likelihood of adverse outcomes, so higher autonomy or opacity requires a more conservative appetite. Option D is correct because the applicable regulatory landscape and sector-specific compliance obligations (for example, EU AI Act risk tiers, HIPAA, or financial-services rules) set non-negotiable limits that the risk appetite cannot exceed. Option A is not essential because market share is a business-performance metric that does not define risk tolerance boundaries. Option E is not essential because historical data volume is an input to model development and data-governance decisions, not a determinant of how much AI risk the enterprise is willing to accept.

Answer

The current market share of the company.

Answer

Historical data volume.

Question 9

Which TWO of the following techniques help mitigate bias in the AI lifecycle?

Accepted Answer

Conducting fairness audits using specialized metrics (e.g., Disparate Impact).. Option C is correct because conducting fairness audits with specialized metrics such as Disparate Impact (the ratio of favorable outcomes for a protected group to those of a reference group, ideally near 1.0 under the 80% rule) quantitatively detects and helps correct disparate treatment or outcomes across groups in the AI lifecycle. Option D is correct because sourcing training data from diverse, representative populations reduces sampling and representation bias, ensuring the model learns patterns that generalize fairly rather than overfitting to a dominant or historically skewed subgroup. Option A is incorrect because the programming language (e.g., Java) is irrelevant to bias; bias arises from data, objectives, and evaluation, not from the implementation language. Option B is incorrect because reducing model monitoring would hide drift and emerging bias, worsening rather than mitigating it. Option E is incorrect because simply adding more features can amplify bias through proxy variables and increase overfitting without addressing underlying data or fairness issues.

Answer

Switching the programming language to Java.

Answer

Reducing the frequency of model monitoring.

Answer

Increasing the number of features in the model.

Question 10

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?

Accepted Answer

Identifying stakeholders, intended use, and potential human impacts.. 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.

Answer

Choosing the most efficient algorithm.

Answer

Selecting the cloud service provider.

Answer

Calculating the total budget for the project.

Question 11

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

Accepted Answer

Link AI risks to overall business impact, strategic goals, and current risk appetite.. 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.

Answer

State that AI is completely safe and requires no oversight.

Answer

Provide a list of all model parameters.

Answer

Give a 4-hour technical lecture on deep learning.

Question 12

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

Accepted Answer

It must be formally documented and periodically reviewed against risk appetite. 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.

Answer

It should be ignored if below a specific performance threshold

Answer

It should always be zero for production systems

Answer

It should be formally accepted by the risk owner if it exceeds appetite

Question 13

When assessing the risk of AI-generated content, which factor is the most important for calculating 'Impact'?

Accepted Answer

The potential for reputational and legal consequences. Impact in risk assessment is defined as the magnitude of consequences if a risk materializes; for AI-generated content, the most significant consequences are reputational damage and legal liability (defamation, copyright infringement, regulatory penalties, discrimination claims). These are the factors that determine severity in a risk matrix, whereas generation speed, font color, and GPU count are technical or cosmetic attributes with no bearing on harm magnitude.

Answer

The speed at which the content is generated

Answer

The color of the font used for the output

Answer

The number of GPUs used to create the content

Question 14

A team is designing an AI lifecycle monitoring system. Which metric is most effective for detecting 'Out-of-Distribution' (OOD) risks?

Accepted Answer

Anomaly scores derived from the model's embedding space or activation layers.. Out-of-distribution inputs produce anomalous patterns in the model's internal representations — embedding vectors or activation layers — so anomaly scores derived from these spaces (e.g., Mahalanobis distance, reconstruction error, or energy scores) are the most direct signal of OOD risk. These scores detect inputs that differ semantically from training distribution even when surface features look normal. This is the standard OOD detection approach in production ML monitoring.

Answer

Total number of users accessing the model.

Answer

Average request latency.

Answer

The number of lines of code in the model file.

Question 15

You are performing a 'Model Drift' analysis. Which metric is most indicative of performance degradation without access to real-time ground truth?

Accepted Answer

Kullback-Leibler (KL) Divergence between training and inference feature distributions.. Statistical distribution shifts (like KL Divergence) in input data serve as a proxy for performance degradation.

Answer

The number of concurrent API requests.

Answer

The memory consumption of the container.

Answer

Model accuracy on historical data.

Question 16

You are integrating AI governance into the existing Corporate Governance framework. Which THREE of the following activities are essential to ensure the Board of Directors maintains adequate oversight?

Accepted Answer

Reviewing and approving the organization's AI Risk Appetite Statement.. Option A is correct because the Board's core governance role includes setting and approving the organization's risk appetite, and an AI Risk Appetite Statement translates that mandate into explicit tolerance thresholds for AI-specific risks (e.g., bias, model drift, data privacy), giving the Board a formal basis for oversight. Option C is correct because designating executive-level accountability (e.g., a Chief AI Officer or an AI risk owner at the C-suite) establishes clear ownership and escalation paths, which is a fundamental governance control ensuring the Board can hold management responsible for AI risk outcomes. Option E is correct because a defined reporting cadence for significant AI-related risk incidents and model performance trends provides the Board with timely, recurring visibility into AI risk posture, enabling informed oversight and intervention consistent with corporate governance principles. Options B and D are not Board-level governance activities: patching GPU servers is an operational IT maintenance task, and daily code reviews of training algorithms are tactical engineering controls, both of which belong to management and technical teams rather than the Board's oversight responsibilities.

Answer

Installing software patches on all GPU servers.

Answer

Conducting daily code reviews of all AI training algorithms.

Question 17

Which THREE factors contribute to 'Data Quality Risk' during the training phase?

Accepted Answer

Incomplete datasets with many missing values.. Data Quality Risk during training arises from defects in the data itself, and option A is correct because incomplete datasets with many missing values degrade the statistical representativeness of the training sample, causing the model to learn from biased or insufficient signal. Option B is correct because systematic errors in target label annotations inject consistent noise into the ground truth, so the model optimizes toward incorrect labels and produces unreliable predictions. Option D is correct because selection bias in the training set creation process means the sampled data does not reflect the real-world distribution the model will encounter, skewing learned patterns. Option C is not a data quality issue but a hardware/resource constraint (GPU memory capacity), and option E is an infrastructure/vendor choice unrelated to the intrinsic quality of the training data.

Answer

The model is too large for the GPU memory.

Answer

Using the wrong cloud vendor.

Question 18

When documenting an AI Risk Program, what should be included in the 'AI Asset Register'?

Accepted Answer

The model purpose, data lineage, stakeholders, and risk classification.. An AI Asset Register is an inventory of AI systems and models that captures governance-relevant metadata: the model's purpose, the data sources and lineage used to train it, the stakeholders accountable for it, and its risk classification. This information enables risk management, regulatory reporting, and lifecycle oversight across the AI portfolio.

Answer

The raw hardware serial numbers of the servers.

Answer

The exact lines of code in the model.

Answer

The home addresses of all developers.

Question 19

Which THREE factors should be monitored to detect 'Model Drift' in the deployment phase?

Accepted Answer

The distribution of prediction outputs.. Model drift detection in the deployment phase focuses on changes in data and model behavior over time. Option A, the distribution of prediction outputs, is correct because a shift in the distribution of predictions (e.g., a classifier suddenly predicting one class far more often) is a direct symptom of model drift, often caused by changing input data or degraded decision boundaries. Option D, the confidence score of the model predictions, is correct because declining or abnormally shifting confidence scores indicate the model is becoming less certain and less aligned with the data it was trained on, a common signal of drift. Option E, the distribution of incoming input features, is correct because monitoring feature distributions (e.g., via statistical tests like PSI or KL divergence) detects data drift, which is the primary driver of model drift in production. Option B, the temperature of the server room, is incorrect because it is a physical infrastructure concern unrelated to statistical model behavior. Option C, the current market price of GPU hardware, is incorrect because it is a procurement/cost factor with no bearing on whether a deployed model's performance is degrading.

Answer

The temperature of the server room.

Answer

The current market price of GPU hardware.

Question 20

Which THREE items should be included in an AI 'Risk Assessment' report for a new project?

Accepted Answer

Identification of potential failure modes and their impact.. A complete report covers the risk identification, the impact analysis, and the control strategy.

Answer

The favorite color of the lead developer.

Answer

A list of all competitors currently using the same software.