Knowledge + Practice

CompTIA AI+ AI0-001 (AI0-001) — Questions 376–450

754 questions total · 11pages · All types, answers revealed

Take a mock exam Exam hub

Page 6 of 11

376

MCQmedium

A data science team is building a model to detect fraudulent transactions. They have a dataset of 1 million normal transactions and 1,000 fraudulent ones. What is the MOST effective data preparation step to handle this imbalance?

A.Delete all normal transactions until the dataset is balanced

B.Duplicate the fraudulent transactions 1,000 times

C.Apply SMOTE to generate synthetic fraudulent transactions and randomly undersample normal transactions

D.Train the model on the original dataset; class imbalance does not affect model performance

AnswerC

SMOTE creates synthetic fraud samples, and undersampling reduces the majority class, creating a more balanced dataset.

Why this answer

Combining oversampling of the minority class (e.g., SMOTE) with undersampling of the majority class is a common and effective approach to balance the dataset.

Full explanation →

377

MCQmedium

A company is deploying a computer vision model to smartphones for offline object detection. The model was trained in PyTorch. Which format should they use for deployment on iOS devices?

A.TorchScript

B.ONNX

C.Core ML

D.TensorFlow Lite

AnswerC

Core ML is Apple's native format for iOS, providing optimized inference.

Why this answer

Core ML is Apple's framework for on-device machine learning on iOS, and PyTorch models can be converted to Core ML format.

Full explanation →

378

MCQmedium

A security team discovers that an AI-based anomaly detection system frequently misclassifies benign network traffic as malicious when the source IP is from a specific geographic region. Which type of AI vulnerability is most likely being exploited?

A.Data poisoning

B.Model inversion

C.Adversarial evasion

D.Membership inference

AnswerC

Adversarial evasion manipulates input features to cause misclassification. The regional bias suggests crafted inputs bypassing detection.

Why this answer

The scenario describes an AI-based anomaly detection system that misclassifies benign traffic from a specific geographic region as malicious. This is a classic example of an adversarial evasion attack, where an attacker crafts inputs (in this case, network traffic) that appear benign to human analysts but cause the AI model to misclassify them. The geographic bias suggests the attacker is exploiting the model's learned decision boundary, likely by manipulating features such as source IP or packet timing to evade detection.

Exam trap

The AI0-001 exam often tests the distinction between data poisoning (training-time attack) and adversarial evasion (inference-time attack), and the trap here is that candidates confuse the geographic bias with a poisoned training set rather than recognizing it as an evasion technique exploiting the model's learned regional patterns.

How to eliminate wrong answers

Option A is wrong because data poisoning involves injecting malicious data into the training set to corrupt the model's learning, not causing misclassification of benign traffic at inference time. Option B is wrong because model inversion attacks aim to reconstruct private training data from the model's outputs, not to cause misclassification of specific inputs. Option D is wrong because membership inference attacks determine whether a specific data point was used in training, not to cause the model to misclassify benign traffic as malicious.

Full explanation →

379

Multi-Selecteasy

A data scientist wants to develop a computer vision model using transfer learning. They need a framework that provides pre-trained models and easy-to-use APIs for data augmentation and training. Which TWO frameworks are best suited for this task?

Select 2 answers

A.Hugging Face Transformers

B.PyTorch

C.scikit-learn

D.TensorFlow

E.Keras

AnswersB, D

PyTorch provides torchvision with pre-trained models and torchvision.transforms for data augmentation, making it ideal for transfer learning in computer vision.

Why this answer

PyTorch (option B) is correct because it offers a rich ecosystem of pre-trained models via `torchvision.models`, along with built-in data augmentation transforms in `torchvision.transforms` and a flexible training loop that is ideal for transfer learning. Its dynamic computation graph makes it easy to modify model architectures for fine-tuning, which is a core requirement for the task.

Exam trap

Candidates often select Hugging Face Transformers because it provides pre-trained models, but it is primarily designed for NLP tasks, not computer vision. Similarly, Keras is a high-level API that runs on top of TensorFlow, so it is not considered a standalone framework for this purpose.

Full explanation →

380

MCQmedium

A social media platform uses an AI system to moderate content. The system incorrectly flags legitimate posts as hate speech, disproportionately affecting minority groups. Which type of bias is likely present?

A.Algorithmic bias

B.Historical bias

C.Selection bias

D.Confirmation bias

AnswerA

The model's algorithm or training process causes systematic errors against certain groups.

Why this answer

The AI system's output (incorrectly flagging legitimate posts as hate speech) is a direct result of the model's design, training data, or deployment choices, which is the definition of algorithmic bias. This bias disproportionately affects minority groups because the algorithm's decision-making process systematically produces unfair outcomes for those groups, even if the training data itself was not historically biased.

Exam trap

The AI0-001 exam often tests the distinction between 'algorithmic bias' (bias introduced by the model's design or deployment) and 'historical bias' (bias present in the training data), so candidates mistakenly choose historical bias when the question describes a system that actively produces unfair outcomes due to its own logic.

How to eliminate wrong answers

Option B (Historical bias) is wrong because historical bias refers to pre-existing societal prejudices reflected in the training data, not the algorithm's own flawed decision-making process; the question states the system 'incorrectly flags' posts, indicating the bias is in the algorithm's logic or thresholds, not just the data. Option C (Selection bias) is wrong because selection bias occurs when the training data is not representative of the real-world population (e.g., overrepresenting certain groups), but the problem here is the algorithm's misclassification of content, not a sampling issue. Option D (Confirmation bias) is wrong because confirmation bias is a human cognitive bias where people favor information that confirms their preexisting beliefs; it does not apply to an AI system's content moderation decisions.

Full explanation →

381

MCQhard

An organization deploys a machine learning model for credit scoring. An attacker submits carefully crafted loan applications that are slightly outside normal ranges but cause the model to approve high-risk loans. What type of attack is this?

A.Model extraction

B.Prompt injection

C.Adversarial example

D.Data poisoning

AnswerC

Adversarial examples are crafted to fool a model during inference by small perturbations.

Why this answer

This is an adversarial example attack, where the attacker crafts inputs with small, carefully chosen perturbations that cause the ML model to misclassify them. In credit scoring, submitting loan applications with values slightly outside normal ranges exploits the model's decision boundary to approve high-risk loans, a classic evasion technique.

Exam trap

CompTIA often tests the distinction between data poisoning (training-time attack) and adversarial examples (inference-time attack), so candidates mistakenly choose data poisoning when they see 'crafted inputs' without recognizing the attack occurs after deployment.

How to eliminate wrong answers

Option A is wrong because model extraction involves querying a model to steal its parameters or architecture, not manipulating inputs to cause misclassification. Option B is wrong because prompt injection targets large language models by injecting malicious instructions into prompts, not numerical input manipulation for tabular ML models. Option D is wrong because data poisons the training data to corrupt the model during training, whereas this attack occurs at inference time on a deployed model.

Full explanation →

382

MCQmedium

An organization is adopting a third-party pre-trained language model for internal use. To assess supply chain security, which document should they request to understand the components and dependencies of the model?

A.OWASP LLM Top 10

B.Model card

C.Data flow diagram

D.Software Bill of Materials (SBOM)

AnswerD

An SBOM lists all third-party components, libraries, and dependencies, enabling supply chain risk assessment.

Why this answer

A Software Bill of Materials (SBOM) is the correct document for assessing supply chain security because it provides a detailed, machine-readable inventory of all components, libraries, and dependencies used to build the model. This allows the organization to identify known vulnerabilities, licensing risks, and transitive dependencies, which is essential for evaluating the security posture of a third-party pre-trained model.

Exam trap

The AI0-001 exam often tests the distinction between a model card (which describes model behavior) and an SBOM (which describes software components), leading candidates to mistakenly choose the model card for supply chain security questions.

How to eliminate wrong answers

Option A is wrong because the OWASP LLM Top 10 is a list of common vulnerabilities and risks for Large Language Model applications, not a document that enumerates the specific components and dependencies of a given model. Option B is wrong because a model card documents the model's intended use, performance, and limitations, but it does not provide a detailed inventory of software components or dependencies needed for supply chain security assessment. Option C is wrong because a data flow diagram illustrates how data moves through a system, but it does not list the software libraries, packages, or third-party components that constitute the model's supply chain.

Full explanation →

383

MCQeasy

A startup has developed a natural language processing model for sentiment analysis. Their CI/CD pipeline includes a step that runs unit tests on the model's output format and a validation step that checks accuracy on a static test dataset. Recently, the pipeline often fails during the validation step, but the failures are inconsistent—sometimes the same model version passes, sometimes fails. The team suspects the test dataset is small and randomly sampled. They need a reliable validation process to deploy models with confidence. Which approach should the team implement?

A.Replace the static test set with k-fold cross-validation in each pipeline run

B.Increase the accuracy threshold to 95% so only very good models pass

C.Remove the validation step and rely on unit tests only

D.Fix the test dataset to be larger and more representative, and use a statistical test to compare against baseline

AnswerD

A fixed dataset and statistical test provide consistent and objective validation.

Why this answer

The core issue is that the static test dataset is too small and randomly sampled, leading to inconsistent validation results. By fixing the dataset to be larger and more representative, and using a statistical test (e.g., a paired t-test or McNemar's test) to compare the model's accuracy against a baseline, the team can reliably determine if performance changes are statistically significant, eliminating the randomness that causes pipeline failures to be inconsistent.

Exam trap

CompTIA often tests the misconception that increasing the accuracy threshold or using cross-validation alone can fix validation instability, when the real solution is to address the root cause of small, non-representative test data with statistical rigor.

How to eliminate wrong answers

Option A is wrong because k-fold cross-validation is computationally expensive and time-consuming for a CI/CD pipeline, and it does not directly address the root cause of a small, randomly sampled test set; it would still suffer from variance if the dataset is small. Option B is wrong because simply raising the accuracy threshold to 95% does not fix the underlying inconsistency from a small test set; it may cause even more frequent failures due to random sampling noise, and it does not provide a statistical basis for decision-making. Option C is wrong because removing the validation step entirely would allow models with poor accuracy to be deployed, undermining the goal of deploying with confidence; unit tests alone cannot assess model performance.

Full explanation →

384

Multi-Selectmedium

A team is developing a natural language processing model to classify customer feedback. The dataset contains text in multiple languages. Which THREE preprocessing steps are essential to ensure the model performs well across all languages?

Select 3 answers

A.One-hot encoding

B.Lowercasing

C.Tokenization

D.Stemming

E.Removing stop words

AnswersB, C, E

Lowercasing reduces vocabulary size and helps generalize across different cases.

Why this answer

Lowercasing is essential because it normalizes text across languages by converting all characters to the same case, reducing vocabulary size and ensuring that words like 'Good' and 'good' are treated identically. This prevents the model from learning separate representations for case variations, which is critical for multilingual datasets where case usage may differ (e.g., German capitalizes nouns). Without lowercasing, the model's performance degrades due to sparsity and increased feature space.

Exam trap

CompTIA often tests the distinction between preprocessing steps (like lowercasing, tokenization, stop word removal) and feature engineering techniques (like one-hot encoding), leading candidates to mistakenly include one-hot encoding as a preprocessing step when it is actually a vectorization method applied after preprocessing.

Full explanation →

385

MCQeasy

A data scientist trains a regression model and notices the training loss is low but validation loss is high. Which technique should be applied FIRST to address this issue?

A.Increase the learning rate.

B.Add more layers to the neural network.

C.Increase the size of the training dataset.

D.Apply L1 or L2 regularization to the model.

AnswerD

Regularization penalizes large weights, reducing overfitting.

Why this answer

The scenario describes overfitting, where the model memorizes the training data but fails to generalize to unseen data. Applying L1 or L2 regularization (Option D) is the correct first step because it adds a penalty to the loss function for large weights, discouraging complexity and reducing overfitting without requiring additional data or architectural changes.

Exam trap

CompTIA often tests the distinction between overfitting and underfitting, and the trap here is that candidates may incorrectly choose to increase dataset size (Option C) as the first action, when regularization is the more immediate and practical first step to address overfitting without requiring new data collection.

How to eliminate wrong answers

Option A is wrong because increasing the learning rate would make training more unstable and could cause the loss to diverge, worsening both training and validation performance. Option B is wrong because adding more layers increases model capacity, which exacerbates overfitting when the training loss is already low and validation loss is high. Option C is wrong because increasing the size of the training dataset can help reduce overfitting, but it is not the first technique to apply; regularization is a simpler, more immediate fix that does not require collecting new data.

Full explanation →

386

Multi-Selecteasy

A company wants to build a system that can generate new product images for an online catalog. Which TWO generative AI approaches are most suitable?

Select 2 answers

A.Diffusion models

B.Variational autoencoders (VAEs)

C.Generative Adversarial Networks (GANs)

D.BERT-based model

E.GPT-style language model

AnswersA, C

Diffusion models like DALL-E and Stable Diffusion produce high-quality images from noise.

Why this answer

Generative Adversarial Networks (GANs) are widely used for image generation, and diffusion models (like Stable Diffusion) have achieved state-of-the-art results in image synthesis. Variational autoencoders (VAEs) can generate images but often produce blurrier outputs. GPT is for text, and BERT is for understanding.

Full explanation →

387

MCQmedium

A healthcare startup deploys an AI model to predict patient readmission rates. An internal audit reveals that the model consistently underestimates readmission risk for non-native English speakers. According to AI ethics principles, what is the most appropriate course of action?

A.Add a confidence score disclaimer to model outputs

B.Reduce the sample size of non-native English speakers to balance the dataset

C.Continue using the model as is, since overall accuracy is acceptable

D.Retrain the model with a more representative dataset that includes diverse language backgrounds

AnswerD

Retraining with balanced data addresses the root cause of bias.

Why this answer

It directly addresses the root cause of the bias: the training data lacks sufficient representation from non-native English speakers, leading to systematic underestimation of readmission risk for that group. Retraining with a more representative dataset aligns with the AI ethics principle of fairness by ensuring the model learns patterns across all demographic groups equally, rather than masking the issue with disclaimers or manipulating sample sizes.

Exam trap

The AI0-001 exam often tests the misconception that adding a disclaimer or adjusting sample sizes post-hoc is sufficient to address bias, when in fact the ethical requirement is to fix the data or model at the training stage to ensure fairness.

How to eliminate wrong answers

Option A is wrong because adding a confidence score disclaimer does not fix the underlying algorithmic bias; it merely informs users of potential inaccuracy without correcting the model's systematic error. Option B is wrong because reducing the sample size of non-native English speakers would exacerbate the bias by further underrepresenting that group, violating the ethical principle of fairness and likely increasing model variance. Option C is wrong because continuing to use a model with known demographic bias, even if overall accuracy is acceptable, violates the AI ethics principle of non-maleficence and could lead to harmful disparities in patient care.

Full explanation →

388

MCQmedium

A healthcare AI system uses patient data to predict disease risk. To comply with privacy regulations, the organization wants to ensure that the model cannot reveal whether a specific patient's data was used in training. Which technique should they implement?

A.Differential privacy

B.Homomorphic encryption

C.Federated learning

D.Model validation

AnswerA

Differential privacy ensures that the inclusion or exclusion of any single record has a limited effect on the model, protecting against membership inference.

Why this answer

Differential privacy adds noise to the training process, making it difficult to determine if any individual was in the training set. This directly addresses membership inference attacks.

Full explanation →

389

MCQmedium

A team is deploying a sentiment analysis model for social media posts. The model currently performs well on English text but poorly on code-switched text (e.g., Spanglish). Which approach is MOST effective for improving performance on code-switched data without starting from scratch?

A.Use a larger base model without additional training

B.Apply data augmentation by translating all code-switched posts to English

C.Train a new model from scratch on a mix of English and code-switched data

D.Fine-tune the existing model on a corpus of code-switched text

AnswerD

Fine-tuning leverages pre-trained knowledge and adapts to the target domain with less data and compute.

Why this answer

Fine-tuning the existing model on a corpus of code-switched text adapts the model to the new language pattern efficiently.

Full explanation →

390

MCQmedium

An AI operations team notices that the accuracy of a deployed fraud detection model has been declining over the past month. Which action should the team take to address this issue proactively?

A.Retrain the model with the most recent data immediately.

B.Manually update the model weights weekly.

C.Replace the model with a rule-based system.

D.Set up automated retraining pipeline triggered by performance degradation thresholds.

AnswerD

This allows continuous monitoring and automated response to drift, keeping the model accurate.

Why this answer

It establishes an automated retraining pipeline triggered by performance degradation thresholds, which aligns with MLOps best practices for maintaining model accuracy in production. This proactive approach ensures the model is retrained when its performance drops below a predefined metric (e.g., AUC or F1 score), without requiring manual intervention. It addresses concept drift, which is a common cause of declining accuracy in deployed fraud detection models.

Exam trap

CompTIA often tests the misconception that retraining with the most recent data immediately is the best proactive action, when in fact automated threshold-based retraining is the correct MLOps practice to avoid overfitting and ensure controlled updates.

How to eliminate wrong answers

Option A is wrong because retraining with the most recent data immediately may introduce data leakage or overfit to recent noise, and it does not address the root cause of performance degradation (e.g., concept drift) in a controlled manner. Option B is wrong because manually updating model weights weekly is not a scalable or reliable practice; it introduces human error and does not leverage automated monitoring or drift detection. Option C is wrong because replacing a machine learning model with a rule-based system would likely reduce the model's ability to detect complex fraud patterns, and it ignores the potential to retrain or update the existing model.

Full explanation →

391

MCQmedium

A machine learning engineer is building a spam filter. The dataset contains 10,000 emails, of which 1,000 are spam. The engineer decides to use a Random Forest classifier. Which preprocessing step is most critical to ensure the model generalizes well to new, unseen emails?

A.Apply Principal Component Analysis (PCA) to reduce dimensionality

B.Normalize the numerical features to have zero mean and unit variance

C.Split the data into training and testing sets before any other preprocessing

D.Encode all features using one-hot encoding

AnswerC

Splitting first prevents data leakage and ensures realistic evaluation.

Why this answer

Splitting the data into training and testing sets before any other preprocessing prevents data leakage. If preprocessing like normalization or PCA is applied to the entire dataset first, the test set information influences the training process, leading to overly optimistic performance estimates and poor generalization to new, unseen emails.

Exam trap

CompTIA often tests the concept of data leakage by presenting preprocessing steps that seem harmless but actually incorporate test set information, tricking candidates into thinking scaling or dimensionality reduction is always necessary for tree-based models.

How to eliminate wrong answers

Option A is wrong because PCA is an unsupervised dimensionality reduction technique that, if applied before splitting, would leak information from the test set into the training set, and Random Forest is robust to high-dimensional sparse data, making PCA unnecessary for generalization. Option B is wrong because Random Forest is a tree-based ensemble method that is invariant to monotonic transformations and does not require feature scaling; normalizing before splitting would also risk data leakage if done on the full dataset. Option D is wrong because one-hot encoding is only relevant for categorical features, and applying it before splitting could introduce data leakage if the encoding uses levels present only in the test set; moreover, not all features in an email dataset are categorical, and Random Forest can handle label encoding without one-hot encoding.

Full explanation →

392

Multi-Selecteasy

A data scientist is training a customer churn prediction model using sensitive customer data. To comply with data privacy regulations, they want to minimize the risk of membership inference attacks. Which TWO techniques should they consider?

Select 2 answers

A.Use k-fold cross-validation to improve model accuracy

B.Deploy the model as a black-box API with no confidence scores

C.Use techniques to reduce overfitting, such as regularization or simpler models

D.Apply differential privacy during training

E.Increase training data size through data augmentation

AnswersC, D

Overfitted models are more susceptible to membership inference because they memorize training examples; reducing overfitting helps generalize and lowers inference risk.

Why this answer

Differential privacy and reducing model complexity (e.g., limiting overfitting) are effective against membership inference. Data augmentation and cross-validation do not directly reduce inference risk. Using a black-box API is about deployment, not training.

Full explanation →

393

MCQeasy

A model serving endpoint is tested using curl commands. Based on the exhibit, what is the most likely issue?

A.The server is returning HTTP 500 errors

B.The input features are malformed

C.The model is experiencing intermittent high latency leading to timeouts

D.The model is not deployed on the server

AnswerC

The third request timed out, suggesting occasional performance degradation.

Why this answer

The exhibit shows that the first curl request succeeds (HTTP 200), but subsequent requests fail with 'curl: (28) Operation timed out' after the default timeout of 30 seconds. This pattern of intermittent success followed by timeouts is characteristic of a model experiencing high latency spikes, not a persistent server error or configuration issue. The server is reachable and the model responds correctly some of the time, ruling out deployment or malformed input issues.

Exam trap

CompTIA often tests the distinction between persistent errors (like 500 or 404) and intermittent timeout failures, where candidates mistakenly attribute timeouts to server errors or input issues rather than recognizing the pattern of variable latency.

How to eliminate wrong answers

Option A is wrong because the exhibit shows HTTP 200 responses for successful requests, not HTTP 500 errors; a server returning 500 errors would consistently fail with a 5xx status code, not timeouts. Option B is wrong because the first request succeeds, proving the input features are correctly formatted and accepted by the model; malformed features would cause persistent failures across all requests. Option D is wrong because the successful first request confirms the model is deployed and serving predictions; an undeployed model would return a 404 or 503 error, not a timeout after a successful response.

Full explanation →

394

MCQhard

An ML team monitors a production model using a dashboard that shows daily performance metrics. Over the past month, the model's accuracy has dropped from 92% to 87%, while the data distribution of input features has remained stable according to statistical tests. Which type of model drift is most likely occurring?

A.Data drift (covariate shift)

B.Model decay

C.Overfitting

D.Concept drift

AnswerD

Concept drift changes the mapping from inputs to outputs, reducing accuracy.

Why this answer

Concept drift occurs when the relationship between input features and the target variable changes, even if the input data distribution remains stable. In this scenario, the model's accuracy declines from 92% to 87% while input feature distributions are unchanged, indicating that the underlying mapping from features to labels has shifted—a classic sign of concept drift.

Exam trap

CompTIA often tests the distinction between data drift and concept drift by presenting a scenario where input distributions are stable but model performance degrades, leading candidates to mistakenly choose data drift (covariate shift) because they focus on the input features rather than the label relationship.

How to eliminate wrong answers

Option A is wrong because data drift (covariate shift) refers to changes in the distribution of input features, which the question explicitly states has remained stable according to statistical tests. Option B is wrong because model decay is a general term for performance degradation over time, but it is not a specific type of drift; the question asks for the type of drift, and concept drift is the precise classification. Option C is wrong because overfitting is a training-time issue where a model fits noise in the training data, leading to poor generalization on new data; it does not explain a gradual performance drop in production while input distributions remain stable.

Full explanation →

395

MCQmedium

A machine learning team notices that their model's performance degrades when deployed to a new geographic region. The data distribution in the new region differs from the training data. Which concept best describes this issue?

A.Covariate shift

B.Data leakage

C.Underfitting

D.Overfitting

AnswerA

Covariate shift happens when the distribution of input features changes between training and deployment.

Why this answer

Covariate shift occurs when the distribution of the input features (covariates) changes between training and deployment, while the conditional relationship P(Y|X) remains the same. In this scenario, the model's performance degrades because the new geographic region has a different data distribution than the training data, which is the classic definition of covariate shift. This is a common issue in machine learning when models are deployed in environments not represented in the training set.

Exam trap

CompTIA often tests the distinction between covariate shift and overfitting, where candidates mistakenly think performance degradation on new data is always due to overfitting, but the key is that overfitting implies poor performance on the same distribution, not a different one.

How to eliminate wrong answers

Option B is wrong because data leakage refers to information from outside the training set (e.g., future data or target information) being used to train the model, which artificially inflates performance, not a distribution shift between training and deployment. Option C is wrong because underfitting occurs when a model is too simple to capture patterns in the training data, resulting in poor performance on both training and test sets, not specifically a degradation due to a change in data distribution. Option D is wrong because overfitting happens when a model learns noise or specific patterns in the training data too well, leading to poor generalization on unseen data from the same distribution, not a shift to a different distribution.

Full explanation →

396

MCQhard

A company is evaluating a vendor's AI system for hiring. The vendor claims the system is fair because it achieves demographic parity. However, the company discovers that the system has significantly different false positive rates across groups. Which fairness issue does this indicate?

A.The system violates individual fairness

B.The system suffers from selection bias

C.The system violates equalised odds

D.The system is not calibrated

AnswerC

Equalised odds demands equal false positive rates across groups; significant differences indicate unfairness.

Why this answer

Equalised odds requires that a model's false positive rates and true positive rates are equal across all demographic groups. Since the vendor's system has significantly different false positive rates across groups, it violates the equalised odds fairness criterion, even if demographic parity (equal selection rates) is satisfied. This is a core fairness metric in AI governance, as it ensures that errors are distributed equitably.

Exam trap

A common CompTIA AI exam trap is the distinction between demographic parity and equalised odds, as candidates may assume that equal selection rates (demographic parity) automatically guarantee fairness across all error types.

How to eliminate wrong answers

Option A is wrong because individual fairness focuses on treating similar individuals similarly, not on group-level error rates like false positives. Option B is wrong because selection bias refers to systematic errors in data collection or sampling that lead to unrepresentative training data, not to post-deployment disparities in model errors across groups. Option D is wrong because calibration measures whether predicted probabilities match actual outcomes within each group, which is a separate property from equalised odds; a model can be calibrated yet still have unequal false positive rates.

Full explanation →

397

Multi-Selecthard

Which TWO are valid techniques to reduce overfitting in a deep neural network? (Choose TWO.)

Select 2 answers

A.Increase batch size

B.Increase learning rate

C.L2 regularization

D.Gradient clipping

E.Dropout

AnswersC, E

L2 regularization adds a penalty for large weights, discouraging complex models.

Why this answer

L2 regularization (option C) is a valid technique to reduce overfitting by adding a penalty term proportional to the square of the weight magnitudes to the loss function. This discourages the network from learning overly complex patterns, effectively shrinking weights and improving generalization. Dropout (option E) randomly drops a fraction of neurons during training, which prevents co-adaptation of features and forces the network to learn more robust representations, also reducing overfitting.

Exam trap

CompTIA often tests the distinction between techniques that improve training stability (like gradient clipping or adjusting batch size/learning rate) versus those that directly regularize the model to reduce overfitting (like L2 regularization and dropout), leading candidates to confuse optimization tricks with regularization methods.

Full explanation →

398

MCQmedium

A company wants to build a conversational agent that can handle complex multi-step tasks such as booking a flight, reserving a hotel, and scheduling a car rental in a single session. The agent must be able to break down the user's request into sub-tasks, call external APIs, and reason about the results. Which design pattern is BEST suited for this requirement?

A.Retrieval-Augmented Generation (RAG) with a vector store

B.An agentic workflow implementing the ReAct pattern with tool use

C.A single large language model prompt with all instructions

D.Fine-tuning a model on a dataset of flight, hotel, and rental conversations

AnswerB

ReAct (Reasoning+Acting) agents iteratively decompose tasks, call APIs, and reason about results, perfectly suiting complex multi-step workflows.

Why this answer

Agentic workflows, particularly the ReAct pattern, combine reasoning and acting (tool calls) allowing the agent to iteratively decompose tasks, use APIs, and adapt based on results.

Full explanation →

399

Multi-Selecthard

A team is using k-fold cross-validation to evaluate a model. They observe high variance in performance scores across folds. Which TWO actions are most likely to reduce this variance? (Choose TWO.)

Select 2 answers

A.Increase the number of folds

B.Use stratified cross-validation

C.Decrease the number of folds

D.Shuffle data before splitting

E.Use a more complex model

AnswersA, B

More folds mean each training set is larger and more similar to the full dataset, reducing variance.

Why this answer

Increasing the number of folds (e.g., from 5 to 10) means each fold contains more training data, which reduces the variance of the performance estimate because the model is trained on a larger portion of the dataset each time. Stratified cross-validation ensures that each fold maintains the same class distribution as the original dataset, which stabilizes performance scores when the dataset is imbalanced, thereby reducing variance across folds.

Exam trap

The AI0-001 exam often tests the misconception that decreasing the number of folds reduces variance, when in fact the opposite is true—fewer folds increase variance because each training set is smaller and more dissimilar.

Full explanation →

400

MCQeasy

A team is building a recommendation system using collaborative filtering. They have a sparse user-item matrix. Which technique should they use to handle the sparsity and improve recommendations?

A.Association rule mining

B.Matrix factorization

C.k-nearest neighbors

D.Content-based filtering

AnswerB

Matrix factorization reduces dimensionality and captures latent features, effectively handling sparsity.

Why this answer

Matrix factorization (B) is the correct technique because it decomposes the sparse user-item matrix into lower-dimensional latent factor matrices, effectively capturing underlying patterns and filling in missing entries. This directly addresses sparsity by learning dense representations that generalize beyond observed interactions, which is a core strength in collaborative filtering for recommendation systems.

Exam trap

CompTIA often tests the misconception that k-nearest neighbors (k-NN) is the go-to for collaborative filtering, but candidates fail to recognize that k-NN's performance collapses under high sparsity, whereas matrix factorization explicitly models latent factors to overcome this.

How to eliminate wrong answers

Option A is wrong because association rule mining (e.g., Apriori algorithm) is designed for market basket analysis to find frequent itemsets and rules, not for handling sparse user-item matrices in collaborative filtering; it fails to generalize from sparse data and does not model latent factors. Option C is wrong because k-nearest neighbors (k-NN) is a memory-based collaborative filtering method that relies on direct similarity computations between users or items, which degrades severely with high sparsity due to lack of overlapping ratings, leading to poor recommendations. Option D is wrong because content-based filtering uses item features (e.g., genre, keywords) to recommend similar items, not the user-item interaction matrix; it does not address sparsity in collaborative filtering and ignores collaborative signals from other users.

Full explanation →

401

Multi-Selectmedium

Which TWO are best practices for deploying AI models in a containerized production environment? (Select TWO.)

Select 2 answers

A.Always pull the latest image tag for automatic updates

B.Store model artifacts inside the container image for portability

C.Use an orchestration platform like Kubernetes for scaling and health management

D.Package the model and its dependencies into a single container image

E.Configure JVM heap arguments inside the container if using Java

AnswersC, D

Kubernetes provides automated scaling and self-healing.

Why this answer

Orchestration platforms like Kubernetes provide automated scaling, self-healing, and rolling updates for containerized AI models. Kubernetes uses liveness and readiness probes to monitor model health and restart failed containers, ensuring high availability in production.

Exam trap

CompTIA often tests the distinction between containerization best practices (e.g., immutable images, external model storage) and generic software deployment habits (e.g., using latest tags, embedding data), so candidates mistakenly select options that seem convenient but violate production reliability principles.

Full explanation →

402

Multi-Selecteasy

Which TWO of the following are common activation functions used in deep neural networks?

Select 2 answers

A.Linear Regression

B.Support Vector Machine

C.K-means

D.ReLU

E.Sigmoid

AnswersD, E

ReLU is the most common activation for hidden layers.

Why this answer

ReLU (Rectified Linear Unit) is a common activation function in deep neural networks because it introduces non-linearity while being computationally efficient, outputting the input directly if positive and zero otherwise. It helps mitigate the vanishing gradient problem, making it a default choice for hidden layers in many architectures.

Exam trap

CompTIA AI often tests the distinction between machine learning algorithms (like Linear Regression, SVM, K-means) and neural network components (like activation functions), so candidates mistakenly select algorithms as activation functions because they recognize them as common ML terms.

Full explanation →

403

MCQeasy

A data engineer needs to process streaming clickstream data for real-time feature engineering in an ML pipeline. Which data pipeline technology is BEST suited for this task?

A.Apache Spark in batch mode

B.Snowflake

C.Apache Kafka

D.Apache Airflow

AnswerC

Kafka is purpose-built for real-time data streaming and can feed into ML pipelines.

Why this answer

Apache Kafka is the best choice because it is a distributed streaming platform designed for high-throughput, fault-tolerant, real-time data ingestion and processing. It can capture clickstream events as they occur and make them immediately available for feature engineering in an ML pipeline, supporting exactly-once semantics and low-latency delivery.

Exam trap

CompTIA AI often tests the distinction between data ingestion/messaging systems (Kafka) and batch processing or storage systems, leading candidates to confuse Airflow's orchestration role with actual stream processing capabilities.

How to eliminate wrong answers

Option A is wrong because Apache Spark in batch mode processes data in static, finite batches with high latency, making it unsuitable for real-time streaming clickstream data. Option B is wrong because Snowflake is a cloud-based data warehouse optimized for analytical queries on structured, stored data, not for real-time stream ingestion or processing. Option D is wrong because Apache Airflow is a workflow orchestration tool for scheduling and monitoring batch jobs, not a stream processing or messaging system capable of handling real-time data streams.

Full explanation →

404

Multi-Selecteasy

A company wants to use machine learning to recommend products to customers based on their purchase history. Which TWO techniques are appropriate for this task? (Select TWO)

Select 2 answers

A.Collaborative filtering

B.Principal Component Analysis (PCA)

C.K-Nearest Neighbors (KNN)

D.Naive Bayes

E.Linear regression

AnswersA, C

Collaborative filtering uses behavior patterns to recommend items.

Why this answer

Collaborative filtering recommends based on user similarities. K-Nearest Neighbors can find similar users or items. Both are suitable for recommendation.

Full explanation →

405

MCQeasy

A startup is building a chatbot for customer service. They have 500 recorded conversations and want to use a pre-trained language model to generate responses. However, they have limited computational resources and need the chatbot to respond in real-time. They are considering fine-tuning a large model like GPT-3 or using a smaller model like DistilBERT. The conversation data contains industry-specific jargon. Which approach should they take?

A.Use GPT-3 via API without fine-tuning

B.Fine-tune DistilBERT on the conversation data

C.Train a custom RNN from scratch on the conversations

D.Implement a rule-based system with keywords

AnswerB

DistilBERT is smaller, faster, and fine-tuning on domain-specific data will adapt it to jargon while meeting real-time requirements.

Why this answer

Fine-tuning DistilBERT on the 500 recorded conversations allows the model to adapt to industry-specific jargon while maintaining real-time responsiveness due to its smaller size. DistilBERT is a distilled version of BERT that retains 97% of BERT’s language understanding with 40% fewer parameters, making it suitable for limited computational resources. Fine-tuning on domain-specific data is essential here, as pre-trained models like GPT-3 lack exposure to the startup’s specialized terminology, and using a smaller model ensures low-latency inference for real-time chatbot responses.

Exam trap

CompTIA often tests the misconception that larger pre-trained models like GPT-3 are always superior for domain adaptation, ignoring the critical trade-offs of computational cost, latency, and the need for fine-tuning on small, specialized datasets.

How to eliminate wrong answers

Option A is wrong because using GPT-3 via API without fine-tuning would not adapt to the industry-specific jargon in the 500 conversations, leading to generic or incorrect responses, and the API call latency and cost are unsuitable for real-time constraints with limited resources. Option C is wrong because training a custom RNN from scratch on only 500 conversations is insufficient for learning complex language patterns, resulting in poor generalization and high risk of overfitting, while also requiring significant computational resources for training. Option D is wrong because a rule-based system with keywords cannot handle the variability and nuance of natural language in customer service conversations, especially with industry-specific jargon, and would fail to generate coherent, context-aware responses beyond predefined patterns.

Full explanation →

406

MCQeasy

A security analyst is reviewing logs from an AI-powered recommendation system and notices an unusually high number of requests for products from a specific vendor. The analyst suspects data poisoning. Which mitigation strategy should be implemented first?

A.Encrypt all training data at rest

B.Deploy an anomaly detection system on model outputs

C.Retrain the model with a smaller, curated dataset

D.Implement input validation and sanitization for training data

AnswerD

Input validation prevents poisoned data from entering the training pipeline.

Why this answer

Input validation and sanitization directly prevent malicious or anomalous data from entering the training pipeline, which is the root cause of data poisoning. In an AI-powered recommendation system, poisoned training data can cause the model to learn biased associations, such as favoring a specific vendor. By validating and sanitizing inputs before they are used for training, the attack vector is blocked at the earliest stage, making it the most effective first mitigation step.

Exam trap

The AI0-001 exam often tests the principle of defense in depth by making candidates choose a reactive or recovery measure (like retraining or monitoring outputs) instead of the proactive control that stops the attack at the input stage.

How to eliminate wrong answers

Option A is wrong because encrypting training data at rest protects confidentiality and integrity during storage, but it does not prevent malicious data from being ingested into the training set; data poisoning occurs before encryption is applied. Option B is wrong because deploying an anomaly detection system on model outputs is a reactive measure that detects poisoning after the model has already been compromised, rather than preventing the attack. Option C is wrong because retraining with a smaller, curated dataset may reduce the impact of poisoning but does not address the underlying vulnerability that allowed poisoned data to enter the system; it is a recovery step, not a first-line mitigation.

Full explanation →

407

MCQeasy

A data analyst wants to predict housing prices based on square footage, number of bedrooms, and location. Which machine learning approach is most suitable?

A.K-means clustering

B.Decision tree regression

C.Association rule mining

D.Linear regression

AnswerD

Linear regression models the linear relationship between input features and a continuous output.

Why this answer

Linear regression is the most suitable approach because the problem involves predicting a continuous numeric target (housing prices) from multiple independent variables (square footage, bedrooms, location). Linear regression models the linear relationship between the features and the target, providing interpretable coefficients and efficient training for this type of regression task.

Exam trap

The trap here is that candidates may confuse regression (predicting a continuous value) with classification or unsupervised learning, and incorrectly select decision tree regression or clustering because they see 'prediction' and assume any tree-based or grouping method works.

How to eliminate wrong answers

Option A is wrong because K-means clustering is an unsupervised learning algorithm used for grouping unlabeled data into clusters, not for predicting a continuous target variable. Option B is wrong because decision tree regression can be used for regression, but it is not the most suitable here; it tends to overfit and lacks the interpretability and simplicity of linear regression for a straightforward linear relationship. Option C is wrong because association rule mining is an unsupervised technique for discovering frequent itemsets and rules in transactional data, not for predicting numeric values.

Full explanation →

408

Multi-Selecteasy

Which TWO are common types of adversarial attacks on AI models?

Select 2 answers

A.Hyperparameter tuning

B.Transfer learning

C.Evasion attack

D.Backdoor attack

E.Data poisoning

AnswersC, E

Evasion attacks craft input perturbations to cause misclassification at test time.

Why this answer

Evasion attacks (Option C) are a common type of adversarial attack where an attacker crafts malicious input data that is intentionally designed to cause a trained AI model to make incorrect predictions or classifications, often by adding imperceptible perturbations to legitimate inputs. This exploits the model's sensitivity to small changes in feature space, leading to misclassification without altering the model itself.

Exam trap

The AI0-001 exam often tests the distinction between attack types that occur during training (data poisoning) versus inference (evasion), and candidates may mistakenly classify hyperparameter tuning or transfer learning as attacks because they sound like active manipulations, but they are standard ML practices.

Full explanation →

409

MCQeasy

Which metric is most appropriate for evaluating a binary classification model where the positive class is rare and false positives are costly?

A.Accuracy

B.F1-score

C.Precision

D.Recall

AnswerC

Correct; precision measures how many predicted positives are actually positive, reducing false positives.

Why this answer

Precision is the most appropriate metric when the positive class is rare and false positives are costly because it measures the proportion of true positive predictions among all positive predictions. In this scenario, minimizing false positives is critical, and precision directly penalizes them by requiring high confidence before labeling an instance as positive. This aligns with the business need to avoid costly false alarms, such as in fraud detection or medical diagnosis for rare diseases.

Exam trap

CompTIA often tests the misconception that accuracy is always the best metric, but the trap here is that candidates overlook how class imbalance and asymmetric costs make precision or recall more relevant, and they fail to distinguish between F1-score and precision when the cost of false positives is explicitly stated.

How to eliminate wrong answers

Option A is wrong because accuracy is misleading for imbalanced datasets; a model that predicts the majority class for all instances can achieve high accuracy while failing to identify any positive cases, which is useless when the positive class is rare. Option B is wrong because F1-score balances precision and recall, but when false positives are costly, precision alone is more appropriate; F1-score would still allow some false positives in favor of recall, which is undesirable here. Option D is wrong because recall focuses on capturing all positive instances, but it does not penalize false positives; in a rare positive class scenario with high cost of false positives, maximizing recall would likely increase false positives, which is counterproductive.

Full explanation →

410

MCQhard

A deep learning model for image classification is overfitting the training data. The team has already tried data augmentation and dropout. Which additional technique should they implement to reduce overfitting?

A.Batch normalization

B.Increase number of epochs

C.Gradient clipping

D.Early stopping

AnswerD

Early stopping monitors validation loss and stops training when it starts to increase, reducing overfitting.

Why this answer

Early stopping (Option D) is the correct additional technique because it halts training when validation performance stops improving, directly preventing the model from memorizing noise in the training data. Since data augmentation and dropout are already in use, early stopping provides a complementary regularization effect by limiting the number of training iterations before overfitting occurs.

Exam trap

CompTIA often tests the distinction between techniques that address overfitting versus those that solve optimization issues, leading candidates to confuse batch normalization or gradient clipping as overfitting solutions when they are not.

How to eliminate wrong answers

Option A is wrong because batch normalization primarily accelerates training and stabilizes learning by normalizing layer inputs, but it does not directly reduce overfitting—it can even have a slight regularizing effect, but it is not a primary overfitting countermeasure. Option B is wrong because increasing the number of epochs would exacerbate overfitting by giving the model more opportunities to memorize training data, making the problem worse. Option C is wrong because gradient clipping is used to prevent exploding gradients in deep networks, especially in RNNs, and does not address overfitting from excessive model capacity or insufficient regularization.

Full explanation →

411

MCQmedium

An AI engineer is tuning a deep learning model and observes that the training loss decreases very slowly. The learning rate is set to 0.001. Which adjustment is most likely to speed up convergence?

A.Increase the learning rate to 0.01

B.Add more hidden layers

C.Decrease the learning rate to 0.0001

D.Increase the batch size

AnswerA

A higher learning rate allows larger weight updates, potentially speeding up convergence.

Why this answer

A learning rate of 0.001 is causing the model to take very small steps toward the minimum of the loss function, resulting in slow convergence. Increasing the learning rate to 0.01 allows larger weight updates per iteration, which typically speeds up training. However, care must be taken not to overshoot the optimum, as an excessively high learning rate can cause divergence.

Exam trap

CompTIA often tests the misconception that decreasing the learning rate always improves training, when in fact a learning rate that is too low is a primary cause of slow convergence, and the correct adjustment is to increase it within a safe range.

How to eliminate wrong answers

Option B is wrong because adding more hidden layers increases model complexity and the number of parameters, which generally slows training and can exacerbate the slow convergence problem rather than solving it. Option C is wrong because decreasing the learning rate to 0.0001 would make the updates even smaller, further slowing convergence. Option D is wrong because increasing the batch size provides a more accurate gradient estimate but reduces the frequency of updates per epoch, which can actually slow convergence in terms of steps needed to reach a given loss.

Full explanation →

412

MCQeasy

A data scientist is building a binary classification model to predict customer churn. The dataset has 90% non-churn and 10% churn. After training, the model achieves 90% accuracy, but the recall for the churn class is only 20%. Which metric should the team primarily focus on to evaluate the model's effectiveness?

A.Recall for the churn class

B.Accuracy

C.Area Under the ROC Curve (AUC-ROC)

D.Precision for the non-churn class

AnswerA

Recall measures how many actual churners are correctly identified, which is the key concern.

Why this answer

When classes are imbalanced, accuracy is misleading. Recall (or F1) for the minority class is more informative.

Full explanation →

413

Multi-Selectmedium

A cybersecurity team is red-teaming their internal LLM-powered code assistant. They want to test the model's resistance to jailbreaking techniques that bypass safety guardrails. Which TWO of the following should they include in their red teaming exercise to effectively evaluate jailbreak resilience?

Select 2 answers

A.Model inversion to reconstruct training data

B.Role-playing scenarios where the model is asked to act as a character with no restrictions (e.g., DAN)

C.Encoding obfuscation, such as base64 encoding malicious instructions

D.Payload splitting across multiple user messages

E.Few-shot prompting with benign examples

AnswersB, C

Role-playing scenarios are a classic jailbreak technique that attempts to override system instructions by assigning the model an unrestricted persona.

Why this answer

Role-playing scenarios (e.g., DAN) and encoding obfuscation (e.g., base64) are common jailbreak techniques. Payload splitting is a type of prompt injection, not specifically jailbreaking. Few-shot prompting is a legitimate technique.

Model inversion is a privacy attack.

Full explanation →

414

MCQhard

A research team is training a deep learning model for image classification using a small dataset of 1,000 labeled images. They are concerned about overfitting. Which combination of regularisation techniques would be MOST effective?

A.Use early stopping without any other regularisation

B.Dropout with a rate of 0.5 and L2 regularisation

C.L1 regularisation and batch normalisation

D.Increase learning rate and use momentum

AnswerB

Dropout and L2 regularisation together effectively reduce overfitting by preventing reliance on specific neurons and penalising large weights.

Why this answer

Dropout randomly disables neurons during training to prevent co-adaptation, and L2 regularisation penalises large weights. Both are standard regularisation techniques. L1 promotes sparsity but is less common for dense layers.

Batch normalisation helps convergence but is not primarily a regularisation method.

Full explanation →

415

MCQmedium

An AI system experiences degraded accuracy over time due to changes in user behavior. Which monitoring metric should be prioritized to detect this issue earliest?

A.API response latency

B.Data drift detection on input features

C.Area under the ROC curve (AUC)

D.Model accuracy on a holdout validation set

AnswerB

Data drift detects changes before performance degrades.

Why this answer

Data drift detection monitors changes in input distribution, which often precedes accuracy drop. Option A is wrong because accuracy is a lagging indicator. Option C is wrong because latency doesn't reflect data shift.

Option D is wrong because AUC is also lagging.

Full explanation →

416

Multi-Selecteasy

A data scientist is monitoring a deployed image classification model. Which TWO actions are best practices for detecting model drift? (Choose 2.)

Select 2 answers

A.Schedule automatic weekly retraining of the model.

B.Increase the model's complexity to improve generalization.

C.Use a holdout test set to periodically evaluate model accuracy.

D.Monitor the average prediction confidence of the model.

E.Track the distribution of input data over time.

AnswersC, E

Comparing performance on a static test set reveals concept drift.

Why this answer

Periodically evaluating the model on a holdout test set that reflects the current production data distribution is a direct method to detect accuracy degradation caused by model drift. This approach measures whether the model's performance on unseen data has declined over time, which is a key indicator of drift.

Exam trap

CompTIA often tests the distinction between detection and remediation actions, so candidates mistakenly choose retraining (Option A) as a detection method when it is actually a corrective action.

Full explanation →

417

MCQhard

Refer to the exhibit. An AI governance review finds that a model was deployed without required ethics approval. Based on the audit log, who is most responsible for the compliance failure?

A.Bob

B.Alice

C.Carol

D.System

AnswerA

Bob deployed the model without ethics approval.

Why this answer

Bob is the data scientist who deployed the model to production. The audit log shows that Bob executed the deployment command without first obtaining the required ethics approval. As the individual who performed the action that violated the governance policy, Bob bears primary responsibility for the compliance failure.

Exam trap

The AI0-001 exam often tests the distinction between who performed the action versus who requested or approved it, leading candidates to incorrectly blame the project manager or ethics officer instead of the deployer.

How to eliminate wrong answers

Option B (Alice) is wrong because Alice is the project manager who requested the model deployment, but she did not perform the actual deployment action; the audit log shows she only submitted the request. Option C (Carol) is wrong because Carol is the ethics officer who approved the model earlier, but the audit log indicates she did not approve this specific deployment; the failure is that Bob bypassed the required approval step. Option D (System) is wrong because the system is an automated deployment pipeline that executed Bob's command; it has no agency or responsibility for compliance decisions, and the governance policy assigns accountability to human actors.

Full explanation →

418

Multi-Selecthard

A security engineer is hardening an LLM application against indirect prompt injection attacks. Which TWO controls are MOST effective? (Select two.)

Select 2 answers

A.Output filtering

B.Input validation and sanitization

C.Rate limiting

D.Differential privacy

E.Federated learning

AnswersA, B

Filtering outputs can block actions that arise from injected instructions.

Why this answer

Input validation and sanitization can filter malicious content in retrieved data, and output filtering can prevent the model from executing injected instructions. Both are key defenses.

Full explanation →

419

MCQhard

A large hospital system deploys an AI triage system for emergency rooms. The system uses patient vitals and symptoms to recommend treatment priority. Six months after deployment, complaints arise that the system frequently underestimates the severity of symptoms for patients from certain ethnic backgrounds. A data scientist runs a bias audit and finds that the model's false negative rate is 20% higher for the minority group. The hospital's AI governance board requires immediate corrective action. The data science team has limited resources and cannot retrain the entire model from scratch. They have access to the training data, which is imbalanced. The model is a gradient boosted tree. Which course of action best addresses the bias while minimizing operational impact?

A.Rebalance the training data using SMOTE and retrain the model

B.Use adversarial debiasing during training to remove protected attribute correlations

C.Post-process the model's predictions by adjusting thresholds for the minority group

D.Replace the model with a simpler logistic regression model to improve interpretability

AnswerC

Threshold adjustment is fast, cheap, and directly minimizes false negative disparity.

Why this answer

Post-processing by adjusting decision thresholds for the minority group directly compensates for the higher false negative rate without requiring retraining. Since the team has limited resources and cannot retrain the entire gradient boosted tree model, this approach minimizes operational impact while addressing the bias. The threshold adjustment effectively lowers the probability cutoff for the minority group, making the model more sensitive to their symptoms and reducing underestimation of severity.

Exam trap

CompTIA often tests the misconception that bias mitigation always requires retraining or complex algorithmic changes, when in fact post-processing threshold adjustments can be a quick, effective fix for deployed models with limited resources.

How to eliminate wrong answers

Option A is wrong because SMOTE rebalances the training data by oversampling the minority class, but retraining the entire gradient boosted tree model from scratch is resource-intensive and contradicts the constraint of limited resources; moreover, SMOTE may introduce synthetic noise that degrades model performance. Option B is wrong because adversarial debiasing is a training-time technique that requires modifying the model architecture and retraining, which is not feasible given the limited resources and the fact that the model is already deployed; it also does not directly address the false negative rate disparity without full retraining. Option D is wrong because replacing the model with a simpler logistic regression model would require retraining and likely reduce predictive performance, especially for complex interactions in patient vitals and symptoms, and does not guarantee bias reduction; interpretability alone does not correct the existing bias.

Full explanation →

420

Multi-Selectmedium

A team is implementing a RAG system. They are designing the document loading and chunking strategy. Which TWO techniques are commonly used for chunking documents? (Select two.)

Select 2 answers

A.Fixed-size chunking with a token limit

B.Frequency-based chunking by term occurrence

C.Character-level chunking with no overlap

D.Semantic chunking using sentence boundaries

E.Hierarchical chunking using document structure

AnswersA, D

Why this answer

Fixed-size chunking (based on token count) and semantic chunking (based on natural boundaries) are both standard approaches. Hierarchical chunking is less common, and character-level is rarely used. Overlap is a parameter, not a chunking strategy.

Full explanation →

421

Multi-Selecteasy

A data scientist is tuning a deep learning model. Which TWO hyperparameters directly affect the model's capacity to overfit?

Select 2 answers

A.Number of layers in the network.

B.Batch size.

C.Optimizer choice (e.g., SGD vs Adam).

D.Dropout rate.

E.Learning rate.

AnswersA, D

More layers increase capacity, raising overfitting risk.

Why this answer

Increasing the number of layers increases the model's depth, which expands its representational capacity and allows it to learn more complex patterns, including noise, thereby directly increasing overfitting risk. Option D is correct because dropout is a regularization technique that randomly drops neurons during training; a low dropout rate (e.g., 0.0) removes this regularization, while a high rate (e.g., 0.5) reduces overfitting by preventing co-adaptation of neurons.

Exam trap

CompTIA often tests the distinction between hyperparameters that affect model capacity (number of layers, dropout rate) versus those that affect training dynamics (batch size, optimizer, learning rate), leading candidates to mistakenly select learning rate or batch size as direct overfitting controls.

Full explanation →

422

MCQhard

A company deploys an AI model for loan approval. The model shows bias against a protected group. The team decides to use adversarial debiasing. What is the PRIMARY advantage of this approach?

A.It guarantees the model's predictions are private.

B.It reduces bias while preserving predictive performance by learning representations that are invariant to sensitive attributes.

C.It is simpler to implement than pre-processing techniques.

D.It ensures equal approval rates across all groups.

AnswerB

This is the core benefit of adversarial debiasing.

Why this answer

Adversarial debiasing is an in-processing technique that trains a primary model to predict the target (e.g., loan approval) while simultaneously training an adversary to predict the sensitive attribute from the model's learned representations. The primary model is penalized when the adversary succeeds, forcing it to learn representations that are invariant to the sensitive attribute. This reduces bias while preserving predictive performance because the model retains the ability to learn task-relevant patterns that are not correlated with the protected attribute.

Exam trap

The trap here is that candidates confuse 'reducing bias' with 'ensuring equal outcomes' (demographic parity), but adversarial debiasing targets equalized odds or equal opportunity by focusing on representation invariance, not strict rate equality.

How to eliminate wrong answers

Option A is wrong because adversarial debiasing does not guarantee privacy; it addresses fairness, not confidentiality, and does not provide differential privacy or encryption. Option C is wrong because adversarial debiasing is an in-processing technique that is generally more complex to implement than pre-processing techniques like reweighing or sampling, which modify the dataset before training. Option D is wrong because adversarial debiasing aims to reduce bias by learning invariant representations, but it does not enforce equal approval rates across groups; equal approval rates would be demographic parity, which is a different fairness metric and may not align with the model's predictive performance.

Full explanation →

423

MCQmedium

A financial services company has a real-time fraud detection system that uses Apache Kafka to stream transaction events, a TensorFlow Serving model for scoring, and a Redis cache for lookup of historical fraud patterns. The system processes 10,000 transactions per second with an SLA of 100ms latency per transaction. Recently, after a model update, the latency for some transactions spiked to over 500ms, causing timeouts. The model uses a deep neural network with 10 million parameters. The engineering team suspects the issue is due to increased model inference time. Which action should be taken to reduce latency without significant loss in accuracy?

A.Add more Redis nodes to the cache cluster

B.Increase the number of Kafka partitions and consumer threads

C.Decrease the inference batch size from 32 to 1

D.Quantize the model weights from FP32 to FP16

AnswerD

FP16 quantization reduces model size and speeds up inference, typically with minimal accuracy impact.

Why this answer

The latency spike is caused by increased model inference time after a model update. Quantizing model weights from FP32 to FP16 reduces memory bandwidth and computation requirements, directly speeding up inference on compatible hardware (e.g., GPUs with Tensor Cores) with minimal accuracy loss. This addresses the root cause—model inference latency—without changing the system architecture.

Exam trap

The trap here is that candidates confuse system-level scaling (adding cache nodes or Kafka partitions) with model-level optimization, failing to recognize that the latency spike originates from the model inference step itself.

How to eliminate wrong answers

Option A is wrong because adding Redis nodes improves cache lookup throughput, but the latency spike is due to model inference time, not cache performance. Option B is wrong because increasing Kafka partitions and consumer threads improves message ingestion parallelism, but does not reduce the per-transaction inference latency of the TensorFlow Serving model. Option C is wrong because decreasing the inference batch size from 32 to 1 reduces throughput and increases per-transaction overhead (e.g., kernel launch latency), which would worsen latency, not improve it.

Full explanation →

424

MCQhard

An AI model is deployed to a mobile app with limited computational resources. The model is a deep neural network with high latency. Which technique is best to reduce inference time?

A.Increase batch size

B.Add more layers

C.Use a larger model

D.Quantization

AnswerD

Quantization reduces model size and speeds up inference by using lower-precision arithmetic.

Why this answer

Quantization reduces the precision of the model's weights and activations (e.g., from 32-bit floating point to 8-bit integer), which decreases memory footprint and speeds up computation on resource-constrained devices like mobile phones. This directly lowers inference latency without requiring additional hardware or architectural changes.

Exam trap

The AI0-001 exam often tests the misconception that increasing batch size or model size improves performance on edge devices, when in fact these techniques increase resource demands and latency in low-resource environments.

How to eliminate wrong answers

Option A is wrong because increasing batch size improves throughput (samples per second) but does not reduce per-sample latency; it actually increases memory usage and can worsen latency on mobile devices with limited resources. Option B is wrong because adding more layers increases the model depth, which increases computational complexity and latency, making inference slower. Option C is wrong because using a larger model (more parameters) increases both memory and compute requirements, directly increasing inference time on constrained devices.

Full explanation →

425

MCQmedium

An AI model's performance drops significantly in production compared to testing. The data shows distribution shift. What is the best first step?

A.Add more features

B.Retrain model with new data

C.Use a different algorithm

D.Reduce model complexity

AnswerB

Retraining with current data addresses drift.

Why this answer

(Retrain model with new data) is correct because retraining with more representative data adapts to distribution shift. Option A (Add more features) may not address the shift. Option C (Change algorithm) is a larger change without addressing data.

Option D (Reduce model complexity) might worsen performance.

Full explanation →

426

MCQhard

A company is training a large language model from scratch and wants to minimise its environmental impact. Which practice aligns with green AI principles?

A.Use model pruning and train on a smaller, representative dataset

B.Use more GPUs to parallelise training and reduce wall-clock time

C.Deploy the model on a cloud provider with renewable energy certificates

D.Train the model on a larger dataset to improve accuracy

AnswerA

Pruning reduces model size and computational cost; training on a smaller dataset also lowers energy consumption, aligning with green AI.

Why this answer

Green AI advocates for resource-efficient AI, including using smaller models, pruning, and efficient architectures to reduce carbon footprint. Training larger models with more data increases environmental impact, not reduces it.

Full explanation →

427

Multi-Selectmedium

A data scientist is using differential privacy to protect individual privacy in a training dataset. Which TWO actions are correct implementations of differential privacy?

Select 2 answers

A.Train the model on a small subset of data to reduce exposure

B.Remove all personally identifiable information (PII) from the dataset

C.Aggregate data into groups before training

D.Set a privacy budget (epsilon) to limit information leakage

E.Add noise to the training data to mask individual contributions

AnswersD, E

The privacy budget epsilon quantifies the privacy guarantee and is a core concept of differential privacy.

Why this answer

Setting a privacy budget (epsilon) is a core mechanism in differential privacy that quantifies and limits the amount of information leaked about any individual in the dataset. By controlling epsilon, the data scientist can formally bound the privacy loss, ensuring that the model's outputs do not reveal whether any specific individual's data was included in training.

Exam trap

A common pitfall in this question is thinking that removing PII or aggregating data is sufficient for differential privacy. In reality, differential privacy requires a formal mathematical framework with noise addition and a privacy budget parameter. CompTIA often tests this distinction.

Full explanation →

428

MCQhard

Based on the exhibit, what is the most likely cause of the accuracy drop?

A.A required feature is missing from the production data pipeline.

B.Data drift in the 'income' feature has caused the model to become less accurate.

C.The model was overfitted to the training data.

D.The model's confidence threshold needs to be adjusted.

AnswerB

The detected distribution shift for 'income' indicates data drift, a common cause of performance degradation.

Why this answer

The exhibit shows a sudden and sustained drop in model accuracy coinciding with a shift in the distribution of the 'income' feature. This is a classic symptom of data drift, where the statistical properties of the input feature change over time, causing the model's learned patterns to no longer match the production data. Option B correctly identifies this as the most likely cause because the model was trained on a prior income distribution and is now encountering values outside that range.

Exam trap

CompTIA often tests the distinction between data drift and model overfitting by presenting a sudden accuracy drop after stable performance, leading candidates to incorrectly attribute it to overfitting when the exhibit clearly shows a distribution shift in a specific feature.

How to eliminate wrong answers

Option A is wrong because a missing feature in the production data pipeline would typically cause a pipeline failure or missing-value error, not a gradual accuracy drop that correlates with a specific feature's distribution shift. Option C is wrong because overfitting would manifest as high training accuracy with poor generalization from the start, not a sudden accuracy drop after a period of stable performance; the exhibit shows a clear change point, not a consistently low accuracy. Option D is wrong because adjusting the confidence threshold changes the precision-recall trade-off but does not address the underlying cause of the model's predictions becoming less reliable due to shifted input distributions; it would not restore the original accuracy level.

Full explanation →

429

MCQmedium

A healthcare startup is developing a deep learning model to detect diabetic retinopathy from retinal images. The model is trained on a dataset of 10,000 labeled images. During initial testing, the model achieves 99% accuracy on the training set but only 85% on the test set. The startup wants to deploy the model in a clinical setting where false negatives (missing a disease) are critical. The team has access to additional unlabeled retinal images from multiple sources. Which strategy should the team use to improve the model's generalization and reduce false negatives?

A.Use semi-supervised learning with the unlabeled images to improve feature representations

B.Apply aggressive data augmentation to the training set

C.Increase the learning rate during training

D.Add more convolutional layers to the model

AnswerA

Semi-supervised learning utilizes unlabeled data to learn generalizable features, reducing overfitting and improving test performance.

Why this answer

Semi-supervised learning leverages the large pool of unlabeled retinal images to learn robust feature representations, which helps the model generalize better to unseen data. By reducing overfitting (the gap between 99% training and 85% test accuracy), this approach directly improves test-set performance. Additionally, semi-supervised methods can be tuned to emphasize recall, thereby reducing false negatives critical in clinical diabetic retinopathy screening.

Exam trap

CompTIA often tests the misconception that simply increasing data or model complexity (augmentation, layers) always improves generalization, when in fact semi-supervised learning is the targeted solution for leveraging unlabeled data to close the train-test accuracy gap and address class-specific metrics like false negatives.

How to eliminate wrong answers

Option B is wrong because aggressive data augmentation, while helpful for generalization, does not directly address the high false-negative rate; it may even distort critical pathological features if applied too aggressively. Option C is wrong because increasing the learning rate typically destabilizes training, leading to divergence or poor convergence, and does not reduce false negatives or improve generalization. Option D is wrong because adding more convolutional layers increases model capacity, which would likely worsen overfitting given the already large gap between training and test accuracy, and does not specifically target false negatives.

Full explanation →

430

MCQhard

A company uses the above policy to control AI model access. A data scientist tries to run inference with model "llama-3-70b" at 150 requests in 30 minutes. What will happen?

A.All requests are allowed because the model is in the allowed list

B.All requests are denied because the rate limit is per minute and 150 exceeds the limit

C.The first 100 requests are allowed; the remaining 50 are denied

D.All requests are denied because the second rule blocks all models

AnswerC

The rate limit allows 100 requests per hour; exceeding requests are denied.

Why this answer

The policy allows up to 100 requests per 30 minutes for models in the allowed list, and 'llama-3-70b' is in that list. The rate limit is applied per 30-minute window, not per minute, so the first 100 requests are allowed, and the remaining 50 exceed the limit and are denied.

Exam trap

The trap here is that candidates often misinterpret the rate limit as a per-minute value (like 100 per minute) rather than the stated 100 per 30 minutes, leading them to incorrectly select Option B.

How to eliminate wrong answers

Option A is wrong because it ignores the rate limit; being in the allowed list does not bypass the 100 requests per 30-minute cap. Option B is wrong because it misinterprets the rate limit as per minute, but the policy specifies a per-30-minute window, so 150 requests in 30 minutes does not exceed a per-minute limit. Option D is wrong because the second rule does not block all models; it only blocks models not in the allowed list, and 'llama-3-70b' is explicitly allowed.

Full explanation →

431

Multi-Selecteasy

A data scientist is preparing a dataset for a classification model. The dataset contains several categorical variables with high cardinality. Which TWO encoding methods are appropriate for converting these categorical variables into numerical features?

Select 2 answers

A.Min-max scaling

B.K-means clustering

C.One-hot encoding

D.Principal component analysis (PCA)

E.Label encoding

AnswersC, E

One-hot encoding converts each category into a binary vector, suitable for categorical variables.

Why this answer

One-hot encoding is appropriate for high-cardinality categorical variables because it creates binary columns for each category, allowing the model to treat each category as an independent feature without imposing an ordinal relationship. This is crucial for classification models that assume numerical inputs, as it prevents the model from misinterpreting arbitrary integer labels as having meaningful order or magnitude.

Exam trap

Candidates may incorrectly discard label encoding (option E) because they assume it always imposes an ordinal relationship, but in many classification contexts (e.g., tree-based models) label encoding can handle high-cardinality nominal variables without issue. The trap is to overgeneralize the rule against label encoding, even when the question lists it as a correct answer alongside one-hot encoding.

Full explanation →

432

Multi-Selecthard

During a security audit of an AI-powered code generation tool, the audit team discovers that the system prompt (which contains sensitive internal instructions) can be leaked through carefully crafted user inputs. Which THREE OWASP LLM Top 10 categories are MOST directly relevant to this finding?

Select 3 answers

A.Model denial of service

B.Prompt injection

C.Insecure output handling

D.Supply chain vulnerabilities

E.Sensitive information disclosure

AnswersB, C, E

Prompt injection (LLM01) is the direct attack technique that tricks the model into revealing the system prompt.

Why this answer

Prompt injection (direct or indirect) is the attack vector that causes the system prompt leak. Sensitive information disclosure is the consequence. Insecure output handling can also be relevant if the leak is due to improper output management.

Model denial of service, supply chain vulnerabilities, and training data poisoning are not directly related to prompt leaking.

Full explanation →

433

MCQeasy

Which technique adds controlled noise to query results or training data to prevent an attacker from inferring whether a specific individual's data was included in the dataset?

A.Anonymisation

B.Federated learning

C.Differential privacy

D.Pseudonymisation

AnswerC

Differential privacy injects noise into computations or outputs to bound the risk of re-identification.

Why this answer

Differential privacy adds calibrated noise to ensure the output does not reveal individual participation. Anonymisation removes identifiers. Pseudonymisation replaces identifiers.

Federated learning decentralises data but does not necessarily add noise.

Full explanation →

434

MCQeasy

A company wants to build a system that automatically tags uploaded images with objects they contain (e.g., 'car', 'tree', 'person'). Which AI application type is this?

A.Image classification/object detection

B.Recommendation system

C.Anomaly detection

D.Document intelligence

AnswerA

Object detection identifies and localizes objects in images, matching the requirement.

Why this answer

The task of identifying and labeling objects (e.g., 'car', 'tree', 'person') within an image is a classic use case for image classification combined with object detection. Image classification assigns a single label to the entire image, while object detection localizes and classifies multiple objects within the image, which is exactly what the system requires.

Exam trap

The AI0-001 exam often tests the distinction between image classification (single label per image) and object detection (multiple localized objects), so candidates may mistakenly choose image classification alone when the question implies multiple objects per image.

How to eliminate wrong answers

Option B is wrong because recommendation systems analyze user behavior and preferences to suggest items (e.g., movies, products), not to identify objects in images. Option C is wrong because anomaly detection identifies unusual patterns or outliers in data (e.g., fraud detection), not the presence of common objects in images. Option D is wrong because document intelligence focuses on extracting text, structure, and information from documents (e.g., OCR, form processing), not on visual object recognition.

Full explanation →

435

MCQeasy

Refer to the exhibit. What is the recall of the model?

A.0.72

B.0.8

C.0.7

D.0.73

AnswerB

TP=80, FN=20, recall=80/100=0.8

Why this answer

Recall is calculated as True Positives divided by (True Positives + False Negatives). From the confusion matrix, True Positives = 80 and False Negatives = 20, so recall = 80 / (80 + 20) = 80 / 100 = 0.8. This measures the model's ability to correctly identify all actual positive cases.

Exam trap

CompTIA often tests recall by providing a confusion matrix and expects candidates to correctly identify the denominator as TP+FN, not total samples, to avoid confusing recall with accuracy or precision. In this question, the key is correctly reading TP=80 and FN=20 from the exhibit.

How to eliminate wrong answers

Option A (0.72) is wrong because it incorrectly uses True Positives divided by total predictions (72/100 = 0.72), which is accuracy, not recall. Option C (0.7) is wrong because it likely results from misreading the matrix (e.g., using 72/102 or confusing with precision). Option D (0.73) is wrong because it may come from a miscalculation such as (72 + 1)/(72 + 18 + 10) = 73/100, which is not a standard metric.

Full explanation →

436

MCQhard

A team is training a deep learning model for image classification. The training loss decreases steadily but the validation loss plateaus after 20 epochs and then starts to increase. Which action is MOST likely to improve generalization?

A.Add more convolutional layers

B.Increase the learning rate

C.Implement early stopping

D.Reduce the batch size

AnswerC

Early stopping monitors validation loss and stops training before overfitting occurs, directly addressing the plateau and rise.

Why this answer

Early stopping halts training when validation loss stops improving, preventing overfitting. Increasing learning rate would worsen divergence; adding more layers increases capacity and overfitting; reducing batch size may help optimization but not directly address overfitting.

Full explanation →

437

MCQmedium

An e-commerce company uses an AI system to set dynamic prices for products. A customer complains that the price they see is higher than the price shown to a friend for the same product at the same time. The company wants to ensure pricing fairness. Which ethical principle should guide the redesign of the pricing algorithm?

A.Transparency and explainability

B.Privacy by design

C.Accountability

D.Beneficence

AnswerA

Transparency requires the company to disclose how prices are determined, helping to ensure fairness and build trust.

Why this answer

Transparency and explainability is the correct principle because the core issue is that the customer cannot understand why the AI system set a different price for them compared to their friend. Redesigning the algorithm to provide clear, understandable reasons for price variations—such as demand, purchase history, or time of day—directly addresses this lack of visibility. This principle ensures that the system's decision-making process is open to scrutiny, which is essential for building trust and resolving fairness complaints in dynamic pricing models.

Exam trap

CompTIA often tests the distinction between 'accountability' (who is responsible) and 'transparency' (how the decision is made), leading candidates to pick accountability when the question explicitly asks for the principle that guides the redesign to ensure fairness through understanding.

How to eliminate wrong answers

Option B (Privacy by design) is wrong because the complaint is about price disparity and lack of understanding, not about how customer data is collected, stored, or protected. Option C (Accountability) is wrong because while accountability is important for assigning responsibility, it does not directly solve the customer's need to understand why the price differs; it focuses on who is responsible rather than making the algorithm's logic visible. Option D (Beneficence) is wrong because beneficence refers to doing good or maximizing benefits, but the immediate ethical failure here is the lack of clarity and justification for the pricing decision, not the absence of overall positive outcomes.

Full explanation →

438

MCQmedium

An AI team is deploying a large language model for a customer-facing application. They need to ensure that the model's output is always in valid JSON format for downstream processing. Which prompt engineering technique should they use?

A.Enable JSON mode in the model's API parameters

B.Use few-shot examples of JSON outputs in the prompt

C.Post-process the output with a JSON validator and reject invalid responses

D.Add a system prompt that says 'You are a helpful assistant.'

AnswerA

JSON mode instructs the model to produce only valid JSON, ensuring downstream parsability.

Why this answer

Structured output via JSON mode (available in many LLM APIs) constrains the model to output only valid JSON, which is critical for programmatic consumption.

Full explanation →

439

MCQmedium

A data scientist is training a resume screening model to rank job applicants. The training data includes historical hiring decisions from the past 10 years. The company wants to avoid unfair bias against underrepresented groups. Which type of bias is most likely present in the training data?

A.Algorithmic bias

B.Selection bias

C.Confirmation bias

D.Historical bias

AnswerD

Historical bias is present when the training data reflects existing societal inequalities, such as discriminatory hiring practices.

Why this answer

Historical bias occurs when the training data reflects past societal biases, such as underrepresentation of certain groups in hiring. The model learns these patterns, perpetuating unfairness. Selection bias arises from non-random sampling, confirmation bias from favoring information that confirms preexisting beliefs, and algorithmic bias from model design choices.

Full explanation →

440

Multi-Selectmedium

Which TWO actions should be taken to ensure an AI model complies with GDPR requirements when processing personal data?

Select 2 answers

A.Limit data collection to only what is necessary for the model

B.Provide a full explanation of model predictions

C.Store all user data for a minimum of 10 years

D.Anonymize all personal data before use

E.Implement user data deletion upon request

AnswersA, E

Data minimization is a GDPR principle.

Why this answer

GDPR's data minimization principle (Article 5(1)(c)) requires that personal data collected be adequate, relevant, and limited to what is necessary for the purpose for which it is processed. In AI model training, this means collecting only the features essential for the model's objective, reducing the risk of processing excessive or irrelevant personal data.

Exam trap

CompTIA often tests the misconception that anonymization is always required before any AI processing of personal data, but GDPR allows processing under lawful bases without anonymization, making Option D a tempting but incorrect choice.

Full explanation →

441

MCQeasy

A developer wants to secure an AI API service. Which practice is MOST effective for preventing unauthorized access to the model?

A.Using a larger context window

B.Enforcing least-privilege API access with proper key management

C.Enabling response logging

D.Implementing rate limiting

AnswerB

Correct. Least-privilege and key management are foundational access controls.

Why this answer

Enforcing least-privilege API access with proper key management is the most effective practice because it ensures that each API key or token has only the minimum permissions necessary for its intended function, reducing the attack surface. Proper key management includes rotating keys, using scoped access tokens (e.g., OAuth 2.0 scopes), and storing keys securely (e.g., using a secrets manager like AWS Secrets Manager or HashiCorp Vault). This directly prevents unauthorized access by limiting what a compromised or misused key can do, unlike other options that address secondary concerns.

Exam trap

The AI0-001 exam often tests the distinction between preventive and detective controls, and the trap here is that candidates confuse rate limiting (a throttling mechanism) with access control, thinking it prevents unauthorized access when it only limits the frequency of requests.

How to eliminate wrong answers

Option A is wrong because using a larger context window increases the amount of input the model can process but does nothing to authenticate or authorize API requests; it is a model configuration parameter, not a security control. Option C is wrong because enabling response logging aids in auditing and detecting breaches after they occur, but it does not prevent unauthorized access in real time; it is a detective control, not a preventive one. Option D is wrong because implementing rate limiting mitigates denial-of-service attacks and abuse by throttling request volume, but it does not verify the identity or permissions of the requester; an attacker with a valid key could still access the model within rate limits.

Full explanation →

442

MCQeasy

A startup wants to add an AI-powered virtual assistant to their mobile app. They have limited in-house AI expertise and need a solution that can be integrated quickly with minimal infrastructure management. Which deployment pattern is MOST suitable?

A.Implement an asynchronous processing queue for all user requests

B.Train and deploy a custom model on an on-premises server

C.Deploy the model on edge devices for offline inference

D.Use a cloud-based AI microservice (e.g., Amazon Lex, Azure Bot Service) with a pre-built model

AnswerD

Cloud AI microservices provide ready-to-use models, easy integration, and managed infrastructure, ideal for rapid development.

Why this answer

Using AI microservices from a cloud provider (e.g., AWS, Azure, GCP) allows quick integration, scalability, and minimal management. Training on-premises requires expertise and resources. Edge deployment is complex.

Async queues are for batch processing, not real-time assistant.

Full explanation →

443

MCQmedium

A data science team uses Git for version control of model code and DVC for data versioning. They want to implement a model registry to track trained models, their hyperparameters, and performance metrics. Which tool is specifically designed for this purpose and integrates with the existing workflow?

A.Apache Airflow

B.Docker

C.MLflow Model Registry

D.Kubernetes

AnswerC

MLflow provides a model registry that stores model versions and metadata.

Why this answer

MLflow Model Registry is specifically designed for managing model versions, tracking metadata, and integrating with Git and DVC. Apache Airflow is for workflow orchestration, not model registry. Kubernetes is for container orchestration.

Docker is for containerization.

Full explanation →

444

MCQeasy

An organisation is developing an AI policy. According to the NIST AI RMF, which function involves establishing policies and procedures to ensure the organisation governs AI responsibly?

A.Manage

B.Measure

C.Govern

D.Map

AnswerC

Govern involves setting policies, roles, and responsibilities for AI governance.

Why this answer

The NIST AI RMF's Govern function focuses on establishing governance structures, policies, and accountability mechanisms. Map, Measure, and Manage are other functions in the framework.

Full explanation →

445

MCQeasy

Which embedding type is MOST suitable for capturing semantic meaning of text in a RAG pipeline?

A.Bag-of-words vectors

B.Dense embeddings from a pre-trained transformer model

C.TF-IDF vectors

D.One-hot encoding

AnswerB

Dense embeddings capture contextualized semantic meaning, enabling effective similarity search.

Why this answer

Dense embeddings represent semantic meaning in a continuous vector space, ideal for similarity search in RAG.

Full explanation →

446

MCQmedium

A company has a TensorFlow model trained on-premises and wants to deploy it on AWS SageMaker for scalable inference. What is the BEST way to package the model for deployment?

A.Convert the model to ONNX and upload to SageMaker

B.Upload the .h5 file to S3 and create a SageMaker endpoint directly

C.Package the model in a Docker container with a TensorFlow serving script and push to Amazon ECR

D.Use SageMaker Studio to train the model again from scratch

AnswerC

This creates an inference container that SageMaker can deploy; it includes the model and serving logic.

Why this answer

SageMaker expects models in a container format; the inference container should include the model artifacts and the serving code, allowing SageMaker to host it on scalable endpoints.

Full explanation →

447

MCQeasy

An AI security team is conducting a threat model for a new document summarization service. They want to identify threats related to spoofing of the AI's identity. Which STRIDE category should they consider?

A.Repudiation

B.Tampering

C.Information disclosure

D.Spoofing

AnswerD

Spoofing involves impersonation, such as an attacker pretending to be the AI service.

Why this answer

Spoofing in STRIDE refers to impersonating something or someone else. In the context of AI, an attacker could spoof the AI service to provide false summaries.

Full explanation →

448

Multi-Selecthard

A team is implementing a RAG system for a legal document Q&A. They need to chunk documents effectively. Which THREE chunking strategies should they consider to improve retrieval accuracy for legal texts that contain hierarchical sections (clauses, sub-clauses, definitions)?

Select 3 answers

A.Hierarchical chunking that indexes chunks at clause and sub-clause levels with parent relationships

B.Overlapping chunks with a 10% overlap between consecutive chunks

C.Fixed-size chunking with a 512-token window and no overlap

D.Chunking based on the document's table of contents and section hierarchy

E.Semantic chunking that splits at natural boundaries (e.g., section headings, paragraph breaks)

AnswersA, D, E

Allows retrieval of granular chunks while maintaining broader context.

Why this answer

Semantic chunking splits at natural boundaries (e.g., paragraphs, sections), preserving meaning. Hierarchical chunking indexes with parent-child relationships for context. Fixed-size chunking is simple but may break sentences or clauses.

Overlapping chunks can help but is not a primary strategy for accuracy; sliding window is a specific technique.

Full explanation →

449

MCQhard

A healthcare startup is developing a deep learning model to detect diabetic retinopathy from retinal fundus images. The dataset contains 50,000 images, but only 5% are labeled as positive for the disease. The team uses a convolutional neural network (CNN) with a final sigmoid layer and binary cross-entropy loss. After training for 20 epochs, the model achieves 95% accuracy on the test set, but the recall for the positive class is only 10%. The team suspects the model is biased toward the negative class due to class imbalance. The data is stored in a secure environment, and no additional labeled data can be obtained. The team has access to the following techniques: oversampling the minority class, undersampling the majority class, using class weights in the loss function, applying data augmentation, and using a different architecture. Which course of action is most likely to improve recall for the positive class while maintaining reasonable overall performance?

A.Undersample the majority class to balance the dataset

B.Oversample the minority class using synthetic image generation

C.Assign higher class weights to the positive class in the loss function

D.Replace the CNN with a transformer-based architecture

AnswerC

Class weights force the model to focus on the minority class, improving recall.

Why this answer

Assigning higher class weights to the positive class in the loss function directly penalizes misclassifications of the minority class during training. This forces the model to pay more attention to positive samples without altering the dataset distribution, which is critical when no additional labeled data can be obtained and the data is in a secure environment. It improves recall by increasing the gradient contribution from positive samples, while maintaining overall performance because the model still sees the original data distribution.

Exam trap

The trap here is that candidates often choose oversampling (Option B) as the default solution for class imbalance, but fail to recognize that synthetic image generation for medical images can introduce unrealistic patterns and is not a standard or safe technique, whereas class weights are a lightweight, data-preserving approach that directly addresses the loss function.

How to eliminate wrong answers

Option A is wrong because undersampling the majority class discards a large number of negative samples, which can lead to loss of valuable information and degrade overall accuracy, especially with a 95% negative class. Option B is wrong because oversampling the minority class using synthetic image generation (e.g., SMOTE) is not directly applicable to high-dimensional image data without careful adaptation, and it may introduce unrealistic artifacts that harm generalization; the question specifies 'synthetic image generation' which is not a standard or safe approach for retinal fundus images. Option D is wrong because replacing the CNN with a transformer-based architecture does not address the class imbalance problem; transformers are not inherently better at handling imbalanced data and would require more data and computational resources, which are not available here.

Full explanation →

450

MCQhard

During inference, a model served via a REST API occasionally returns high latency due to cold starts. The team uses a containerized service on Kubernetes with horizontal pod autoscaling. Which solution minimizes cold start impact while controlling cost?

A.Configure the autoscaler based on request count with a shorter cooldown period

B.Increase CPU and memory requests for the inference container

C.Switch to vertical pod autoscaling

D.Use a sidecar container that pre-warms the model and set a minimum replica count

AnswerD

Pre-warming ensures the model is loaded; minimum replicas keep pods ready, reducing cold starts.

Why this answer

A sidecar warm-up agent and a minimum replica count keep pods ready. Increasing resources may not fix cold starts; autoscaling based on request count may lag; vertical scaling helps but not directly.

Full explanation →

Page 6 of 11

All pages

1 2 3 4 5 6 7 8 9 10 11

Practice AI0-001 by domain

Target a specific domain to shore up weak areas.

AI Infrastructure and Technologies AI Security AI Concepts and Foundations AI Concepts and Techniques Machine Learning and Deep Learning AI Models and Data Engineering Implementing AI Solutions AI Implementation and Operations AI Security, Ethics and Governance AI Governance and Ethics

See all domains with question counts →