I85PREP
⌘ K
PECB Certified Artificial Intelligence Professional study materials
PECB Certified Artificial Intelligence Professional study resources

PECB Certified Artificial Intelligence Professional practice questions

One question at a time. Build your PECB Certified Artificial Intelligence Professional understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Artificial Intelligence

In the ISO/IEC 22989 vocabulary, which statement best captures the relationship between an AI system and its environment?

  1. An AI system is any database that stores data about an environment and returns it unchanged on request.
  2. An AI system is a statistical model that can only be evaluated offline and never interacts with its environment.
  3. An AI system is a software or hardware system that, for explicit or implicit objectives, infers from received input how to generate outputs such as predictions that can influence its environment. Correct answer.
  4. An AI system is a physical robot whose behaviour is entirely scripted by a human operator in real time.

Explanation

The standard definition of an AI system emphasises inference from input toward explicit or implicit objectives, producing outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. It does not require a physical embodiment, nor does it exclude models evaluated offline.

Question 2Artificial Intelligence

A team must decide whether a task is a good candidate for machine learning rather than hand-coded rules. Which characteristic most strongly favours a machine learning approach?

  1. The output must be provably identical to a specification for every possible input.
  2. The problem involves a fixed, closed set of inputs with no historical examples.
  3. The rules are stable, fully documented and simple to express as conditionals.
  4. The underlying patterns are complex, change over time and are difficult to specify explicitly. Correct answer.

Explanation

Machine learning is valuable when the relationship between inputs and outputs is complex, dynamic or hard to express as explicit rules, and when representative data is available to learn from. Simple, stable, fully specified logic is better served by conventional programming.

Question 3Artificial Intelligence

Which description correctly distinguishes supervised from unsupervised learning?

  1. Supervised learning clusters data into groups, while unsupervised learning predicts a continuous target value.
  2. Supervised learning requires no data, while unsupervised learning requires labelled data.
  3. Supervised learning always uses neural networks, while unsupervised learning always uses decision trees.
  4. Supervised learning uses labelled examples to map inputs to targets, while unsupervised learning finds structure in unlabelled data. Correct answer.

Explanation

Supervised learning learns a mapping from inputs to known target labels, whereas unsupervised learning seeks patterns such as clusters or latent structure in data without target labels. The choice of algorithm family is independent of this distinction.

Question 4Artificial Intelligence

During training, a model performs very well on the training set but poorly on held-out validation data. What does this most directly indicate?

  1. The model has overfitted and is not generalising to unseen data. Correct answer.
  2. The learning rate was set to zero for the entire run.
  3. The validation set is too large relative to the training set.
  4. The model has underfitted because it has not learned the training data.

Explanation

A large gap between strong training performance and weak validation performance is the classic signature of overfitting, where the model captures noise specific to the training sample. The remedy lies in regularisation, more or better data, or reduced model capacity.

Question 5Artificial Intelligence

What is the primary purpose of splitting data into training, validation and test sets?

  1. To remove the need for any monitoring after deployment.
  2. To estimate generalisation performance on data not used for fitting or tuning. Correct answer.
  3. To increase the total volume of data available to the model.
  4. To guarantee that the model cannot make errors on future inputs.

Explanation

Holding out data lets practitioners tune hyperparameters on validation data and obtain an unbiased estimate of performance on the untouched test set. This estimates generalisation rather than guaranteeing correctness.

Question 6Artificial Intelligence

Which statement best describes the role of a loss function in training a model?

  1. It encrypts model parameters before they are written to disk.
  2. It defines the hardware resources allocated to each training job.
  3. It quantifies the discrepancy between model predictions and target values so that parameters can be optimised. Correct answer.
  4. It stores the training data in compressed form for later reuse.

Explanation

A loss function converts prediction errors into a scalar objective that optimisation algorithms such as gradient descent minimise by adjusting parameters. It is the mathematical signal that drives learning, not a storage or infrastructure mechanism.

Question 7Artificial Intelligence

A recruitment screening tool shows much lower selection rates for one demographic group than another. Which concept does this scenario most directly illustrate?

  1. Overfitting of the model to the training data.
  2. Inference latency exceeding the service level objective.
  3. Bias in model outcomes affecting a protected group. Correct answer.
  4. Model drift caused by seasonal input changes.

Explanation

Disparate selection rates across demographic groups are a fairness concern, commonly described as outcome bias or disparate impact. Detecting it requires disaggregated performance measurement across groups, not just aggregate accuracy.

Question 8Artificial Intelligence

Which practice best supports the principle of data minimisation in an AI project?

  1. Sharing the full dataset with all third parties to increase transparency.
  2. Collecting and retaining only the attributes needed for the stated purpose and for the defined retention period. Correct answer.
  3. Retaining raw personal data indefinitely to allow future retraining.
  4. Collecting every available attribute in case it proves useful later.

Explanation

Data minimisation limits collection and retention to what is adequate, relevant and necessary for the specified purpose, with a defined retention period. Collecting or retaining more than needed increases privacy risk without a justified purpose.

Question 9Artificial Intelligence

A bank wants an AI system to explain each declined credit application to the applicant. Which requirement does this most directly impose on the solution?

  1. The system must provide local, case-level explanations of the decision factors. Correct answer.
  2. The system must run entirely on edge devices.
  3. The model must be a deep neural network with many layers.
  4. The model must be trained without any historical lending data.

Explanation

Explaining an individual decision requires local interpretability, such as reason codes or feature attributions for that case. Global model transparency or a particular architecture does not by itself satisfy a case-level explanation duty.

Question 10Artificial Intelligence

Which pair of metrics is most appropriate for evaluating a binary classifier on a highly imbalanced dataset?

  1. R-squared and adjusted R-squared.
  2. Precision and recall, summarised by the F1 score or the precision-recall curve. Correct answer.
  3. Silhouette score and Davies-Bouldin index.
  4. Accuracy and mean squared error.

Explanation

On imbalanced data, accuracy can look high while the minority class is ignored, so precision, recall and their combination are more informative. A precision-recall curve shows the trade-off across thresholds.

Question 11Artificial Intelligence

A deployed model's input distribution gradually shifts as customer behaviour changes, degrading its predictions. What is the most appropriate response?

  1. Increase the batch size used during inference.
  2. Immediately delete the model and stop the service.
  3. Monitor input and output distributions, detect drift and trigger retraining or recalibration. Correct answer.
  4. Assume the model is still valid because the code has not changed.

Explanation

Data drift means the statistical properties of inputs change relative to training, so ongoing monitoring of distributions and performance is needed to detect degradation. Detected drift should trigger investigation and retraining or recalibration under governance.

Question 12Artificial Intelligence

Which statement correctly describes the difference between classification and regression tasks?

  1. Classification and regression differ only in the choice of programming language.
  2. Classification predicts a continuous value, while regression predicts a discrete label.
  3. Classification predicts a discrete label, while regression predicts a continuous numeric value. Correct answer.
  4. Classification requires unlabelled data, while regression requires labelled data.

Explanation

Classification assigns inputs to discrete categories, whereas regression estimates continuous quantities such as price or demand. Both are typically supervised tasks and both can be implemented with the same tooling.

Question 13Artificial Intelligence

When is stratified k-fold cross-validation preferable to plain k-fold cross-validation?

  1. When the target variable is not defined for the task.
  2. When class proportions are imbalanced and must be preserved in each fold. Correct answer.
  3. When the model has no hyperparameters to tune.
  4. When the dataset is extremely large and training time is the main constraint.

Explanation

Stratification keeps the class distribution consistent across folds, which reduces variance in the performance estimate when one class is rare. Plain k-fold can produce folds with very few minority examples.

Question 14Artificial Intelligence

Which control best mitigates the risk of prompt injection in an LLM-based application that processes untrusted user content?

  1. Treat retrieved and user-supplied content as untrusted data, isolate it from instructions and constrain tool permissions. Correct answer.
  2. Rely solely on the model's built-in safety training and remove all external validation.
  3. Increase the model temperature so that responses are more creative.
  4. Store the system prompt in the same channel as user input so the model can compare them.

Explanation

Prompt injection exploits the blending of trusted instructions with untrusted content, so defences separate instruction and data channels and limit what tools the model can invoke. Least-privilege tool access and output validation reduce impact when injection succeeds.

Question 15Artificial Intelligence

What is the main benefit of using SHAP values to explain a model's predictions?

  1. They guarantee that the model is fair with respect to all protected groups.
  2. They convert any regression model into a classifier.
  3. They eliminate the need for a validation dataset.
  4. They attribute a prediction to individual features using a game-theoretic notion of contribution. Correct answer.

Explanation

SHAP values allocate the difference between a prediction and a baseline among features based on Shapley values from cooperative game theory. They explain contributions but do not certify fairness or change the task type.

Question 16Artificial Intelligence

Which statement best describes the concept of a human oversight mechanism in an AI management system?

  1. A mechanism that only applies to physical robots in manufacturing.
  2. A mechanism that removes humans from the decision loop to increase throughput.
  3. A mechanism that automatically approves every model update without review.
  4. A mechanism that lets a human review, intervene in or override AI system decisions where appropriate. Correct answer.

Explanation

Human oversight means designing the socio-technical system so that people can monitor, intervene and, where necessary, override AI outputs. The appropriate degree of oversight is proportionate to the risk and context of use.

Question 17Artificial Intelligence

An organisation wants to deploy a high-risk AI system in a regulated sector. Which sequence best reflects a sound governance approach?

  1. Perform a one-off risk assessment at deployment and never revisit it.
  2. Deploy first, then perform a risk assessment only if an incident occurs.
  3. Classify the use case and risk, assess impacts, implement controls, document evidence, then deploy with ongoing monitoring. Correct answer.
  4. Skip documentation to accelerate delivery and rely on vendor assurances.

Explanation

Sound governance front-loads risk classification and impact assessment, then applies proportionate controls with documented evidence before deployment. Because systems and contexts change, monitoring and periodic review continue after go-live.

Question 18Artificial Intelligence

Which practice best reduces the risk of data leakage between training and test sets in a machine learning pipeline?

  1. Duplicating near-identical records across both splits to increase sample size.
  2. Fitting the scaler on the full dataset and then splitting into train and test.
  3. Tuning hyperparameters repeatedly on the test set until performance improves.
  4. Splitting the data before any preprocessing that learns parameters, such as scaling or imputation. Correct answer.

Explanation

Preprocessing steps that estimate parameters from data must be fitted only on the training portion, otherwise information from the test set leaks into the model. Splitting first and fitting transformations within each fold preserves an honest estimate.

Question 19Artificial Intelligence

What is the primary function of a feature store in an ML platform?

  1. To store the model's source code in a version control system.
  2. To serve consistent, reusable feature definitions for both training and inference. Correct answer.
  3. To provide GPU acceleration for model training.
  4. To replace the need for a data catalog.

Explanation

A feature store centralises feature computation and serves the same definitions online and offline, which helps prevent training-serving skew and encourages reuse. It complements, rather than replaces, data cataloguing and compute infrastructure.

Question 20Artificial Intelligence

A team is choosing between a simpler interpretable model and a more accurate complex model for a context where decisions must be explained to regulators. Which reasoning is most sound?

  1. Always choose the simplest model regardless of the performance loss.
  2. Always choose the complex model because accuracy outweighs every other consideration.
  3. Weigh the accuracy gain against explainability requirements and consider post hoc explanation or a constrained model. Correct answer.
  4. Choose based on which model has more hyperparameters.

Explanation

Model selection is a trade-off among performance, interpretability and regulatory obligations. Where explanations are required, a slightly less accurate but transparent model, or a complex model with validated post hoc explanations, may be the defensible choice.

Question 21Artificial Intelligence

In a reinforcement learning setup, what distinguishes the reward signal from the environment state?

  1. The reward is defined only by the agent's source code, while the state is defined only by the learning rate.
  2. The reward is a scalar feedback indicating desirability of an action, while the state describes the situation the agent observes. Correct answer.
  3. The reward is the full observation vector, while the state is a scalar score.
  4. The reward and state are identical and always interchangeable.

Explanation

In reinforcement learning the agent observes a state, takes an action and receives a scalar reward that signals how good the outcome was. The state carries situational information, whereas the reward drives policy improvement.

Question 22Artificial Intelligence

A vendor claims its model is robust because accuracy on a random test split is high. Which additional evaluation is most relevant before deployment in an adversarial setting?

  1. Increasing the number of classes in the output layer.
  2. Testing performance under adversarial perturbations and distribution shift, with subgroup analysis. Correct answer.
  3. Measuring only training time and memory footprint.
  4. Removing the validation set to speed up evaluation.

Explanation

High average accuracy on an in-distribution split says little about behaviour under deliberate perturbation or shifted data. Adversarial testing and subgroup analysis reveal failure modes that aggregate accuracy hides.

Question 23Artificial Intelligence

An organisation applies differential privacy to a model trained on personal data. What does the privacy budget epsilon control?

  1. The size of the neural network's hidden layers.
  2. The trade-off between privacy protection and utility, with smaller values giving stronger privacy. Correct answer.
  3. The encryption key length used for data at rest.
  4. The number of training epochs before convergence.

Explanation

Epsilon bounds how much any single record can influence the output, so smaller values mean stronger privacy guarantees and typically more noise and lower utility. It is a formal privacy parameter, not a compute or encryption setting.

Question 24Artificial Intelligence

A model trained on historical hiring decisions reproduces past discriminatory patterns. Which intervention most directly addresses the root cause rather than the symptom?

  1. Removing the protected attribute from the feature set and assuming the problem is solved.
  2. Auditing label generation and historical decision processes, and correcting biased targets or sampling before retraining. Correct answer.
  3. Increasing model complexity so it captures more of the historical pattern.
  4. Lowering the decision threshold uniformly for all applicants.

Explanation

Bias encoded in labels or sampling reflects the historical process that produced the data, so remedying it requires auditing and correcting target definitions and data collection. Simply dropping a protected attribute leaves proxy variables that still carry the bias.

Question 25Artificial Intelligence

Which statement best characterises the relationship between an AI management system and AI risk management in ISO/IEC 42001 and the ISO/IEC 23894 guidance?

  1. An AI management system guarantees that deployed models will be free of bias and error.
  2. Risk management replaces the need for an AI management system because risk registers are sufficient.
  3. Risk management applies only to generative models and not to other AI systems.
  4. A management system provides the organisational framework of policies, roles and continual improvement within which risk management processes are applied to AI. Correct answer.

Explanation

An AI management system establishes the governance structure, responsibilities and continual improvement cycle, while risk management supplies the process for identifying, assessing and treating AI-specific risks within it. The two are complementary, and neither guarantees perfect outcomes.

Question 26Artificial Intelligence

In supervised learning, what does the term 'label' refer to?

  1. The hyperparameter controlling learning rate
  2. The set of input features fed to the model
  3. The random seed used to split the dataset
  4. The target output associated with a training example Correct answer.

Explanation

A label is the ground-truth target value attached to each training example, which the model learns to predict from the input features. Features are the inputs, while learning rate and random seed are training configuration settings rather than data labels.

Question 27Artificial Intelligence

Which activity is characteristic of the inference phase rather than the training phase of a machine learning lifecycle?

  1. Tuning hyperparameters on a validation set
  2. Computing gradients over a mini-batch
  3. Updating model weights through backpropagation
  4. Serving predictions on new inputs using fixed parameters Correct answer.

Explanation

Inference uses a trained model with frozen parameters to produce predictions on new data. Backpropagation, gradient computation and hyperparameter tuning all belong to training or model development, where parameters or configuration are still being adjusted.

Question 28Artificial Intelligence

An AI system is described as a 'narrow' or 'weak' AI. Which capability does this description imply?

  1. It requires no training data to function
  2. It can transfer general reasoning to any unseen task
  3. It is limited to a specific task or domain Correct answer.
  4. It possesses conscious self-awareness

Explanation

Narrow AI is designed to perform a defined task or narrow set of tasks well, such as image classification or recommendation. General reasoning across arbitrary tasks and self-awareness describe hypothetical general or strong AI, which current deployed systems do not exhibit.

Question 29Artificial Intelligence

Which statement best describes the purpose of a validation set during model development?

  1. To guide model selection and hyperparameter choices without touching the test set Correct answer.
  2. To permanently retrain the model after deployment
  3. To store raw data before any preprocessing occurs
  4. To provide the final unbiased estimate of production performance

Explanation

A validation set is used to compare model configurations and tune hyperparameters, keeping the test set untouched for a final estimate. Reusing the test set for tuning would leak information and bias the reported performance.

Question 30Artificial Intelligence

Which practice most directly supports accountability in an AI system's operation?

  1. Deploying without a designated responsible owner
  2. Increasing the model's parameter count
  3. Removing logging to reduce storage costs
  4. Documenting decisions, owners and change history in an audit trail Correct answer.

Explanation

Accountability requires that decisions and changes can be attributed to named owners and reviewed after the fact, which an audit trail with decision logs and change history provides. Larger models or reduced logging do not create accountability and can undermine it.

Question 31Artificial Intelligence

What is the primary purpose of anonymisation or pseudonymisation in an AI training dataset?

  1. To increase the model's accuracy on the test set
  2. To convert labels into one-hot vectors
  3. To compress the dataset for faster storage
  4. To reduce the risk of identifying individuals in the data Correct answer.

Explanation

Anonymisation and pseudonymisation reduce the link between records and identifiable individuals, lowering privacy risk when data is used for training. They are privacy safeguards, not accuracy, compression or encoding techniques.

Question 32Artificial Intelligence

A team needs to compare several candidate models quickly before committing to full evaluation. Which approach fits this goal?

  1. Deploying all candidates to production simultaneously
  2. Ignoring validation results and using intuition
  3. Skipping evaluation and choosing the largest model
  4. Running a small pilot evaluation on a representative sample Correct answer.

Explanation

A pilot evaluation on a representative sample gives early comparative signal at low cost, allowing the team to shortlist candidates before full evaluation. Skipping evaluation, deploying all candidates, or relying on intuition removes the evidence needed for a sound choice.

Question 33Artificial Intelligence

Which element is essential for a meaningful AI impact assessment before deployment?

  1. The number of employees in the IT department
  2. Identification of affected stakeholders and potential harms Correct answer.
  3. The marketing budget for the system
  4. A list of competitor product names

Explanation

An impact assessment examines who is affected by the system and what harms could arise, so that risks can be weighed and mitigated. Competitor names, marketing budgets and department headcount do not describe impacts on people or society.

Question 34Artificial Intelligence

A model achieves 99% accuracy on a dataset where 99% of records belong to the majority class. Which conclusion is most defensible?

  1. The accuracy figure is misleading and class-sensitive metrics are needed Correct answer.
  2. The model reliably detects the minority class
  3. The model is overfitted to the minority class
  4. The dataset is balanced enough for accuracy to be sufficient

Explanation

When one class dominates, a trivial predictor that always outputs the majority class already reaches the majority share, so accuracy alone hides poor minority detection. Metrics such as recall, precision or balanced accuracy for the minority class are required to judge usefulness.

Question 35Artificial Intelligence

A team notices that a feature derived from the target variable is available at training time but would not be known at prediction time. What problem does including it cause?

  1. Underfitting due to too few features
  2. Reduced need for a test set
  3. Target leakage, producing optimistic validation results Correct answer.
  4. Slower inference from fewer parameters

Explanation

A feature that encodes information about the target but is unavailable at prediction time causes target leakage, inflating validation performance that will not hold in production. The correct remedy is to exclude such features or reconstruct them using only information available at inference.

Question 36Artificial Intelligence

Which metric set best captures performance for a retrieval-augmented generation system answering questions from a document corpus?

  1. Retrieval recall together with answer faithfulness to retrieved passages Correct answer.
  2. Training loss alone
  3. Model parameter count
  4. Number of indexed documents

Explanation

Such a system has two failure points: whether the right passages are retrieved and whether the generated answer is grounded in them. Retrieval recall plus faithfulness to the retrieved context measures both, whereas loss, index size and parameter count do not measure answer quality.

Question 37Artificial Intelligence

When fine-tuning a large language model on sensitive internal documents, which control most directly limits memorisation of those documents?

  1. Removing the validation split
  2. Applying deduplication and privacy filtering to the fine-tuning data Correct answer.
  3. Increasing the number of training epochs
  4. Using a larger batch size

Explanation

Memorisation risk rises when sensitive text appears repeatedly or verbatim in training data, so deduplication and privacy filtering reduce the chance the model reproduces it. More epochs and larger batches tend to increase fitting, and removing validation removes oversight.

Question 38Artificial Intelligence

A team must decide whether an AI feature requires a full risk assessment or a lighter review. Which factor most appropriately drives that decision?

  1. The programming language used to build the system
  2. The number of meetings held during development
  3. The colour scheme of the user interface
  4. The potential severity and scale of harm to affected people Correct answer.

Explanation

Risk-based proportionality ties the depth of assessment to the severity and scale of potential harm, which is the core logic of risk classification in AI governance frameworks. Implementation language, meeting counts and interface styling do not indicate risk to people.

Question 39Artificial Intelligence

Which approach best addresses a model that performs well overall but poorly for a specific user subgroup?

  1. Remove the subgroup from the dataset
  2. Report only the aggregate metric and proceed
  3. Evaluate disaggregated performance by subgroup and mitigate the identified gap Correct answer.
  4. Increase the overall dataset size without further analysis

Explanation

Aggregate metrics can mask subgroup failures, so performance must be disaggregated to reveal disparities before targeted mitigation such as rebalancing or group-aware evaluation. Hiding the gap or deleting the subgroup conceals the problem rather than solving it.

Question 40Artificial Intelligence

In an AI management system, what distinguishes a 'corrective action' from a preventive action?

  1. Corrective action is optional; preventive action is mandatory
  2. Corrective action is always automated; preventive action is always manual
  3. Corrective action applies only to hardware; preventive action applies only to software
  4. Corrective action addresses a nonconformity that has occurred; preventive action addresses a potential one Correct answer.

Explanation

Corrective action responds to a nonconformity or incident that has already happened, aiming to remove its cause and prevent recurrence. Preventive action targets a potential nonconformity before it occurs, so the two differ by whether the issue has materialised.

Question 41Artificial Intelligence

A production model's error rate rises sharply after a vendor changes the format of an upstream data feed. Which root-cause category does this most likely represent?

  1. Data drift caused by a change in input schema or pipeline Correct answer.
  2. Concept drift caused by a change in the input-to-output relationship
  3. Overfitting to the training set
  4. Insufficient model parameter count

Explanation

A schema or feed change alters the inputs the model receives, which is a data drift and pipeline-integrity problem rather than a change in the underlying relationship between inputs and labels. The fix lies in monitoring input contracts and validating data quality, not in retraining for concept drift alone.

Question 42Artificial Intelligence

Which practice best supports reproducibility of an experiment that produced a strong model result?

  1. Recording data version, code commit, hyperparameters and random seeds Correct answer.
  2. Relying on a developer's memory of the settings
  3. Sharing only the final accuracy figure
  4. Overwriting previous experiment logs

Explanation

Reproducibility requires capturing every input that affects the outcome, including data version, code revision, hyperparameters and seeds, so the run can be repeated and verified. Sharing only a metric or overwriting logs destroys the ability to reconstruct the result.

Question 43Artificial Intelligence

A team wants to reduce hallucination in a customer-facing LLM assistant. Which design choice most directly addresses the cause?

  1. Grounding responses in retrieved, cited source documents Correct answer.
  2. Increasing the maximum response length
  3. Removing the system prompt
  4. Raising the sampling temperature

Explanation

Hallucination often stems from the model generating unsupported content, so grounding answers in retrieved sources with citations constrains outputs to verifiable material. Higher temperature increases randomness, and removing instructions or lengthening outputs does not improve factual grounding.

Question 44Artificial Intelligence

An organisation wants assurance that a third-party AI component meets its risk requirements before integration. Which artefact is most relevant to request from the supplier?

  1. The supplier's social media calendar
  2. A conformity or assurance documentation package covering intended use, testing and limitations Correct answer.
  3. The personal CVs of the supplier's sales team
  4. The supplier's office floor plan

Explanation

Third-party assurance depends on documented evidence about intended use, evaluation results, known limitations and conformity with agreed requirements. Office layouts, sales CVs and social media plans provide no information about the component's risk profile or performance.

Question 45Artificial Intelligence

Which statement best explains why an AI system's intended use must be defined before conformity assessment?

  1. Requirements and risk thresholds depend on the context and purpose of use Correct answer.
  2. Conformity assessment is independent of how the system is used
  3. Intended use only affects the marketing slogan
  4. Intended use is determined solely by the model architecture

Explanation

Risk and applicable requirements are context-dependent, so the intended purpose determines which hazards matter and which controls are proportionate. Conformity assessment cannot be meaningful without a defined use context, and architecture alone does not define purpose.

Question 46Artificial Intelligence

A model card is being prepared for a released system. Which content most improves downstream users' ability to judge fitness for their context?

  1. Documented intended use, training data characteristics, evaluation results and known limitations Correct answer.
  2. A list of the development team's hobbies
  3. The source code licence alone
  4. The release date without further detail

Explanation

Model cards communicate the information downstream users need to judge whether a model suits their context, including purpose, data, measured performance and limitations. Hobbies, licence alone or a bare release date do not support fitness-for-purpose judgements.

Question 47Artificial Intelligence

An organisation wants to detect harmful outputs from a generative model in near real time. Which combination is most appropriate?

  1. Periodic annual review only
  2. Disabling logging of model outputs
  3. Relying solely on user complaints after publication
  4. Automated content classifiers plus sampled human review with escalation paths Correct answer.

Explanation

Near-real-time detection needs automated filters for scale, backed by sampled human review and escalation to catch classifier misses and handle edge cases. Annual review is too slow, and removing logs or waiting for complaints leaves harms undetected until after impact.

Question 48Artificial Intelligence

A credit-scoring model uses a proxy variable that is strongly correlated with a protected attribute. Why is removing only the protected attribute insufficient for fairness?

  1. Protected attributes are never present in training data
  2. Proxy variables can reconstruct the protected information and preserve disparate impact Correct answer.
  3. Fairness is guaranteed once the attribute is absent
  4. Removing any variable always improves model accuracy

Explanation

Correlated proxies can carry the same information as the protected attribute, so disparate outcomes may persist even after its removal. Fairness analysis therefore examines outcomes and proxy effects rather than relying on feature exclusion alone.

Question 49Artificial Intelligence

A red-team exercise finds that an LLM agent with tool access can be manipulated into exfiltrating data through a crafted document. Which control most directly addresses this class of risk?

  1. Least-privilege tool permissions with output validation and human approval for sensitive actions Correct answer.
  2. Increasing the model's context window
  3. Adding more few-shot examples of normal tasks
  4. Reducing the temperature to zero

Explanation

The risk arises from the agent's ability to act on untrusted content with broad tool permissions, so limiting privileges, validating outputs and requiring approval for sensitive actions constrains the impact. Context size, examples and temperature do not remove the agent's capability to exfiltrate data.

Question 50Artificial Intelligence

An organisation must demonstrate that an AI system remains compliant after deployment. Which monitoring design best supports that claim?

  1. Collecting logs without defined thresholds or owners
  2. Monitoring only server uptime
  3. One-off testing performed before release only
  4. Continuous logging of inputs, outputs, decisions and drift indicators mapped to defined thresholds and review triggers Correct answer.

Explanation

Post-deployment compliance requires ongoing evidence that the system stays within defined performance and risk bounds, which means monitoring relevant signals against thresholds and assigning review triggers and owners. Pre-release testing alone or unowned logs cannot demonstrate continuing conformity.