Question bank
AI Governance Professional
40 AI Governance Professional questions with the correct answer and a written explanation, organised across 4 exam domains.
- 40 questions
- 4 domains
- Foundation level
certilumQuestion index
Jump straight to any question. Answers and explanations are inline.
Understanding how laws, standards, and frameworks apply to AI
10 questions in this domain
- Question 1Single Choice
What is the defining characteristic of supervised learning?
- AModels learn by trial and error with a reward function.
- BModels are trained on labeled datasets containing input features and desired outputs.Correct answer
- CModels find hidden patterns in unlabeled data.
- DModels generate synthetic data from learned distributions.
Why this is the answer
Supervised learning is defined by training on labeled data with both input features and desired output. The other options describe different learning paradigms.
- Question 2Multiple Select
Which factors are primary risk vectors for introducing systemic bias in supervised learning? Select all that apply.
- AQuality of the ground truth dataCorrect answer
- BRepresentativeness of the training dataCorrect answer
- CLabeling accuracyCorrect answer
- DNumber of hidden layers in the model
Why this is the answer
The playbook identifies ground truth quality, representativeness, and labeling accuracy as primary risk vectors for bias. Model architecture is not the cause identified here.
- Question 3Interactive
Based on the configuration in the code block, which machine learning paradigm is described?
- ASupervised Learning
- BUnsupervised Learning
- CReinforcement LearningCorrect answer
- DGenerative Modeling
Why this is the answer
The configuration shows an agent using a reward function and trial-and-error feedback, which is the defining structure of reinforcement learning.
- Question 4Single Choice
In machine learning, what happens during the training phase?
- AThe model learns from the data.Correct answer
- BThe model is applied to new, unseen data.
- CThe model is validated against production traffic.
- DThe model is tested for drift.
Why this is the answer
Training is the phase where a model learns from data. Applying the model to unseen data occurs during inference.
- Question 5Multiple Select
According to governance guidance, which risks are associated with the inference phase? Select all that apply.
- AModel driftCorrect answer
- BReal-time reliabilityCorrect answer
- CData privacy during training
- DBias in training labels
Why this is the answer
Inference risks involve model drift and real-time reliability. Data privacy and training-label bias are risks associated with the training phase.
- Question 6Interactive
Given the accuracy metrics in the code block, which condition is most likely present?
- AOverfittingCorrect answer
- BUnderfitting
- CModel drift
- DReward hacking
Why this is the answer
High training accuracy with much lower validation accuracy indicates the model memorized noise in the training data rather than generalizing.
- Question 7Single Choice
A biased model shows high accuracy. What is the most effective initial governance response?
- AReplace the algorithm with a more advanced model.
- BInspect the training data for quality and labeling issues.Correct answer
- CAdd more hidden layers to the model.
- DIgnore the bias because accuracy is strong.
Why this is the answer
Bias is often rooted in the training data, not the architecture. Improving data quality and labeling accuracy is the better governance response.
- Question 8Multiple Select
Which statements align with the governance treatment of the black box fallacy? Select all that apply.
- AXAI tools make it possible to implement governance for complex models.Correct answer
- BDeep learning models cannot be governed and should be abandoned.
- CThe opacity of deep learning is an acceptable reason to ignore the model.
- DGovernance requires explainability measures rather than abandoning the technology.Correct answer
Why this is the answer
The black box fallacy is the belief that complex models are inherently unmanageable. XAI tools and explainability measures allow governance, not abandonment.
- Question 9Interactive
What phenomenon is most likely indicated by the accuracy trend and changing data distribution in the code block?
- AModel driftCorrect answer
- BUnderfitting
- CReward hacking
- DHallucination
Why this is the answer
Declining performance over time while the real-world data distribution changes is the natural statistical phenomenon known as model drift.
- Question 10Single Choice
How does a discriminative model make predictions?
- AIt focuses on the boundary between classes to distinguish data points.Correct answer
- BIt focuses on the full distribution of the data to capture how the data was created.
- CIt uses a reward function to generate new examples.
- DIt relies on unlabeled data to form clusters.
Why this is the answer
Discriminative models focus on the boundary between classes to distinguish data points, while the other choices describe generative or unsupervised behavior.
Understanding how to govern AI deployment and use
12 questions in this domain
- Question 11Single Choice
In supervised learning, what are the primary governance risk vectors for introducing systemic bias?
- AQuality, representativeness, and labeling accuracy of ground truth dataCorrect answer
- BComplexity of the model architecture
- CVolume of data generated during inference
- DDesign of the agent's reward function
Why this is the answer
Supervised learning relies on labeled ground truth, so data quality, representativeness, and labeling accuracy are the key bias risk vectors.
- Question 12Multiple Select
Which governance considerations apply to unsupervised learning? Select all that apply.
- AInterpretability of the hidden patterns it discoversCorrect answer
- BAuditing clusters for unintended discriminatory correlationsCorrect answer
- CLabeling accuracy of ground truth data
- DReward hacking prevention
Why this is the answer
Unsupervised learning uses unlabeled data; governance centers on interpretability and auditability of discovered patterns, not labels or rewards.
- Question 13Interactive
The provided code block defines a learning signal for an AI agent. Which governance risk is most directly associated with this mechanism?
- AReward hackingCorrect answer
- BOverfitting
- CModel drift
- DSynthetic data
Why this is the answer
Reinforcement learning relies on reward functions, and reward hacking occurs when the agent meets the objective through harmful shortcuts.
- Question 14Single Choice
Which governance risk is most associated with the inference phase?
- AModel drift and real-time reliabilityCorrect answer
- BData privacy and bias from training data
- CLabeling accuracy of ground truth
- DCopyright of training content
Why this is the answer
Training risks involve privacy and bias; inference risks center on model drift and real-time reliability.
- Question 15Multiple Select
Which statements about overfitting and underfitting are correct? Select all that apply.
- AOverfitting means the model memorizes noise in the training data and fails to generalize.Correct answer
- BUnderfitting means the model fails to capture the underlying trend.Correct answer
- CBoth conditions create reliability risks in production.Correct answer
- DBoth conditions are caused by reward hacking.
Why this is the answer
Overfitting memorizes noise, underfitting misses trends, and both compromise accurate, predictable production outcomes.
- Question 16Interactive
The provided code block represents a phase in the AI lifecycle. Which governance risks should be reviewed for this phase?
- AData privacy and biasCorrect answer
- BReal-time reliability
- CProvenance of generated content
- DReward hacking
Why this is the answer
The code block shows training, where the main risks are data privacy and bias in the training data.
- Question 17Single Choice
Which statement correctly describes model drift?
- AIt is a natural statistical phenomenon caused by changing real-world data distributions.Correct answer
- BIt is a software bug in the model code.
- CIt can be permanently fixed by selecting a better algorithm.
- DIt only occurs in generative AI.
Why this is the answer
Model drift is not a code error; it reflects natural degradation caused by changes in real-world data distributions.
- Question 18Multiple Select
Which of the following are common AI governance traps? Select all that apply.
- AConfusing AI with machine learningCorrect answer
- BAssuming accuracy equals fairnessCorrect answer
- CTreating hallucination as only a bugCorrect answer
- DSelecting technical validation as a governance structureCorrect answer
Why this is the answer
The playbook identifies all of these as common traps, including conceptual confusion, fairness assumptions, hallucination framing, and misuse of validation.
- Question 19Interactive
The provided code block shows an optimization goal. Which model category does this objective represent?
- ADiscriminative model minimizing classification errorCorrect answer
- BGenerative model maximizing data likelihood
- CReinforcement learning agent maximizing reward
- DUnsupervised clustering model
Why this is the answer
Discriminative models aim to minimize classification error; the code compares predictions against labels.
- Question 20Single Choice
From a probability modeling perspective, what is the difference between discriminative and generative models?
- ADiscriminative models focus on boundaries between classes; generative models focus on the distribution of the data.Correct answer
- BDiscriminative models generate text; generative models classify labels.
- CDiscriminative models use unlabeled data; generative models use labeled data.
- DThere is no meaningful difference.
Why this is the answer
Discriminative models distinguish classes, while generative models model the distribution to capture how data was created.
- Question 21Multiple Select
Which governance concerns are expanded or emphasized in generative AI compared to discriminative AI? Select all that apply.
- AProvenance and copyrightCorrect answer
- BMitigation of non-factual outputsCorrect answer
- CHallucinated or stereotyped contentCorrect answer
- DAccuracy, fairness, and explainability as the only focus
Why this is the answer
Generative AI governance expands to provenance, copyright, and non-factual outputs instead of focusing only on accuracy and explainability.
- Question 22Interactive
The provided code block describes an output emitted by a model. Which AI category does this output pattern best represent?
- AGenerative AICorrect answer
- BDiscriminative AI
- CUnsupervised learning
- DReinforcement learning
Why this is the answer
High-dimensional, complex outputs such as images are characteristic of generative AI.
Understanding how to govern AI development
10 questions in this domain
- Question 23Single Choice
Which machine learning paradigm trains models on labeled datasets that contain both input features and the desired output for each example?
- ASupervised learningCorrect answer
- BUnsupervised learning
- CReinforcement learning
- DGenerative AI
Why this is the answer
Supervised learning uses labeled datasets with inputs and desired outputs. Governance must audit ground truth quality because labeling errors can introduce systemic bias.
- Question 24Multiple Select
Which governance risks are specifically associated with unsupervised learning, according to the risk profile in the playbook?
- ALack of interpretability of discovered patternsCorrect answer
- BUnintended discriminatory correlations in clustersCorrect answer
- CReward hacking from a malfunctioning reward function
- DLabeling inaccuracies in ground truth
Why this is the answer
Unsupervised learning risks focus on interpretability and hard-to-audit clusters that may encode discriminatory correlations. Reward hacking belongs to reinforcement learning; labeling issues belong to supervised learning.
- Question 25Interactive
The code block shows an agent selecting actions and receiving rewards while updating its policy. Which learning paradigm does this represent?
- AReinforcement learningCorrect answer
- BSupervised learning
- CUnsupervised learning
- DDiscriminative AI
Why this is the answer
The loop uses actions, rewards, and policy updates, which is the hallmark of reinforcement learning. Governance should watch for reward hacking in such systems.
- Question 26Single Choice
How should model drift be interpreted from a governance perspective, according to the playbook's description?
- AA statistical phenomenon from changing data distributionsCorrect answer
- BA software bug requiring code changes
- CThe same as overfitting during training
- DA labeling error in the training dataset
Why this is the answer
Model drift is a natural statistical phenomenon from changing real-world data, not a code bug. Governance should treat it as an ongoing monitoring concern.
- Question 27Multiple Select
Which statements correctly match governance risks with either the training phase or the inference phase of a model lifecycle?
- AData privacy is a training-phase riskCorrect answer
- BModel drift is an inference-phase riskCorrect answer
- CReal-time reliability is a training-phase risk
- DLabeling accuracy is an inference-phase risk
Why this is the answer
Training risks include data privacy and bias, while inference risks include model drift and real-time reliability. The other pairings reverse these relationships.
- Question 28Interactive
The code block reports training accuracy as high and validation accuracy as lower. Which model condition does this pattern suggest?
- AOverfittingCorrect answer
- BUnderfitting
- CModel drift
- DReward hacking
Why this is the answer
High training accuracy with lower validation accuracy indicates the model memorized noise and failed to generalize, which is overfitting.
- Question 29Single Choice
What belief characterizes the governance trap of confusing data quality with model architecture when outcomes are biased?
- AA better algorithm automatically fixes biased outcomesCorrect answer
- BBias is usually rooted in the model architecture
- CData quality is irrelevant once the model is deployed
- DAlgorithm choice has no effect on bias
Why this is the answer
The trap is assuming a better algorithm fixes biased outcomes; bias often is rooted in the training data itself.
- Question 30Multiple Select
Which statements reflect sound governance responses to the Black Box Fallacy for highly complex deep learning models?
- AImplement explainability (XAI) toolsCorrect answer
- BGovern the model through appropriate explainability methodsCorrect answer
- CAbandon complex models because they cannot be governed
- DTreat high complexity as making the model unmanageable
Why this is the answer
The Black Box Fallacy is the belief that complex models are inherently unmanageable. Governance should use XAI tools to address complexity.
- Question 31Interactive
The code block shows one model returning a classification result and another model returning generated content. Which output distinction applies?
- AA is discrete; B is complex and high-dimensionalCorrect answer
- BA is high-dimensional; B is discrete
- CBoth outputs are discrete labels
- DBoth outputs are synthetic content
Why this is the answer
Discriminative AI produces discrete outputs like labels; generative AI produces high-dimensional outputs like text, images, or audio.
- Question 32Single Choice
In probability modeling, what do discriminative models primarily focus on when they distinguish between classes of data points?
- AThe boundary between classesCorrect answer
- BThe distribution of the data itself
- CThe reward function design
- DThe generation of synthetic examples
Why this is the answer
Discriminative models focus on the boundary between classes to distinguish data points, while generative models capture the data distribution.
Understanding the foundations of AI governance
8 questions in this domain
- Question 33Single Choice
Which machine learning paradigm uses labeled datasets that contain both input features and the desired output?
- ASupervised LearningCorrect answer
- BUnsupervised Learning
- CReinforcement Learning
- DSemi-supervised Learning
Why this is the answer
Supervised learning trains on labeled ground truth. Governance risk vectors include the quality, representativeness, and labeling accuracy of the ground truth data.
- Question 34Multiple Select
Which governance risks are associated with reinforcement learning? Select all that apply.
- AReward hacking, where the agent achieves the mathematical objective through unintended shortcutsCorrect answer
- BAgent behavior that violates human safety constraints while optimizing the reward functionCorrect answer
- CInaccurate ground truth labels introducing systemic bias
- DHidden clusters in unlabeled data that are difficult to audit
Why this is the answer
Reinforcement learning governance risks center on reward hacking and agents taking unintended or harmful shortcuts that violate human safety constraints. Label bias is a supervised learning concern, and hidden clusters are an unsupervised learning concern.
- Question 35Interactive
Review the code excerpt below and select the governance concern that should be evaluated.
- AThe agent may find unintended shortcuts that maximize reward but violate safety constraintsCorrect answer
- BThe model is memorizing noise in the training data
- CThe ground truth labels are inaccurate
- DThe model cannot generalize to new inputs
Why this is the answer
A reward function that only rewards reaching the target can lead to reward hacking, where the agent achieves the mathematical objective through unintended or harmful shortcuts that violate safety constraints.
- Question 36Single Choice
During the inference phase, what is the model doing?
- AIt applies what it learned during training to new, unseen dataCorrect answer
- BIt learns from labeled training examples
- CIt creates synthetic data for future training
- DIt evaluates the loss function on training data
Why this is the answer
Inference is the phase where the model applies its learned patterns to new, unseen data. Governance risks at inference include model drift and real-time reliability.
- Question 37Multiple Select
Which risks require governance oversight during the inference phase? Select all that apply.
- AModel driftCorrect answer
- BReal-time reliabilityCorrect answer
- CBias in training labels
- DPrivacy violations during model training
Why this is the answer
Inference risks include model drift and real-time reliability. Bias in training labels and privacy violations during training are primarily training-phase concerns.
- Question 38Interactive
Review the code output comparison below and determine which reliability risk is indicated.
- AOverfitting: the model memorized noise in the training data and failed to generalizeCorrect answer
- BUnderfitting: the model failed to capture the underlying trend
- CModel drift: the real-world data distribution changed
- DHallucination: the model generated non-factual content
Why this is the answer
A large gap between high training accuracy and lower validation accuracy indicates overfitting, where the model memorizes training noise and fails to generalize to unseen data.
- Question 39Single Choice
What is the key governance trap associated with model drift?
- ATreating model degradation as a software bug instead of a natural statistical phenomenonCorrect answer
- BImplementing explainability tools before deployment
- CConfusing training data quality with model architecture
- DUsing unsupervised learning instead of supervised learning
Why this is the answer
The trap is misinterpreting model drift: it should be treated as a natural statistical phenomenon caused by changing real-world data distributions, not as a code error or software bug.
- Question 40Multiple Select
Which statements accurately describe discriminative and generative models? Select all that apply.
- ADiscriminative models focus on the boundary between classes to distinguish data pointsCorrect answer
- BGenerative models focus on the distribution of the data itself to capture how the data was createdCorrect answer
- CGenerative AI can amplify biases by creating hallucinated or stereotyped contentCorrect answer
- DDiscriminative models maximize the likelihood of generated samples matching the training data distribution
Why this is the answer
Discriminative models separate classes, while generative models capture data distributions. Generative AI can amplify bias through hallucinated or stereotyped content. Maximizing likelihood of generated samples is a generative objective, not a discriminative one.
Frequently asked questions
Everything about the AI Governance Professional question bank and how to use it.