Courseiva

CompTIA AI+ AI0-001 (AI0-001) — Questions 1–75

962 questions total · 13pages · All types, answers revealed

Take a mock examExam hub

Page 1 of 13

Page 2
1
MCQeasy

When implementing a vector store for a RAG system, which similarity search metric is MOST commonly used to find the most relevant document chunks for a given query embedding?

A.Manhattan distance
B.Euclidean distance
C.Dot product
D.Cosine similarity
AnswerD

Cosine similarity measures the angle between query and chunk embeddings, ignoring magnitude, which suits text embeddings where direction encodes semantic meaning. It is the default metric in most vector stores such as Azure AI Search.

Why this answer

Cosine similarity is the most common metric for comparing embedding vectors in RAG because it measures the angle between vectors, which works well for high-dimensional semantic embeddings.

Full explanation →
2
MCQmedium

An AI risk manager is applying the NIST AI Risk Management Framework (AI RMF). In which function would the organization establish a risk management process and assign roles and responsibilities for AI oversight?

A.Map
B.Manage
C.Govern
D.Measure
AnswerC

The Govern function establishes the policies, processes, roles and responsibilities for AI risk management across the organisation. Assigning oversight accountability and defining the risk process therefore sits in Govern, which satisfies the stem's requirement before Map, Measure and Manage activities occur.

Why this answer

The Govern function in the NIST AI RMF is specifically designed to establish organizational structures, policies, and accountability mechanisms for AI risk management. This includes defining roles and responsibilities, setting risk management processes, and ensuring oversight across the AI lifecycle. The other functions (Map, Measure, Manage) focus on different aspects such as understanding context, assessing risks, and treating risks, respectively.

Exam trap

A common trap is to assume that the 'Manage' function covers all risk management activities including establishing processes and roles, because its name implies broad oversight. However, in the NIST AI RMF, 'Govern' is the specific function for setting up risk management processes and accountability structures, while 'Manage' is reserved for risk treatment after assessment.

How to eliminate wrong answers

Option A is wrong because the Map function focuses on understanding the AI system's context, including its intended use, stakeholders, and potential impacts, not on establishing governance structures or assigning roles. Option B is wrong because the Manage function deals with prioritizing, responding to, and treating identified risks after they have been assessed, not with setting up the initial risk management process or assigning oversight roles. Option D is wrong because the Measure function involves quantitative and qualitative assessment of AI risks, including metrics and monitoring, but does not cover the establishment of governance processes or role assignment.

Full explanation →
3
MCQeasy

A data scientist is training a binary classification model to detect fraudulent transactions. The dataset has 99% legitimate transactions and 1% fraudulent. The model achieves 99% accuracy but fails to catch most fraud. Which metric should the team prioritize to evaluate model performance?

A.F1 score
B.Precision
C.Accuracy
D.Recall
AnswerD

With 99% legitimate transactions, a model predicting everything as legitimate scores 99% accuracy while catching no fraud. Recall measures the proportion of actual frauds detected, exposing that failure, whereas accuracy is misleading under such class imbalance.

Why this answer

Recall (sensitivity) measures the proportion of actual positive cases (fraud) correctly identified. With 99% accuracy but failing to catch most fraud, the model is biased toward the majority class (legitimate transactions), so recall is the critical metric to ensure fraud detection improves.

Exam trap

CompTIA often tests the misconception that high accuracy implies good model performance, especially in imbalanced datasets, leading candidates to overlook recall as the appropriate metric for minority class detection.

How to eliminate wrong answers

Option A is wrong because F1 score is the harmonic mean of precision and recall; while useful, it does not isolate the model's ability to catch fraud, and in this imbalanced dataset, a high F1 could still mask poor recall if precision is high. Option B is wrong because precision measures how many predicted frauds are actually fraud, but the model's failure to catch most fraud means recall is the primary concern, not the false positive rate. Option C is wrong because accuracy is misleading in imbalanced datasets; 99% accuracy can be achieved by simply predicting 'legitimate' for all transactions, which explains why the model fails to detect fraud.

Full explanation →
4
Multi-Selectmedium

A team is designing a deep learning pipeline for a computer vision task. They want to reduce overfitting. Which two techniques are specifically effective for this purpose? (Select TWO.)

Select 2 answers
A.Dropout
B.Using a smaller batch size
C.Adding more layers
D.L2 weight regularization
E.Increasing the learning rate
AnswersA, D

Dropout randomly deactivates neurons, reducing overfitting by preventing reliance on specific features.

Why this answer

Dropout is a regularization technique that randomly drops a fraction of neurons during training, which prevents the network from relying too heavily on any single neuron and forces it to learn more robust features. This reduces overfitting by introducing noise that improves generalization.

Exam trap

This exam often tests the misconception that increasing model capacity (more layers) or adjusting batch size directly reduces overfitting, when in fact these changes typically require additional regularization to be effective.

Full explanation →
5
MCQmedium

An AI engineer is tuning a large language model for a summarization task. The output summaries are too verbose and include irrelevant details. Which technique should be applied to encourage concise outputs?

A.Provide a few-shot example with concise summaries
B.Use chain-of-thought prompting
C.Decrease the top-k value
D.Increase the temperature
AnswerA

Few-shot prompting supplies in-context demonstrations that steer the model's output distribution toward the desired style. By including concise summary exemplars in the prompt, the model infers the expected length and level of detail, directly countering the verbosity and irrelevant content described in the stem without retraining.

Why this answer

Providing a few-shot example with concise summaries (Option A) directly demonstrates the desired output format to the model, leveraging in-context learning to bias generation toward brevity and relevance. This is the most effective technique for controlling output style without altering the model's underlying parameters.

Exam trap

CompTIA often tests the misconception that adjusting sampling parameters (top-k, temperature) is the primary way to control output length, when in fact these parameters affect randomness and diversity, not the explicit length or relevance of the generated text.

How to eliminate wrong answers

Option B is wrong because chain-of-thought prompting encourages step-by-step reasoning, which typically increases verbosity and is designed for complex reasoning tasks, not for reducing output length. Option C is wrong because decreasing the top-k value restricts the sampling pool to the k most likely tokens, which can reduce randomness but does not inherently enforce conciseness or relevance; it may even produce repetitive or incomplete summaries. Option D is wrong because increasing the temperature raises the randomness of token selection, often leading to more diverse but also more verbose and irrelevant outputs, the opposite of the desired effect.

Full explanation →
6
Multi-Selectmedium

When evaluating a binary classification model, which two metrics are most appropriate for imbalanced datasets? (Choose two.)

Select 2 answers
A.Accuracy
B.Mean absolute error
C.Recall
D.R-squared
E.Precision
AnswersC, E

Recall measures the proportion of actual positives correctly identified, so on a 99.9:0.1 dataset it exposes how many rare fraudulent or minority cases the model misses. Accuracy would look deceptively high, making recall essential for imbalanced evaluation.

Why this answer

Recall (C) is correct because it measures the proportion of actual positives that were correctly identified (TP / (TP + FN)), which directly reflects how well the model catches the minority class that would otherwise be swamped in an imbalanced dataset. Precision (E) is correct because it measures the proportion of predicted positives that are truly positive (TP / (TP + FP)), exposing the false-positive cost that becomes critical when the positive class is rare. Together, precision and recall (often summarized by F1 or PR-AUC) focus on minority-class performance rather than being dominated by the majority class.

Accuracy (A) is not appropriate because a trivial majority-class predictor can score very high accuracy while completely missing the minority class. Mean absolute error (B) and R-squared (D) are regression metrics and do not apply to binary classification evaluation.

Exam trap

CompTIA often tests the misconception that accuracy is always the best metric, but the trap here is that accuracy fails on imbalanced datasets, and candidates must recognize that recall and precision are the appropriate pair for evaluating minority class performance.

Full explanation →
7
MCQhard

A self-driving car company is testing a perception model that detects pedestrians. The model achieves 99% accuracy on the test set but fails to detect pedestrians wearing dark clothing at night. The company wants to improve the model's robustness. Which action should the team take to best address this specific weakness?

A.Collect and annotate more nighttime images of pedestrians wearing dark clothing and add them to the training set.
B.Increase the model's depth by adding more layers to capture more complex patterns.
C.Apply data augmentation by randomly rotating and flipping existing daytime images.
D.Reduce the model's confidence threshold for pedestrian detection to increase sensitivity.
AnswerA

Adding targeted training data that represents the failure case directly addresses the model's weakness. By exposing the model to more examples of dark-clothed pedestrians at night, it can learn the relevant features and improve detection. This data-centric approach is often the most effective for specific robustness gaps.

Why this answer

Collecting and annotating more nighttime images of pedestrians in dark clothing is the most direct way to address the specific weakness. The model's failure indicates a gap in the training distribution, so adding representative data will help it learn the necessary visual cues. Other options either do not target the root cause or introduce trade-offs without improving fundamental recognition.

Exam trap

The trap here is thinking that increasing model complexity or adjusting the decision threshold can compensate for missing training data, when the core issue is a data distribution gap.

Full explanation →
8
MCQhard

A data scientist trains a deep learning model on a large dataset. The training loss decreases steadily but the validation loss starts increasing after 20 epochs. The scientist uses early stopping with patience=5. Which of the following is the MOST likely cause and best corrective action?

A.Model is overfitting; add dropout regularization.
B.Training data is not representative; collect more data.
C.Model is underfitting; increase model capacity.
D.Learning rate too high; reduce learning rate.
AnswerA

Overfitting is the cause: training loss falls while validation loss rises, so the model memorises training data rather than generalising. Dropout randomly deactivates neurons during training, constraining the network's capacity and reducing this divergence, satisfying the stem's requirement to address the validation-loss increase directly.

Why this answer

The training loss decreasing while validation loss increasing after 20 epochs is a classic sign of overfitting, where the model memorizes training data noise instead of generalizing. Early stopping with patience=5 would halt training after 5 epochs of no validation improvement, but the root cause is overfitting. Adding dropout regularization randomly drops neurons during training, forcing the network to learn more robust features and reducing overfitting.

Exam trap

CompTIA often tests the distinction between overfitting and underfitting by showing a diverging validation loss curve, and the trap here is that candidates may confuse overfitting with a learning rate issue or data quality problem, leading them to choose 'reduce learning rate' or 'collect more data' instead of the correct regularization technique.

How to eliminate wrong answers

Option B is wrong because the validation loss increasing while training loss decreases indicates overfitting, not unrepresentative data; collecting more data might help but is not the most direct corrective action for overfitting. Option C is wrong because underfitting would show high training loss that does not decrease, not a decreasing training loss with increasing validation loss. Option D is wrong because a high learning rate would typically cause training loss to oscillate or diverge, not steadily decrease; reducing learning rate addresses convergence issues, not overfitting.

Full explanation →
9
MCQmedium

A company is training a large language model and wants to reduce its carbon footprint. Which practice is MOST effective for reducing training energy consumption while maintaining model quality?

A.Increase the batch size to the maximum the GPU memory allows
B.Use a larger model architecture to achieve higher accuracy faster
C.Use mixed-precision training and prune unnecessary parameters
D.Train the model on CPUs instead of GPUs
AnswerC

Mixed-precision training halves memory and compute per operation, while pruning removes redundant parameters, cutting training energy and carbon emissions directly. Both techniques preserve model quality, satisfying the requirement to reduce energy consumption without degrading accuracy.

Why this answer

Mixed-precision training (e.g., FP16/BF16 with FP32 master weights) reduces memory bandwidth and compute cost per operation, while pruning removes redundant parameters, lowering FLOPs and energy per training step. Together they cut energy consumption substantially without materially degrading model quality when done carefully. This is the most effective listed practice for reducing training energy while preserving quality.

Exam trap

AI0-001 often tests the misconception that bigger batch sizes or larger models automatically improve efficiency; candidates must recognize that precision reduction and pruning directly lower energy per useful training step.

How to eliminate wrong answers

Option A is wrong because simply maximizing batch size can improve hardware utilization but does not inherently reduce total energy per unit of model quality and may require more epochs or cause convergence issues. Option B is wrong because larger models increase compute and energy consumption, directly worsening the carbon footprint. Option D is wrong because CPUs are far less energy-efficient than GPUs for the parallel matrix operations in LLM training, so training on CPUs would dramatically increase energy use and time.

Full explanation →
10
MCQeasy

A healthcare organization uses an AI model to recommend treatment plans. The model was trained on data from a single hospital, and now treats patients from multiple demographics. Which ethical concern is most critical?

A.Accountability for treatment outcomes
B.Lack of transparency in model decisions
C.Privacy violations in training data
D.Fairness and bias in predictions
AnswerD

Training on a single hospital's data embeds that population's demographics, so predictions for other groups inherit skewed patterns. This directly creates disparate performance across demographics, making fairness and bias the critical ethical concern the multi-demographic scenario raises.

Why this answer

The model was trained on data from a single hospital, which likely has a homogeneous demographic profile. When deployed across multiple demographics, the model may produce biased or unfair predictions for underrepresented groups, making fairness and bias the most critical ethical concern. This directly violates the principle of distributive justice in AI ethics.

Exam trap

The AI0-001 exam often tests the distinction between general ethical principles (like accountability or transparency) and the specific, root-cause ethical violation triggered by the scenario, which here is fairness and bias due to demographic mismatch in training data.

How to eliminate wrong answers

Option A is wrong because accountability for treatment outcomes is a general ethical concern but not the most critical here; the primary issue is that the model's training data lacks demographic diversity, which leads to biased predictions before accountability can even be assessed. Option B is wrong because lack of transparency (black-box nature) is a separate concern; while it can exacerbate bias, the core problem is that the model's training data does not represent the target population, not that the model's decisions are opaque. Option C is wrong because privacy violations in training data are a valid concern but not directly triggered by the scenario; the scenario describes using data from a single hospital, which does not inherently imply privacy breaches, whereas the demographic shift introduces bias.

Full explanation →
11
MCQmedium

A product team wants a system that can generate high-quality synthetic images of furniture in different room settings for an online catalog. The images must be photorealistic and vary in style. Which generative AI approach is BEST suited for this task?

A.Variational autoencoder (VAE)
B.Diffusion model
C.Generative adversarial network (GAN)
D.Recurrent neural network (RNN)
AnswerB

Diffusion models generate images by iteratively denoising random noise, producing photorealistic outputs with fine detail and controllable stylistic variation. This directly satisfies the requirement for high-quality, style-varied furniture images, unlike GANs' instability or VAEs' blurrier results.

Why this answer

Diffusion models are the best choice because they iteratively denoise random noise to produce high-quality, photorealistic images with diverse styles. Unlike GANs, they avoid mode collapse and training instability, and they generate more detailed and varied outputs than VAEs, making them ideal for furniture catalog images in different room settings.

Exam trap

CompTIA AI often tests the misconception that GANs are always the best for image generation, but the trap here is that GANs' mode collapse and training instability make diffusion models superior for high-quality, diverse outputs in production systems.

How to eliminate wrong answers

Option A is wrong because VAEs generate blurry images due to their variational lower bound objective, which smooths over fine details, making them unsuitable for photorealistic furniture images. Option C is wrong because GANs can suffer from mode collapse, where they generate limited variations (e.g., only one style of room), and training instability, reducing reliability for diverse catalog images. Option D is wrong because RNNs are designed for sequential data (e.g., text, time series) and cannot generate high-dimensional spatial images like furniture in room settings.

Full explanation →
12
MCQeasy

A developer is using Hugging Face Transformers to fine-tune a BERT model for sentiment analysis. They want to track experiments, log metrics, and compare runs. Which MLOps tool should they integrate?

A.Apache Airflow
B.Docker
C.Kubeflow
D.MLflow
AnswerD

MLflow provides experiment tracking, metric logging and run comparison, integrating directly with Hugging Face Transformers training loops. It satisfies the stated need to track experiments and compare runs, which raw training scripts alone cannot deliver.

Why this answer

MLflow is the correct choice because it is purpose-built for experiment tracking, metric logging, and run comparison in machine learning workflows. It provides an API to log parameters, metrics, and artifacts, and its UI allows easy comparison of different fine-tuning runs, which directly matches the developer's need to track experiments and compare runs for a BERT sentiment analysis model.

Exam trap

CompTIA often tests the distinction between infrastructure tools (Airflow, Docker, Kubeflow) and ML-specific experiment tracking tools (MLflow), trapping candidates who confuse orchestration or containerization with MLOps tracking capabilities.

How to eliminate wrong answers

Option A is wrong because Apache Airflow is a workflow orchestration tool for scheduling and managing DAGs (Directed Acyclic Graphs) of tasks, not for experiment tracking or metric logging; it lacks native ML run comparison capabilities. Option B is wrong because Docker is a containerization platform for packaging applications and dependencies, not an MLOps tool for logging metrics or comparing experiments; it provides environment consistency but no tracking or logging features. Option C is wrong because Kubeflow is a Kubernetes-native platform for deploying and managing ML pipelines at scale, but it is overkill for simple experiment tracking and does not offer the lightweight, focused metric logging and run comparison that MLflow provides out of the box.

Full explanation →
13
MCQmedium

A financial services company trains a gradient-boosted classification model on a dataset that includes customer account balances. The security team wants to limit how much any single customer's balance can influence the model's learned parameters, because an attacker who obtains the trained model could otherwise probe it to recover specific training values. Which technique should they apply during training to cap the influence of individual records?

A.Increase the model's L2 regularization coefficient until training accuracy drops significantly.
B.Apply differential privacy with a bounded per-record gradient clipping norm and calibrated noise.
C.Hash each customer account balance with SHA-256 before feeding it to the training pipeline.
D.Encrypt the model artifacts at rest with a customer-managed key in the cloud KMS.
AnswerB

Differential privacy bounds each training record's contribution by clipping per-example gradients to a fixed norm, then adds calibrated noise to the aggregate update. This mathematically limits how much any one account balance can shift the learned parameters, so an attacker probing the released model cannot reliably infer whether a specific customer's record was present or recover its exact value.

Why this answer

Differential privacy with per-record gradient clipping and calibrated noise is the only listed technique that formally bounds how much any single training record can change the model's parameters. That bound is what prevents an attacker with access to the trained model from reliably reconstructing or confirming individual customer balances, whereas hashing, regularization, and at-rest encryption leave the influence of individual records unbounded.

Exam trap

The trap here is assuming that any privacy-preserving preprocessing step, such as hashing or encryption, limits how much a training record influences the model, when only differential privacy provides that formal bound.

Full explanation →
14
MCQmedium

A healthcare AI startup is developing a model to predict patient readmission risk. The company wants to ensure the model's decisions can be understood by clinicians. Which explainability technique provides local, model-agnostic explanations by fitting a simple surrogate model around a prediction?

A.SHAP values
B.LIME
C.Attention visualisation
D.Model cards
AnswerB

LIME fits a sparse linear surrogate around each individual prediction, so clinicians receive a local, model-agnostic explanation of which features drove that specific readmission risk score, satisfying the interpretability requirement without exposing the underlying model's internals.

Why this answer

LIME (Local Interpretable Model-agnostic Explanations) is the correct technique because it generates local explanations by fitting a simple, interpretable surrogate model (e.g., linear regression or decision tree) around a specific prediction. It is model-agnostic, meaning it works with any black-box classifier, and it perturbs the input data near the instance of interest to understand which features most influenced the prediction.

Exam trap

Candidates may confuse SHAP with LIME because both provide local, model-agnostic explanations, but LIME fits a surrogate model (e.g., linear regression) around the prediction, whereas SHAP uses Shapley values from game theory.

How to eliminate wrong answers

Option A is wrong because SHAP values provide both local and global explanations based on cooperative game theory (Shapley values), but they are not a surrogate model; they compute additive feature importance scores directly from the model's output. Option C is wrong because attention visualization is a technique specific to neural network architectures (e.g., transformers) and is not model-agnostic; it relies on internal attention weights, which are not available for arbitrary models. Option D is wrong because model cards are documentation artifacts that describe a model's intended use, performance, and limitations; they do not generate local explanations for individual predictions.

Full explanation →
15
MCQmedium

A company uses an AI model to generate personalized marketing emails. They want to prevent the model from leaking the system prompt used to configure its behavior. Which attack should they guard against?

A.Prompt leaking
B.Model inversion
C.Membership inference
D.Data poisoning
AnswerA

Prompt leaking is the extraction of the hidden system prompt through crafted queries, so the model reveals its configuration instructions. Guarding against it directly addresses the stated goal of preventing disclosure of the system prompt that shapes the model's behaviour.

Why this answer

Prompt leaking is an attack where an adversary crafts inputs to trick the model into revealing its system prompt or hidden instructions. Since the system prompt defines the model's behavior and often contains proprietary or sensitive configuration details, preventing its disclosure is critical. Guarding against prompt leaking directly addresses the goal of keeping the system prompt confidential.

Exam trap

CompTIA often tests the distinction between attacks on training data (model inversion, membership inference, data poisoning) versus attacks on the inference-time configuration (prompt leaking), so candidates mistakenly choose a training-data attack when the question explicitly targets the system prompt.

How to eliminate wrong answers

Option B is wrong because model inversion attacks aim to reconstruct training data from the model's outputs, not to extract the system prompt which is part of the model's runtime configuration, not its training data. Option C is wrong because membership inference attacks determine whether a specific data point was used in the model's training set, which is unrelated to leaking the system prompt. Option D is wrong because data poisoning involves corrupting the training data to alter the model's behavior, not extracting the system prompt that is provided at inference time.

Full explanation →
16
MCQeasy

Which machine learning paradigm is best suited for training a model to play a game by learning from its own actions and rewards, without labeled data?

A.Unsupervised learning
B.Semi-supervised learning
C.Reinforcement learning
D.Supervised learning
AnswerC

Reinforcement learning trains an agent through trial-and-error interaction with an environment, using reward signals rather than labelled examples. This matches the game-playing constraint exactly: the model learns optimal actions from its own experience and cumulative rewards, with no pre-labelled dataset required.

Why this answer

Reinforcement learning is designed for agents that learn by interacting with an environment and receiving rewards or penalties, with no labeled dataset required. The model improves its policy through trial and error, maximizing cumulative reward over time — exactly the paradigm for game-playing agents like AlphaGo and DQN.

Exam trap

AI0-001 often tests the distinction between reinforcement learning (reward-driven, no labels) and supervised learning (label-driven) — candidates pick supervised learning because games seem to have 'correct' moves, but RL learns from rewards, not labels.

How to eliminate wrong answers

Option A is wrong because unsupervised learning finds patterns or clusters in unlabeled data but has no reward signal or environment interaction, so it cannot learn a game-playing policy. Option B is wrong because semi-supervised learning uses a small amount of labeled data plus unlabeled data, which still requires labels and does not involve reward-based learning. Option D is wrong because supervised learning requires labeled input-output pairs, which are not available when an agent learns purely from its own actions and rewards.

Full explanation →
17
MCQhard

An ML team deploys a model on edge devices using INT8 quantization. They notice a significant drop in accuracy on a subset of classes. Which technique should they apply to recover accuracy without increasing model size?

A.Use pruning to remove less important weights
B.Increase the model architecture size
C.Switch to FP16 quantization
D.Apply quantization-aware training (QAT)
AnswerD

Quantization-aware training simulates INT8 rounding during the forward pass while keeping weights in higher precision for gradient updates, letting the model learn to compensate for that error. This recovers accuracy on the affected classes while the deployed artefact stays INT8, so model size is unchanged.

Why this answer

Quantization-aware training (QAT) simulates INT8 quantization effects during the forward pass of training, allowing the model to learn weights and activations that are more robust to the lower precision. This recovers accuracy lost during post-training quantization without increasing the model's size, as the architecture and number of parameters remain unchanged.

Exam trap

CompTIA AI often tests the misconception that post-training quantization is always lossless, leading candidates to overlook the need for QAT when accuracy drops on specific classes due to uneven weight distributions.

How to eliminate wrong answers

Option A is wrong because pruning reduces model size by removing less important weights, which does not directly address the accuracy drop caused by INT8 quantization and may further degrade performance. Option B is wrong because increasing the model architecture size would increase the model's memory footprint and latency, contradicting the requirement to not increase model size. Option C is wrong because switching to FP16 quantization uses 16-bit floating point, which increases the model size compared to INT8 and does not meet the constraint of maintaining the same model size.

Full explanation →
18
MCQhard

A team is implementing a RAG system for legal document retrieval. The documents are long (50-100 pages) with clear section headings. They want to ensure that retrieved chunks are semantically coherent and respect document structure. Which chunking strategy is MOST appropriate?

A.Semantic chunking based on sentence embeddings
B.Fixed-size chunking with 256 tokens and no overlap
C.Recursive character text splitting with chunk size 1000 and chunk overlap 200
D.Hierarchical chunking: first split by sections, then further split each section into fixed-size chunks with overlap
AnswerD

Hierarchical chunking splits first on section headings, preserving the document's logical structure, then subdivides oversized sections with overlap. This keeps chunks semantically coherent and respects structure, which fixed-size or naive splitting would fragment across headings.

Why this answer

Hierarchical chunking preserves document structure by first splitting into sections, then further into chunks, maintaining semantic coherence.

Full explanation →
19
MCQmedium

A developer is using a large language model via an API. They want the model to solve a math problem step by step. Which prompt engineering technique should they use?

A.Set temperature to 0.9
B.Chain-of-thought prompting
C.Few-shot prompting
D.Zero-shot prompting
AnswerB

Chain-of-thought prompting instructs the model to emit intermediate reasoning steps before the final answer, improving multi-step arithmetic accuracy. This directly satisfies the stem's requirement to solve a maths problem step by step rather than jumping to a result.

Why this answer

Chain-of-thought (CoT) prompting explicitly instructs the model to reason through intermediate steps before producing a final answer, which is exactly what's needed for multi-step math problems. By asking the model to 'think step by step,' it decomposes the problem into manageable reasoning stages, significantly improving accuracy on arithmetic and logic tasks.

Exam trap

AI0-001 often tests the confusion between few-shot prompting (providing examples) and chain-of-thought prompting (eliciting reasoning steps) — candidates see 'step by step' and pick few-shot because both involve guiding the model, but only CoT explicitly structures the reasoning process.

How to eliminate wrong answers

Option A is wrong because temperature controls randomness in token sampling — setting it to 0.9 increases creativity and variability, which is the opposite of what you want for deterministic math reasoning. Option C is wrong because few-shot prompting provides examples of input-output pairs but doesn't inherently force step-by-step reasoning; it's a demonstration technique, not a reasoning technique. Option D is wrong because zero-shot prompting simply asks the model to answer without examples or reasoning guidance, which performs poorly on complex math problems.

Full explanation →
20
MCQeasy

A logistics company uses a machine learning model to predict delivery times based on historical data. The model was performing well, but recently it started making inaccurate predictions, especially for routes that have experienced new traffic patterns and road closures. The data engineering team receives an alert that the model's accuracy has dropped by 15% over the last week. They suspect data drift. The team has access to the original training data and a continuous stream of new data. What is the most appropriate first step for the team to take?

A.Roll back the model to the previous stable version and schedule a full audit of the data pipeline.
B.Compare the distributions of key features between the training data and the recent data to quantify data drift.
C.Immediately retrain the model using the most recent data to adapt to the new patterns.
D.Add more features to the model to capture the new traffic patterns and road closures.
AnswerB

Distribution comparison directly quantifies drift by contrasting feature statistics between the original training data and recent streamed data, confirming whether new traffic patterns and closures shifted inputs. This diagnostic precedes retraining, isolating whether the 15% accuracy drop stems from input drift rather than label or concept change.

Why this answer

The first step in diagnosing a suspected data drift is to statistically compare the distributions of key features between the training data and the recent streaming data. This quantifies whether the input data distribution has changed, which directly explains the accuracy drop. Without this analysis, any corrective action (like retraining or rollback) would be premature and could mask the root cause.

Exam trap

CompTIA often tests the misconception that the immediate response to a performance drop should be retraining or rollback, rather than first diagnosing the type of drift (data drift vs. concept drift) through distribution comparison.

How to eliminate wrong answers

Option A is wrong because rolling back the model without first confirming data drift wastes time and may not address the new traffic patterns; it assumes the previous model is still valid, which is false if drift is present. Option C is wrong because immediately retraining on recent data without verifying drift could introduce bias or overfit to transient noise, and it ignores the need to first understand what changed. Option D is wrong because adding features without first analyzing drift is a blind attempt that may not solve the distribution shift and could increase model complexity unnecessarily.

Full explanation →
21
MCQeasy

An organization deploys an AI system that processes personal data of EU citizens. Which regulatory framework imposes strict requirements on automated decision-making and profiling?

A.Payment Card Industry Data Security Standard (PCI DSS)
B.General Data Protection Regulation (GDPR)
C.Health Insurance Portability and Accountability Act (HIPAA)
D.Sarbanes-Oxley Act (SOX)
AnswerB

The GDPR governs processing of EU citizens' personal data and, under Article 22, restricts solely automated decisions and profiling that produce legal or similarly significant effects, demanding safeguards such as human intervention and the right to contest. That directly satisfies the stem's constraint of strict automated decision-making requirements.

Why this answer

The General Data Protection Regulation (GDPR) is the correct regulatory framework because it specifically governs the processing of personal data of EU citizens and imposes strict requirements on automated decision-making and profiling under Article 22. This article grants individuals the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects or similarly significant effects. The GDPR also mandates data protection impact assessments and transparency obligations for such AI-driven processing.

Exam trap

The AI0-001 exam often tests candidates' ability to distinguish between data privacy regulations (GDPR) and industry-specific security standards (PCI DSS, HIPAA, SOX), trapping those who confuse data security with data protection governance for AI systems.

How to eliminate wrong answers

Option A is wrong because PCI DSS is a security standard for protecting payment card data, not a framework for regulating automated decision-making or profiling of EU citizens' personal data. Option C is wrong because HIPAA applies to protected health information in the United States and does not address automated decision-making or profiling under EU law. Option D is wrong because SOX is a US federal law focused on financial reporting and corporate governance, with no provisions for personal data processing or AI-driven profiling.

Full explanation →
22
MCQeasy

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

A.0.44
B.0.80
C.0.90
D.0.99
AnswerA

Recall equals true positives divided by the sum of true positives and false negatives. Reading the confusion matrix, 0.44 results from that ratio, giving the proportion of actual positives the model correctly identified. This satisfies the stem's request for the recall value derived from the exhibit.

Why this answer

Recall is calculated as True Positives divided by (True Positives + False Negatives). From the confusion matrix in the exhibit, True Positives = 400 and False Negatives = 500, so recall = 400 / (400 + 500) = 400 / 900 ≈ 0.44. Option A is correct because this matches the computed recall value.

Exam trap

CompTIA often tests the confusion between recall and precision, so candidates mistakenly compute precision (TP/(TP+FP)) instead of recall, leading them to choose 0.80.

How to eliminate wrong answers

Option B (0.80) is wrong because it likely results from incorrectly using precision (TP/(TP+FP) = 40/50 = 0.80) instead of recall. Option C (0.90) is wrong because it may come from dividing TP by total predictions (40/100 = 0.40) or misreading the matrix, but 0.90 is not supported by any standard metric from the given values. Option D (0.99) is wrong because it is far too high and could stem from confusing recall with accuracy or ignoring the false negatives entirely.

Full explanation →
23
Multi-Selectmedium

A company wants to adopt green AI practices to reduce the environmental impact of training large models. Which TWO actions are most effective?

Select 2 answers
A.Use efficient model architectures (e.g., pruning, quantization)
B.Use larger datasets to improve accuracy
C.Train models only on weekends
D.Train models in the cloud to offload energy costs
E.Use energy-efficient hardware (e.g., TPUs or optimized GPUs)
AnswersA, E

Pruning and quantization shrink parameter counts and numeric precision, cutting the compute and energy consumed during training and inference. This directly satisfies the stem's goal of reducing environmental impact, since training cost scales with model size and floating-point operations.

Why this answer

Using efficient model architectures (A) and energy-efficient hardware (E) directly reduce energy consumption. Using larger datasets (B) increases energy use. Training models on weekends (C) does not affect total energy consumption.

Training models in the cloud (D) may offload energy costs to the provider but does not reduce total energy used.

Full explanation →
24
MCQmedium

A machine learning engineer trains a decision tree to predict customer churn. The tree achieves 99 percent accuracy on the training set but only 68 percent on a held-out test set. The engineer wants to reduce this gap. Which single action is most likely to improve test performance?

A.Limit the tree depth and require a minimum number of samples per leaf
B.Train the tree on the test set as well so it learns the held-out distribution
C.Increase the maximum depth of the tree so it can fit more training examples
D.Add more features derived from the training set to give the tree more signal
AnswerA

Constraining depth and requiring a minimum samples per leaf restricts the tree from creating tiny, noise-driven splits. This regularization reduces variance and typically narrows the train-test gap. For a tree already at 99 percent training accuracy, these hyperparameters directly address the overfitting that causes the 31-point drop on held-out data.

Why this answer

The large train-test gap signals high variance, meaning the unconstrained decision tree memorized training noise. Pre-pruning through maximum depth and minimum samples per leaf limits how finely the tree can split, reducing variance and improving held-out accuracy without discarding the model or leaking test data.

Exam trap

The trap here is equating higher training accuracy with a better model, when a near-perfect training score alongside poor test performance is the classic signature of overfitting.

Full explanation →
25
MCQeasy

A retail analytics team wants to group customers into segments based on purchase frequency, average order value, and recency, without having any predefined segment labels. They plan to use an algorithm that partitions customers into a fixed number of groups by minimizing within-cluster variance. Which technique should they use?

A.Linear regression
B.K-means clustering
C.Logistic regression
D.Principal component analysis
AnswerB

K-means clustering partitions data into a fixed number of groups by minimizing within-cluster variance, exactly matching the described goal. It is unsupervised, so no predefined labels are needed. Given features like purchase frequency, average order value, and recency, K-means will assign customers to the nearest centroid, producing the desired segments.

Why this answer

K-means clustering is the correct technique because it is an unsupervised algorithm that partitions observations into a predefined number of clusters by minimizing the sum of squared distances to cluster centroids. The team's goal of grouping customers without labels aligns perfectly with K-means. The other options are either supervised prediction methods or dimensionality reduction techniques that do not produce discrete customer segments.

Exam trap

The trap here is confusing dimensionality reduction with clustering, since principal component analysis is often mentioned alongside K-means in preprocessing pipelines but does not itself create segments.

Full explanation →
26
Multi-Selecthard

A bank is deploying an LLM-based assistant that drafts responses to customer complaints. The assistant retrieves relevant policy passages from an internal vector database and includes them in the prompt. The security team wants to reduce the risk that an attacker can cause the assistant to reveal the full system prompt or internal policy text that the customer should not see. (Choose two.)

Select 2 answers
A.Increase the model's temperature setting so responses vary and attackers cannot reliably reproduce extracted content.
B.Apply the principle of least privilege so the assistant only retrieves policy passages relevant to the specific customer complaint.
C.Fine-tune the model on the bank's complete policy manual so it no longer needs retrieval at inference time.
D.Store the system prompt and policy passages in a separate encrypted database and grant the LLM read access only during inference.
E.Implement input and output filtering that detects and blocks attempts to extract system instructions or restricted policy content.
AnswersB, E

Limiting retrieval to passages needed for the current complaint shrinks the amount of sensitive policy text placed in the prompt, so even a successful extraction attempt exposes far less. This reduces the blast radius of prompt leakage and complements output filtering by minimizing what the model can potentially reveal.

Why this answer

Reducing prompt and policy leakage in a retrieval-augmented assistant requires limiting what sensitive content enters the context and inspecting what leaves it. Filtering input and output catches extraction attempts and redacts restricted text, while least-privilege retrieval minimizes the sensitive passages available to the model in the first place. Together they shrink both the likelihood and the impact of disclosure, whereas storage encryption, full fine-tuning, and temperature changes do not close the generation channel.

Exam trap

The trap here is treating encryption at rest or higher sampling temperature as protections against prompt leakage, when the actual disclosure path is the model reproducing content that was placed in its context window.

Full explanation →
27
MCQeasy

A company wants to build a real-time anomaly detection system for IoT sensor data using edge AI. The model must run on resource-constrained devices with minimal power consumption. Which model optimization technique is MOST important?

A.Use FP32 precision
B.Model quantization (INT8)
C.Increase the number of layers
D.Use a larger batch size
AnswerB

INT8 quantization shrinks weights and activations to 8-bit integers, cutting memory footprint and enabling integer arithmetic that draws far less power than FP32. This directly satisfies the stem's constraint of running on resource-constrained edge devices with minimal power consumption.

Why this answer

Quantization reduces model precision (e.g., FP32 to INT8), decreasing model size and computation, which is critical for resource-constrained edge devices.

Full explanation →
28
MCQhard

A financial institution uses an AI model to approve loans. The model uses features including credit score and ZIP code. During an audit, it is discovered that the model has a high false positive rate for loan default predictions in certain ZIP codes. What should the institution do to address this?

A.Remove the ZIP code feature from the model
B.Increase the decision threshold for those ZIP codes
C.Discontinue use of the model for those ZIP codes
D.Retrain the model with fairness constraints
AnswerD

Retraining with fairness constraints directly targets the ZIP-code disparity by adding a regularisation term that penalises disparate false positive rates across protected groups during optimisation. This satisfies the audit finding: the model's error rates vary by geography, so the constraint forces the learner to equalise them.

Why this answer

Retraining the model with fairness constraints directly addresses the root cause of the bias—the model's learned correlations between ZIP code and default risk. Fairness constraints, such as demographic parity or equalized odds, are applied during training to ensure the model's predictions are not systematically skewed against certain groups. This approach preserves the predictive power of legitimate features while mitigating discriminatory outcomes, aligning with AI governance principles.

Exam trap

The AI0-001 exam often tests the misconception that removing a sensitive feature (like ZIP code) is sufficient to eliminate bias, when in reality correlated proxy features can perpetuate discrimination—a concept known as 'fairness through unawareness' being a flawed approach.

How to eliminate wrong answers

Option A is wrong because removing the ZIP code feature may not eliminate bias if other features (e.g., income, credit history) are correlated with ZIP code, and it could reduce model accuracy by discarding legitimate predictive information. Option B is wrong because increasing the decision threshold for those ZIP codes is a post-hoc adjustment that treats the symptom (high false positives) without fixing the underlying bias, and it may introduce new disparities or violate regulatory requirements for consistent lending standards. Option C is wrong because discontinuing use of the model for those ZIP codes abandons the model's utility entirely for those areas, which is operationally impractical and does not address the bias—it simply avoids the problem rather than correcting it.

Full explanation →
29
Multi-Selectmedium

An e-commerce company operates an AI recommendation service. After a marketing campaign, the operations team notices that inference costs have tripled while request volume has only doubled. They need to reduce cost per inference without degrading recommendation quality. Which two actions should the team take? (Choose two.)

Select 2 answers
A.Disable caching of recommendation results so every request is computed fresh from the model.
B.Move the recommendation service to a larger GPU instance type with more memory.
C.Apply model quantization or a distilled smaller model for the recommendation ranking stage, validating quality against offline metrics.
D.Enable dynamic batching at the inference server so concurrent requests are grouped into a single model execution.
E.Increase the number of replicas behind the load balancer so each instance handles fewer requests.
AnswersC, D

Quantization or distillation reduces compute and memory per inference, directly lowering cost. Validating against offline metrics such as recall at k or NDCG ensures quality stays within acceptable bounds. This is a standard optimization when cost grows faster than traffic, and it complements batching by reducing the work per execution.

Why this answer

Cost per inference falls when each execution does more useful work or requires fewer resources. Dynamic batching amortizes execution overhead across concurrent requests, and quantization or distillation reduces the compute needed per prediction. Adding replicas, disabling caching, or upgrading instance size raises capacity or work without improving efficiency, so they do not meet the goal.

Exam trap

The trap here is equating more capacity with lower cost, when the objective is specifically cost per inference rather than raw throughput.

Full explanation →
30
MCQmedium

A data scientist is preparing a dataset for a binary classification model. The dataset has 95% majority class and 5% minority class. Which data preparation technique is BEST to address the class imbalance?

A.Min-max normalization of all features
B.Random undersampling of the majority class
C.Removing all minority class samples
D.SMOTE oversampling of the minority class
AnswerD

SMOTE generates synthetic minority-class samples by interpolating between existing minority neighbours, rebalancing the 95:5 split without discarding majority data. This gives the classifier more minority examples to learn from, unlike random undersampling which loses information.

Why this answer

SMOTE (Synthetic Minority Over-sampling Technique) is the best choice because it generates new, synthetic minority class samples by interpolating between existing minority instances and their nearest neighbors, rather than simply duplicating them. This directly addresses severe class imbalance (95:5) by enriching the minority class without discarding valuable majority data. Unlike random oversampling, SMOTE reduces the risk of overfitting to exact copies and helps the model learn a more generalizable decision boundary.

Exam trap

AI0-001 often tests the misconception that any resampling technique is equally valid for class imbalance, but the key is recognizing that SMOTE is preferred for severe imbalance because it synthesizes new minority samples without discarding majority data, unlike random undersampling which loses information.

How to eliminate wrong answers

Option A is wrong because min-max normalization only rescales feature values to a [0,1] range and has no effect on class distribution or imbalance. Option B is wrong because random undersampling discards a large portion of the majority class, which with a 95:5 ratio would leave very few majority samples and cause significant information loss, likely degrading model performance. Option C is wrong because removing all minority class samples eliminates the positive class entirely, making binary classification impossible and destroying the dataset's utility.

Full explanation →
31
MCQeasy

A marketing team wants to segment customers into groups based on purchasing behavior without predefined categories. Which algorithm should they use?

A.K-means clustering
B.Naive Bayes classifier
C.Logistic regression
D.Support vector machine
AnswerA

K-means partitions unlabelled data into k clusters by minimising within-cluster variance, discovering natural groupings in purchasing behaviour without predefined labels. This satisfies the stem's unsupervised, no-categories constraint, unlike classification algorithms that require labelled segment definitions upfront.

Why this answer

K-means clustering is an unsupervised learning algorithm that groups data points into clusters based on similarity without requiring predefined labels. Since the marketing team wants to segment customers based on purchasing behavior without predefined categories, K-means is the correct choice as it discovers natural groupings in the data.

Exam trap

CompTIA often tests the distinction between supervised and unsupervised learning, and the trap here is that candidates may confuse clustering (unsupervised) with classification (supervised) algorithms, leading them to pick a classifier like Naive Bayes or logistic regression instead of K-means.

How to eliminate wrong answers

Option B (Naive Bayes classifier) is wrong because it is a supervised learning algorithm that requires labeled training data to classify instances into predefined categories, making it unsuitable for discovering unknown segments. Option C (Logistic regression) is wrong because it is a supervised learning algorithm used for binary classification tasks, not for unsupervised clustering or segmentation without predefined groups. Option D (Support vector machine) is wrong because it is a supervised learning algorithm that separates data into predefined classes using hyperplanes, not for discovering hidden patterns or groupings in unlabeled data.

Full explanation →
32
MCQmedium

An AI agent is designed to book flights by calling an external API. The agent must decide which tool to call based on user input, then generate the correct API parameters. Which pattern is MOST appropriate for this workflow?

A.Chain-of-thought prompting only
B.Zero-shot prompting with JSON mode
C.ReAct pattern with tool descriptions and function calling
D.Simple prompt with no tool descriptions
AnswerC

ReAct interleaves reasoning with tool invocation, while function calling supplies the schema that constrains generated API parameters. Together they let the agent select the correct booking tool from its descriptions and emit valid arguments, satisfying the stem's decision and parameter-generation requirements.

Why this answer

The ReAct (Reasoning + Acting) pattern interleaves reasoning traces with tool calls, allowing the agent to decide which tool to invoke based on user input, observe the result, and iterate. Combined with function calling and tool descriptions, the model can select the correct API and generate structured parameters (e.g., JSON schema-conformant arguments) for the flight booking API. This is the canonical pattern for tool-using agents that must choose among multiple tools and produce correct parameters.

Exam trap

AI0-001 often tests the misconception that chain-of-thought or JSON mode alone enables tool use, when in fact tool selection and parameter generation require the ReAct pattern with tool descriptions and function calling.

How to eliminate wrong answers

Option A is wrong because chain-of-thought prompting only elicits reasoning steps; it does not provide a mechanism for the model to invoke external tools or generate structured API parameters, so the agent cannot actually call the flight API. Option B is wrong because zero-shot prompting with JSON mode can produce JSON output but does not give the model tool descriptions or a selection mechanism — it cannot reliably choose among multiple tools or know their parameter schemas. Option D is wrong because a simple prompt with no tool descriptions gives the model no information about available APIs or their parameters, making correct tool selection and parameter generation essentially impossible.

Full explanation →
33
Multi-Selectmedium

Which TWO techniques are commonly used to handle missing data in a machine learning dataset? (Choose TWO.)

Select 2 answers
A.Normalization
B.Imputation with mean or median
C.Deletion of rows with missing values
D.One-hot encoding
E.Dimensionality reduction
AnswersB, C

Replacing missing values with mean/median is a common imputation method.

Why this answer

Imputation with mean or median is a standard technique for handling missing numerical data because it preserves the dataset size and avoids introducing bias from simply discarding rows. By replacing missing values with the central tendency of the observed data, the model can still learn patterns without losing information, though it may reduce variance slightly.

Exam trap

CompTIA often tests the distinction between data preprocessing techniques (like normalization and encoding) and actual missing data handling methods, so candidates mistakenly select normalization or one-hot encoding as solutions for missing values.

Full explanation →
34
MCQeasy

A data scientist is building a classification model to detect fraudulent transactions. The dataset is highly imbalanced with only 1% fraudulent cases. Which approach should the scientist use to evaluate model performance most effectively?

A.F1 score
B.Accuracy
C.Recall
D.Precision
AnswerA

F1 score suits this imbalanced fraud scenario because it combines precision and recall into a single harmonic mean, so strong performance on the 1% fraudulent minority cannot be masked by the 99% legitimate majority. Accuracy would mislead here, since predicting every transaction as legitimate already yields 99%.

Why this answer

In highly imbalanced datasets like fraud detection (1% positive class), accuracy is misleading because a model that predicts all transactions as legitimate would achieve 99% accuracy yet fail to detect any fraud. The F1 score (harmonic mean of precision and recall) is the most effective metric because it balances both false positives and false negatives, providing a single score that reflects the model's ability to correctly identify the minority class without being skewed by class imbalance.

Exam trap

CompTIA often tests the misconception that accuracy is always the best metric for classification, but in imbalanced datasets, accuracy is a trap because it does not reflect performance on the minority class, leading candidates to overlook metrics like F1 score that directly address class imbalance.

How to eliminate wrong answers

Option B (Accuracy) is wrong because it is dominated by the majority class (99% legitimate transactions), so a trivial model that never predicts fraud can still achieve 99% accuracy, masking poor fraud detection performance. Option C (Recall) is wrong because it only measures the proportion of actual fraud cases correctly identified (true positives / (true positives + false negatives)), ignoring false positives; a model that flags every transaction as fraud would have perfect recall but be unusable in practice. Option D (Precision) is wrong because it only measures the proportion of predicted fraud cases that are actually fraud (true positives / (true positives + false positives)), ignoring false negatives; a model that makes very few fraud predictions but with high precision would miss many actual frauds, which is unacceptable in fraud detection.

Full explanation →
35
MCQeasy

A machine learning engineer wants to prevent unauthorized users from querying a deployed AI model. Which access control measure is MOST appropriate to secure the API?

A.Rate limiting
B.API key authentication
C.Input sanitization
D.IP whitelisting
AnswerB

API key authentication binds each request to a unique credential, so the API gateway rejects unauthenticated callers before they reach the model. This directly satisfies the stem's constraint of preventing unauthorised users from querying the deployed model endpoint.

Why this answer

API key authentication is the most appropriate access control measure because it requires each request to include a unique key that identifies and authorizes the caller. This directly prevents unauthorized users from querying the model by validating the key against a pre-approved list before processing the request. Unlike other options, API keys provide a dedicated authentication layer for API access.

Exam trap

Candidates often confuse rate limiting with access control. Rate limiting only restricts the number of requests, not who can make them. API key authentication is the correct method to ensure only authorized users can query the model.

How to eliminate wrong answers

Option A is wrong because rate limiting controls the frequency of requests, not who can make them; it prevents abuse but does not authenticate users. Option C is wrong because input sanitization protects against injection attacks (e.g., SQLi, XSS) by cleaning user input, but it does not enforce identity verification or access control. Option D is wrong because IP whitelisting restricts access based on source IP addresses, which is brittle (IPs can be spoofed or changed) and does not provide per-user authentication or granular access control.

Full explanation →
36
Multi-Selecthard

A company wants to deploy an LLM-based chatbot that can handle sensitive customer information. Which THREE measures should be implemented to mitigate prompt injection attacks? (Choose 3)

Select 3 answers
A.Use a system prompt that instructs the model to ignore any instructions in the user input
B.Implement output filtering to detect and block harmful responses
C.Sanitize user inputs to remove special characters and escape sequences
D.Use a smaller model with fewer parameters
E.Set temperature to a low value
AnswersA, B, C

A system prompt establishes instruction hierarchy, telling the model to treat user input as data rather than commands. This directly mitigates injection attempts that try to override the chatbot's original behaviour, though it must be combined with other defences.

Why this answer

Option A is correct because a hardened system prompt that explicitly tells the model to treat user input as data and ignore embedded instructions is a primary defense against prompt injection, helping the model distinguish trusted developer instructions from untrusted user content. Option B is correct because output filtering adds a defense-in-depth layer that inspects the model's responses for harmful, leaked, or policy-violating content before it reaches the user, catching injections that bypass input controls. Option C is correct because sanitizing user inputs by stripping or escaping special characters, control tokens, and injection-style delimiters (e.g., markdown fences or role-play markers) reduces the attack surface for crafted prompts.

Option D is not correct because model size does not determine resistance to prompt injection; smaller models can still be manipulated and may even be more susceptible. Option E is not correct because temperature controls randomness in token sampling, not the model's adherence to injected instructions, so lowering it does not mitigate prompt injection.

Exam trap

AI0-001 often tests the misconception that model size or temperature settings affect security, when in fact prompt injection mitigation requires input/output controls and system-level instructions.

Full explanation →
37
MCQmedium

A machine learning team is deploying a sentiment analysis model for customer reviews. The model was trained on reviews from an e-commerce site but will be used for a social media platform. The team observes a drop in accuracy. Which concept best explains this issue?

A.Data drift
B.Concept drift
C.Bias-variance tradeoff
D.Overfitting
AnswerA

The model encounters social media text whose vocabulary, length and style differ from the e-commerce reviews it was trained on, so the input distribution shifts between training and deployment. This covariate shift is data drift, explaining the accuracy drop without any change in the underlying sentiment-label relationship.

Why this answer

Data drift occurs when the statistical properties of the input data change between the training and production environments. Here, the model was trained on e-commerce reviews but is now processing social media posts, which have different vocabulary, tone, and structure, causing a mismatch in the input distribution and leading to accuracy degradation.

Exam trap

CompTIA often tests the distinction between data drift (input distribution change) and concept drift (relationship change), and candidates mistakenly choose concept drift when the scenario describes a change in the input data source rather than a change in the underlying mapping from inputs to outputs.

How to eliminate wrong answers

Option B is wrong because concept drift refers to a change in the underlying relationship between input features and the target variable over time, not a change in the input data distribution itself. Option C is wrong because bias-variance tradeoff is a model selection concept describing the balance between underfitting and overfitting, not an explanation for performance drop due to data distribution shift. Option D is wrong because overfitting occurs when a model learns training data too well, including noise, and fails to generalize to new data from the same distribution, not to a different distribution.

Full explanation →
38
MCQhard

An AI team is deploying a predictive maintenance model for industrial equipment. The model predicts failure within a 30-day window. The cost of a false positive is 10% of the cost of a false negative. Which evaluation metric should the team prioritize?

A.F2 score (beta=2) to prioritize recall over precision.
B.Area under the ROC curve (AUC-ROC) to measure overall discrimination.
C.F1 score to balance precision and recall equally.
D.Precision to minimize false positives.
AnswerA

F2 score puts more weight on recall, aligning with the higher cost of false negatives.

Why this answer

The F2 score (beta=2) weights recall four times more than precision, which is appropriate because a false negative (missing a failure) costs 10 times more than a false positive (unnecessary maintenance). Prioritizing recall ensures the model captures as many true failures as possible, minimizing the higher-cost error type.

Exam trap

The trap here is that candidates may default to F1 score as a 'balanced' metric without considering the asymmetric cost structure, or they may incorrectly think AUC-ROC captures cost-sensitive performance.

How to eliminate wrong answers

Option B is wrong because AUC-ROC measures overall discrimination across all thresholds and does not account for the asymmetric cost structure between false positives and false negatives. Option C is wrong because the F1 score balances precision and recall equally, which is suboptimal when the cost of a false negative is 10 times higher than a false positive. Option D is wrong because minimizing false positives (maximizing precision) would increase false negatives, leading to higher overall cost due to the 10:1 cost ratio.

Full explanation →
39
MCQmedium

An organization wants to detect if someone is trying to steal their proprietary machine learning model by querying its API. Which monitoring technique is MOST effective?

A.Output filtering to remove sensitive information from responses
B.Rate limiting on the number of API requests per user
C.Monitoring for anomalous query patterns, such as high volume or systematic variations
D.Input validation to reject malformed requests
AnswerC

Monitoring anomalous query patterns detects model extraction, where attackers probe an API with systematic input variations to reconstruct decision boundaries. High-volume or structured querying satisfies the scenario's requirement to identify theft attempts against the proprietary model, since legitimate users rarely exhibit such repetitive, exhaustive probing behaviour.

Why this answer

Model extraction attacks rely on systematically querying the API to reconstruct the model's decision boundary. Monitoring for anomalous query patterns—such as high request volume, uniform input distributions, or systematic variations (e.g., grid-like sampling of feature space)—directly detects the behavioral signature of extraction attempts, unlike passive controls that do not address the attack vector.

Exam trap

The trap here is that candidates confuse generic security controls (rate limiting, input validation) with the specific detection technique needed for model extraction, overlooking that extraction attacks use legitimate, well-formed queries in a systematic pattern.

How to eliminate wrong answers

Option A is wrong because output filtering removes sensitive information from responses but does not prevent an attacker from collecting enough outputs to reconstruct the model; it only obscures specific data points. Option B is wrong because rate limiting reduces request throughput but does not detect or prevent extraction via low-and-slow queries or distributed attacks; it can be bypassed by using multiple IPs or accounts. Option D is wrong because input validation rejects malformed requests but extraction attacks use well-formed, legitimate queries to probe the model; validation does not flag the systematic, high-volume patterns indicative of extraction.

Full explanation →
40
Multi-Selectmedium

A hospital has deployed an AI triage assistant that summarizes patient intake notes and suggests an acuity level for the emergency department. Clinicians report that the assistant sometimes produces confident but unsupported acuity suggestions. The operations team must add safeguards appropriate for a high-stakes clinical deployment. (Choose two.)

Select 2 answers
A.Display the source note spans that the model used for each acuity suggestion so clinicians can verify the evidence before accepting it.
B.Retrain the assistant on the hospital's historical triage outcomes and redeploy it as the primary acuity assigner.
C.Lower the model's temperature and top-p sampling values so its generated summaries are more deterministic.
D.Increase the model's context window so it can ingest the patient's entire longitudinal chart in one request.
E.Require a clinician to explicitly confirm or override every suggested acuity level before it is written to the patient record.
AnswersA, E

Grounding each suggestion in the specific note spans it used gives the clinician a fast way to confirm or reject the recommendation. In a high-stakes setting this turns an opaque output into an auditable claim, and it lets the reviewer notice when the model leaned on an irrelevant or misread passage, which is exactly the failure mode described.

Why this answer

The reported failure is confident but unsupported suggestions, so the safeguards must both expose the evidence behind each suggestion and keep a clinician as the accountable decision-maker. Attribution to source note spans makes the claim checkable, and mandatory confirmation or override prevents an unverified acuity level from entering the patient record.

Exam trap

The trap here is treating a high-stakes clinical hallucination problem as a prompt-tuning or model-size problem, when the effective controls are evidence attribution and a human decision checkpoint.

Full explanation →
41
MCQmedium

Refer to the exhibit. A machine learning pipeline configuration is shown. During a deployment, the model evaluation passes with accuracy 0.86 and precision 0.79. However, the pipeline proceeds to deploy. What is the most likely reason for this behavior?

A.The precision metric is not included in the evaluation script
B.The deployment only checks the accuracy threshold for rollback condition
C.The deployment target is set to staging instead of production
D.The operator manually overrode the threshold
AnswerB

The pipeline gates deployment solely on the accuracy threshold, so a 0.86 accuracy score passes even though precision of 0.79 falls below its own gate. The rollback condition never evaluates precision, allowing deployment despite the weaker metric.

Why this answer

The pipeline configuration shows a rollback condition that only checks the accuracy metric (accuracy < 0.85). Since the model achieved accuracy 0.86, which is above the threshold, the condition is not triggered, and the pipeline proceeds to deploy regardless of the precision value. The precision metric is not part of the rollback evaluation logic in this configuration.

Exam trap

CompTIA often tests the misconception that all evaluation metrics automatically trigger rollback conditions, when in fact only metrics explicitly listed in the condition logic are checked.

How to eliminate wrong answers

Option A is wrong because the evaluation script clearly outputs precision (0.79), and the exhibit shows precision is being calculated; the issue is that the rollback condition does not reference precision. Option C is wrong because the deployment target (staging vs. production) does not affect whether a rollback condition is evaluated; the pipeline proceeds based on the condition logic, not the environment name. Option D is wrong because there is no evidence or indication in the exhibit or scenario that an operator manually overrode the threshold; the behavior is fully explained by the configured rollback condition.

Full explanation →
42
Multi-Selecthard

A company is deploying an AI model that processes financial transactions. They want to implement privacy-preserving machine learning. Which THREE techniques achieve this goal? (Select three.)

Select 3 answers
A.Model pruning
B.Differential privacy
C.Data augmentation
D.Homomorphic encryption
E.Federated learning
AnswersB, D, E

Differential privacy adds calibrated noise to computations or training data so that any single individual's contribution cannot be inferred from outputs. This provides a mathematical privacy guarantee for the financial transaction data, directly achieving the privacy-preserving machine learning objective.

Why this answer

Differential privacy (B) is correct because it adds calibrated noise (e.g., via the Laplace or Gaussian mechanism) to computations or gradients so that any single individual's transaction data has a bounded influence on the model output, providing a formal privacy guarantee. Homomorphic encryption (D) is correct because it allows computations to be performed directly on encrypted financial data (e.g., using schemes like Paillier, BFV, or CKKS), so the model can train or infer without ever decrypting sensitive values. Federated learning (E) is correct because it keeps raw transaction data on local devices or silos and only shares model updates (often combined with secure aggregation or differential privacy), minimizing centralized exposure of private records.

Model pruning (A) merely removes redundant weights to reduce model size and compute, and data augmentation (C) synthetically expands training data for robustness; neither provides a privacy guarantee, so they do not belong.

Exam trap

CompTIA often tests the distinction between techniques that improve model performance (pruning, augmentation) versus those that actively protect data privacy (differential privacy, encryption, federated learning), so candidates mistakenly select performance-enhancing options as privacy-preserving ones.

Full explanation →
43
Multi-Selectmedium

An organisation is developing a document intelligence system that extracts information from scanned invoices. Which THREE data preparation steps are critical to ensure high extraction accuracy? (Choose THREE.)

Select 3 answers
A.Cleaning and correcting OCR output
B.Removing punctuations and stopwords
C.Normalising all text to lowercase
D.Annotating bounding boxes and field labels
E.Image preprocessing (e.g., deskewing, binarisation)
AnswersA, D, E

Cleaning and correcting OCR output directly addresses the scanned-invoice constraint: OCR introduces character errors on low-quality scans, and those errors propagate into extraction models. Normalising recognised text before training or inference raises accuracy, since the system's inputs are images rather than typed digital text.

Why this answer

Option A (Cleaning and correcting OCR output) is correct because OCR on scanned invoices inevitably introduces character-level errors, and fixing those errors before feeding text to the extraction model directly improves field-level accuracy. Option D (Annotating bounding boxes and field labels) is correct because supervised document intelligence models need labelled ground truth that ties specific fields (e.g., invoice number, total) to their spatial locations to learn accurate extraction. Option E (Image preprocessing such as deskewing and binarisation) is correct because scanned invoices often suffer from rotation, noise, and uneven lighting, and correcting these at the image level raises OCR quality and downstream extraction accuracy.

Option B (Removing punctuations and stopwords) is not appropriate because invoice fields such as dates, currency amounts, and vendor names rely on punctuation and specific tokens, so removing them would destroy critical information. Option C (Normalising all text to lowercase) is not appropriate because case can carry meaning in invoice data (e.g., currency codes, product identifiers, proper names), and lowercasing everything can reduce extraction fidelity.

Exam trap

AI0-001 often tests the distinction between general NLP preprocessing steps (like stopword removal and lowercasing) and document-specific preprocessing (like OCR correction and image enhancement), causing candidates to incorrectly select B or C as critical for extraction accuracy.

Full explanation →
44
MCQhard

A machine learning team is developing a model to predict loan defaults using sensitive customer financial data. They need to share the model with third-party auditors without exposing individual customer records. Which privacy-preserving technique allows auditors to query the model while providing mathematical guarantees about the privacy of the training data?

A.Differential privacy
B.Federated learning
C.k-anonymity
D.Homomorphic encryption
AnswerA

Differential privacy provides a formal epsilon guarantee bounding how much any single customer record changes query outputs, letting auditors query the model without exposing individual records. This satisfies the demand for mathematical privacy guarantees on the training data.

Why this answer

Differential privacy is correct because it adds calibrated noise to the model's training process or query responses, providing a formal mathematical guarantee (ε-differential privacy) that the inclusion or exclusion of any single individual's data does not significantly affect the output. This allows auditors to query the model without exposing individual customer records, as the noise bounds the information leakage from the training data.

Exam trap

A common misconception is that federated learning inherently provides privacy guarantees, when in fact it only addresses data locality and does not prevent model inversion or membership inference attacks without additional differential privacy mechanisms.

How to eliminate wrong answers

Option B (Federated learning) is wrong because it is a distributed training technique that keeps raw data on local devices and shares only model updates, but it does not provide mathematical privacy guarantees for the training data against inference attacks from the shared updates. Option C (k-anonymity) is wrong because it is a data anonymization technique that generalizes or suppresses attributes to ensure each record is indistinguishable from at least k-1 others, but it does not provide a formal mathematical guarantee against membership inference or attribute disclosure when the model is queried. Option D (Homomorphic encryption) is wrong because it allows computations on encrypted data, protecting data in transit and at rest, but it does not prevent the model from leaking training data through its outputs when queried, and it does not provide a mathematical privacy guarantee for the training data against the auditor.

Full explanation →
45
MCQmedium

An organization's AI system uses a decision tree model for loan approval. The compliance team requires explanations for each decision. Which property of decision trees makes them suitable for this requirement?

A.They can handle nonlinear relationships
B.They are robust to outliers
C.The decision rules are transparent and can be visualized as a tree
D.They can handle missing values
AnswerC

Decision trees expose their learned logic as a hierarchy of if-then splits on individual features, so each loan decision can be traced to explicit, human-readable conditions. This inherent interpretability satisfies the compliance requirement for per-decision explanations, unlike opaque ensembles or neural networks needing post-hoc surrogates.

Why this answer

Decision trees inherently provide interpretable decision rules by splitting data based on feature thresholds at each node. The entire model can be visualized as a tree structure, allowing compliance teams to trace the exact path and logic behind each loan approval or rejection, which directly satisfies explainability requirements.

Exam trap

CompTIA often tests the distinction between model performance properties (e.g., handling nonlinearity, robustness) and interpretability properties, leading candidates to select a technically true but irrelevant advantage instead of the one that directly satisfies the compliance requirement.

How to eliminate wrong answers

Option A is wrong because handling nonlinear relationships is a general capability of many models (e.g., neural networks, SVMs with kernels) and is not unique to decision trees, nor does it directly address the need for transparent explanations. Option B is wrong because decision trees are not inherently robust to outliers; in fact, they can be sensitive to outliers that cause splits to be skewed, and robustness is not related to explainability. Option D is wrong while decision trees can handle missing values through surrogate splits or other imputation methods, this property does not provide the transparency or traceability required for compliance explanations.

Full explanation →
46
MCQhard

A retail company deploys a machine learning model to predict customer churn. The model outputs a probability between 0 and 1, and churn is predicted if probability > 0.5. After deployment, the model has a high false positive rate (many non-churning customers labeled as churn), which leads to unnecessary retention offers and increased costs. The data science team confirms the model was trained on historical data with a balanced class distribution. The business team wants to reduce false positives while maintaining a reasonable true positive rate. However, they cannot retrain the model because the original training data is no longer available. What is the best course of action to reduce false positives?

A.Retrain the model using only the most recent three months of data.
B.Increase the decision threshold to a higher value, such as 0.7.
C.Collect new labeled data and perform transfer learning from the original model.
D.Decrease the decision threshold to a lower value, such as 0.3.
AnswerB

Raising the threshold above 0.5 requires stronger churn evidence before labelling a customer positive, directly reducing false positives. It needs no retraining, so it works despite the unavailable training data, though some true positives are lost.

Why this answer

Increasing the decision threshold to a higher value, such as 0.7, reduces false positives because the model will only predict churn when it is more confident. Since the model cannot be retrained, adjusting the threshold is the only way to trade off between precision and recall without modifying the model itself.

Exam trap

CompTIA often tests the misconception that retraining or collecting more data is the only way to fix model performance issues, when in fact threshold tuning is a valid post-deployment technique that does not require retraining.

How to eliminate wrong answers

Option A is wrong because retraining is not possible as the original training data is no longer available, and using only the most recent three months of data would likely introduce data drift and require retraining resources. Option C is wrong because transfer learning typically requires access to the original model architecture and training data, and collecting new labeled data does not directly reduce false positives without retraining or threshold adjustment. Option D is wrong because decreasing the threshold to 0.3 would increase the false positive rate, making the problem worse, not better.

Full explanation →
47
MCQmedium

A machine learning engineer is building a recommendation system for an e-commerce platform. The system should suggest products based on user purchase history and browsing behavior. Which model selection is BEST suited for this task?

A.Image classification model (e.g., CNN)
B.Linear regression
C.Random forest classifier
D.Collaborative filtering model (e.g., matrix factorization)
AnswerD

Collaborative filtering exploits the interaction matrix between users and items, learning latent factors from purchase and browsing history to predict unseen preferences. This directly matches the scenario's reliance on behavioural signals rather than item content, making matrix factorisation the best-suited approach for personalised product suggestions.

Why this answer

Collaborative filtering models (e.g., matrix factorization) are effective for recommendation tasks using user-item interaction data. Linear regression is for regression, not recommendation. Image classification is unrelated.

Random forests can be used but are less common for collaborative filtering.

Full explanation →
48
MCQmedium

An image classification model misclassifies a stop sign as a speed limit sign after a few pixels are altered. What is the most effective defense against such attacks?

A.Use a larger validation dataset
B.Reduce the input image resolution
C.Increase the model's complexity
D.Adversarial training
AnswerD

Adversarial training augments the training set with perturbed examples, such as stop signs with altered pixels, so the model learns to classify them correctly. This directly hardens the decision boundary against the small, deliberate pixel-level perturbations described in the stem, unlike input sanitisation, which cannot anticipate every crafted variant.

Why this answer

Adversarial training is the most effective defense because it explicitly incorporates adversarial examples—like the perturbed stop sign—into the model's training data. By training on both clean and adversarially altered images, the model learns to be robust against small, malicious perturbations that cause misclassification. This directly addresses the root cause of the vulnerability, unlike other options that only mitigate symptoms or ignore the attack vector.

Exam trap

The AI0-001 exam often tests the misconception that increasing dataset size or model complexity improves security, when in fact adversarial training is the only listed option that directly hardens the model against input perturbations.

How to eliminate wrong answers

Option A is wrong because a larger validation dataset does not protect against adversarial perturbations; it only improves the statistical estimate of model performance on clean data, not robustness to crafted attacks. Option B is wrong because reducing input resolution may actually increase vulnerability by discarding fine-grained features that help distinguish objects, and it does not prevent pixel-level manipulations from fooling the model. Option C is wrong because increasing model complexity often makes the model more susceptible to overfitting and adversarial examples, as deeper networks can have larger linear regions that attackers exploit.

Full explanation →
49
MCQeasy

A company is considering using an open-source large language model for a commercial application. Which intellectual property consideration is MOST important when deciding between open-source and proprietary models?

A.The model's license terms and any restrictions on commercial use
B.The model's accuracy on benchmark tasks
C.The size of the model's parameter count
D.The model's training data provenance
AnswerA

Open-source licences vary widely: some permit unrestricted commercial deployment, while others impose copyleft, attribution, or usage caps that could block a commercial product. Reviewing the specific licence terms is therefore the decisive intellectual property check when weighing open-source against proprietary models.

Why this answer

The license terms of an open-source model dictate whether and how it can be used commercially, modified, or redistributed. Some licenses (e.g., Apache 2.0, MIT) are permissive, while others (e.g., GPL, AGPL, or custom licenses like Llama 2's) impose restrictions such as copyleft, attribution, or limits on commercial use. For a commercial application, failing to comply with these terms can lead to legal liability, making license review the most critical IP consideration.

Exam trap

AI0-001 often tests the distinction between technical performance metrics and legal/IP considerations, causing candidates to overlook license terms in favor of accuracy or model size.

How to eliminate wrong answers

Option B is wrong because accuracy on benchmarks is a performance metric, not an intellectual property concern; it does not address legal rights to use the model. Option C is wrong because parameter count is a technical specification indicating model size and capacity, not a legal or IP restriction. Option D is wrong because while training data provenance can raise IP issues (e.g., copyright infringement), it is a secondary consideration; the model's license explicitly governs the terms of use and is the primary IP factor for commercial adoption.

Full explanation →
50
MCQmedium

A company deploys a chatbot using a large language model (LLM). After launch, users report that the chatbot sometimes generates plausible but false information. This phenomenon is known as:

A.Gradient explosion
B.Overfitting
C.Concept drift
D.Hallucination
AnswerD

Hallucination describes an LLM producing fluent, confident output that is factually wrong because the model predicts plausible token sequences rather than retrieving verified facts. The stem's "plausible but false" wording maps directly onto this term, not bias, overfitting or latency.

Why this answer

Hallucination in LLMs refers to the generation of plausible but factually incorrect or nonsensical information. This occurs when the model's probabilistic next-token prediction produces confident-sounding outputs that deviate from training data or real-world facts, often due to insufficient grounding or training data gaps.

Exam trap

The trap here is that candidates may confuse hallucination with overfitting, thinking the model is 'making up' data due to memorization errors, but overfitting is about poor generalization to new inputs, not confident false outputs from a well-generalized model.

How to eliminate wrong answers

Option A is wrong because gradient explosion is a training instability issue in deep neural networks where gradients become excessively large, causing weight updates to diverge; it does not relate to post-deployment output inaccuracies. Option B is wrong because overfitting describes a model that memorizes training data too well, performing poorly on unseen data, not generating false information that seems plausible. Option C is wrong because concept drift refers to a change in the statistical properties of the target variable over time, requiring model retraining, not a static LLM generating false outputs.

Full explanation →
51
Multi-Selectmedium

A startup is building a retrieval-augmented generation (RAG) application that must answer questions over a 500,000-document internal knowledge base with low query latency. They plan to use a vector database. Which TWO design choices best support fast, scalable similarity search? (Choose two.)

Select 2 answers
A.Store the full text of each document inside the vector database and run keyword search alongside vector search.
B.Choose embeddings with a smaller dimensionality that still preserve semantic quality for the domain.
C.Recompute embeddings for all documents on every query to ensure freshness.
D.Use an approximate nearest neighbor (ANN) index such as HNSW instead of exact brute-force search.
E.Disable metadata filtering so the index can scan all vectors uniformly.
AnswersB, D

Vector search cost scales with dimensionality because distance computations and index memory grow accordingly. Selecting a lower-dimensional embedding that retains domain-relevant semantics reduces memory footprint and speeds up both index construction and query time. This is a legitimate way to support fast, scalable similarity search without abandoning semantic fidelity.

Why this answer

ANN indexing with HNSW and lower-dimensional embeddings both reduce the computational cost of similarity search on a large corpus. HNSW avoids exhaustive comparison, while reduced dimensionality shrinks distance calculations and memory use. Together they enable low-latency retrieval at the 500,000-document scale without sacrificing meaningful semantic accuracy.

Exam trap

The trap here is assuming that adding more features, such as keyword search or full-text storage, automatically improves performance, when the question is about latency and scalability of similarity search.

Full explanation →
52
MCQmedium

An e-commerce company uses a gradient boosting model to forecast daily sales. Recently, the model's predictions have become less accurate, showing a significant drop in R-squared on validation data. The data scientist checks for data drift but finds no significant changes in feature distributions. The model was trained on data from the past 24 months and is retrained monthly. Upon inspecting the feature importance, the data scientist notices that the top feature 'promotion_flag' has decreased in importance over time. What is the most likely cause of the performance degradation, and what should be done?

A.The model is overfitting to historical promotions; apply more regularization
B.Concept drift has occurred; retrain the model more frequently with recent data only, or use an online learning approach
C.The model's hyperparameters need tuning; perform a grid search
D.The promotion_flag feature is leaking future information; remove it
AnswerB

Stable feature distributions but a shifting feature-target relationship indicate concept drift, not data drift. The declining importance of promotion_flag shows the old mapping no longer holds, so retraining on recent data or adopting online learning restores accuracy.

Why this answer

The model's performance degradation, despite no data drift in feature distributions, is likely due to concept drift—the relationship between features and the target variable has changed over time. The decreasing importance of 'promotion_flag' suggests that promotions no longer influence sales as they once did. Retraining more frequently with recent data or using online learning can help the model adapt to the new concept.

Exam trap

The trap is that candidates may confuse concept drift with data drift or overfitting, but the key clue is the change in feature importance over time without changes in feature distributions, pointing to concept drift rather than other issues.

How to eliminate wrong answers

Option A is wrong because overfitting would typically show good performance on training data but poor on validation; here, the issue is a drop in validation R-squared over time, not necessarily overfitting. Option C is wrong because hyperparameter tuning would not address a change in the underlying data relationship; it might yield marginal improvements but not solve concept drift. Option D is wrong because feature leakage would cause overly optimistic performance during training, not a gradual degradation; also, 'promotion_flag' is a legitimate feature, and removing it without evidence of leakage is not appropriate.

Full explanation →
53
Multi-Selecthard

An insurance company operates an AI claims-triage model that flags suspicious claims for human review. After six months in production, the operations team observes that the model's precision has fallen steadily while recall has stayed roughly constant, and the volume of false-positive flags has grown. The data science team suspects the input data pipeline is the cause rather than the model weights. Which TWO operational checks should the team perform first to diagnose the problem? (Choose two.)

Select 2 answers
A.Increase the model's decision threshold so fewer claims are flagged as suspicious.
B.Audit the feature engineering and ETL pipeline for silent failures such as nulls, default values, or unit changes that alter feature semantics.
C.Compare the distribution of each production input feature against the training baseline to detect upstream data drift or schema changes.
D.Retrain the model immediately on the last six months of production data to restore precision.
E.Expand the human review team so more flagged claims can be examined manually while the model remains unchanged.
AnswersB, C

A pipeline defect that substitutes nulls, defaults, or differently scaled values changes the meaning of features without changing the model, which degrades precision while recall holds. Auditing the ETL and feature engineering stages for silent failures directly tests the team's hypothesis that the input data pipeline, not the model weights, is the root cause.

Why this answer

Stable recall with falling precision points to inputs that no longer match training conditions rather than to the model weights. Comparing production feature distributions against the training baseline detects drift or schema changes, and auditing the ETL and feature engineering stages uncovers silent failures such as nulls, defaults, or unit changes. Retraining, threshold changes, or more reviewers treat symptoms without identifying the pipeline defect.

Exam trap

The trap here is responding to a precision drop by retraining or thresholding the model, when the evidence points to upstream data corruption that must be diagnosed before any model change.

Full explanation →
54
MCQeasy

A social media company's AI recommendation system pushes extreme content to users, causing harm. Which ethical principle is most violated?

A.Autonomy
B.Justice
C.Beneficence
D.Non-maleficence
AnswerD

Non-maleficence obliges developers to avoid causing harm, and the recommender's amplification of extreme content inflicts direct user harm. This principle is violated more precisely than beneficence, which concerns actively promoting wellbeing rather than refraining from damage.

Why this answer

Non-maleficence (do no harm) is the principle most directly violated because the AI system actively causes harm by pushing extreme content that damages users' mental health or incites harmful behavior. Unlike beneficence (doing good), non-maleficence focuses on avoiding harm, and the system's design fails to prevent foreseeable negative outcomes.

Exam trap

CompTIA often tests the distinction between beneficence and non-maleficence, where candidates mistakenly choose beneficence because they think the system failed to do good, but the actual violation is causing direct harm.

How to eliminate wrong answers

Option A is wrong because autonomy concerns user self-determination and informed consent, not the direct harm from content amplification. Option B is wrong because justice relates to fairness and equitable treatment across user groups, not the specific harm caused by extreme content. Option C is wrong because beneficence requires actively doing good, whereas the core violation here is causing harm, not failing to provide a benefit.

Full explanation →
55
Multi-Selectmedium

A financial institution uses a machine learning model to approve loans. They want to protect against membership inference attacks. Which THREE techniques are effective?

Select 3 answers
A.Applying model truncation or output perturbation
B.Training with differential privacy
C.Limiting the granularity of model outputs (e.g., returning scores instead of probabilities)
D.Implementing federated learning
E.Using shadow models to distract attackers
AnswersA, B, C

Truncating outputs or perturbing returned values reduces the confidence information an adversary needs to infer whether a specific record was in the training set, satisfying the requirement to blunt membership inference against the loan model.

Why this answer

Option A (model truncation or output perturbation) is correct because reducing the precision or adding calibrated noise to the model's outputs limits the information an attacker can extract about whether a specific record was in the training set, directly mitigating membership inference. Option B (training with differential privacy) is correct because DP-SGD and related mechanisms provide a formal guarantee that the inclusion or exclusion of any single training record has a bounded effect on the model's behavior, which is the canonical defense against membership inference. Option C (limiting the granularity of model outputs, e.g., returning scores instead of probabilities) is correct because coarse, bucketed outputs reduce the signal an attacker can use to distinguish members from non-members, lowering attack success rates.

Option D (federated learning) is not inherently a membership-inference defense: it keeps raw data local but the shared model updates can still leak membership information, so it does not by itself provide the required protection. Option E (shadow models to distract attackers) is not a recognized defense; shadow models are an attacker technique used to train attack classifiers, not a mitigation, so it does not belong here.

Exam trap

AI0-001 often tests the difference between techniques that directly mitigate membership inference (differential privacy, output perturbation) and those that are unrelated or even detrimental (federated learning alone, shadow models). Candidates may confuse federated learning as a privacy panacea.

Full explanation →
56
MCQmedium

A retail analytics team is preparing a dataset of product reviews for a sentiment classification model. The dataset contains 50,000 reviews, but only 2,000 are labeled as positive or negative. The team wants to use the unlabeled reviews to improve model performance. Which approach best leverages the unlabeled data?

A.Apply principal component analysis (PCA) to reduce the dimensionality of the unlabeled reviews and then train a classifier on the reduced features.
B.Apply semi-supervised learning using a self-training algorithm that iteratively labels high-confidence unlabeled examples and retrains the model.
C.Perform data augmentation by generating synthetic reviews using a generative adversarial network (GAN) trained on the labeled data.
D.Use transfer learning by fine-tuning a pre-trained language model on the 2,000 labeled reviews only, ignoring the unlabeled data.
AnswerB

Self-training is a semi-supervised technique that uses a small labeled set to train an initial model, then predicts labels for unlabeled data, adds the most confident predictions to the training set, and repeats. This directly uses the 48,000 unlabeled reviews to improve the model without manual labeling, making it the best fit for the scenario.

Why this answer

Semi-supervised learning, specifically self-training, is designed to exploit a small labeled set alongside a large unlabeled set. By iteratively adding high-confidence pseudo-labels from the unlabeled reviews, the model can learn from the broader data distribution, improving generalization. The other options either ignore the unlabeled data or use techniques that are not appropriate for text sentiment classification with limited labels.

Exam trap

The trap here is assuming that any technique using unlabeled data, such as PCA, is sufficient, when in fact only semi-supervised methods directly incorporate unlabeled data into the training process to improve classification.

Full explanation →
57
Multi-Selectmedium

A data science team is preparing a dataset of customer support tickets to train a supervised model that routes each ticket to the correct department. They have 40,000 tickets labeled with one of eight departments. Which TWO preprocessing steps are most appropriate before training? (Choose two.)

Select 2 answers
A.Split the data into training, validation, and test sets before fitting any preprocessing.
B.Delete all tickets shorter than 20 words to reduce noise.
C.Standardize the department labels to zero mean and unit variance.
D.Tokenize the ticket text and convert it to numerical vectors using an embedding or TF-IDF representation.
E.Apply one-hot encoding to the raw ticket text strings.
AnswersA, D

Holding out validation and test data before fitting vectorizers or scalers prevents information leakage from the evaluation sets into training. If TF-IDF vocabulary or normalization statistics are learned on the full dataset, reported performance becomes optimistic and unreliable. This step ensures the routing model is evaluated on genuinely unseen tickets.

Why this answer

Text classification requires converting raw strings into numeric features and preserving an honest evaluation split. Tokenization with TF-IDF or embeddings supplies the numerical representation, while splitting before fitting preprocessing prevents leakage. Together these steps prepare the ticket data so the routing model can be trained and fairly assessed on unseen examples.

Exam trap

The trap here is fitting the vectorizer on the entire dataset before splitting, which leaks vocabulary and statistics from the test set into training and inflates reported accuracy.

Full explanation →
58
MCQhard

A company uses a large language model (LLM) to generate customer support responses. They notice the model sometimes produces harmful outputs. Which implementation strategy best reduces this risk while maintaining performance?

A.Implement a keyword-based output filter
B.Use a smaller, less capable model
C.Add system prompts instructing the model to be safe
D.Fine-tune the model using reinforcement learning from human feedback
AnswerD

Reinforcement learning from human feedback trains a reward model on human preference rankings, then optimises the LLM against it, directly suppressing harmful generations while preserving fluency and task performance. This targets the harmful-output risk at the alignment layer rather than filtering responses after generation.

Why this answer

Reinforcement learning from human feedback (RLHF) directly trains the model to align its outputs with human preferences for safety and helpfulness, reducing harmful outputs while preserving performance. Unlike superficial filters or prompts, RLHF adjusts the model's internal behavior through reward modeling and policy optimization, making it the most effective strategy for sustained safety improvements.

Exam trap

CompTIA often tests the misconception that simple output filtering or prompt engineering is sufficient for safety, when in fact only training-based alignment methods like RLHF can meaningfully change model behavior without sacrificing performance.

How to eliminate wrong answers

Option A is wrong because keyword-based output filters are brittle and can be bypassed by paraphrasing or context-dependent harmful content, while also risking false positives that degrade performance by blocking legitimate responses. Option B is wrong because using a smaller, less capable model reduces overall performance and may still produce harmful outputs if not specifically trained for safety, as capability and safety are not directly correlated. Option C is wrong because system prompts are easily overridden by the model's training distribution and do not provide robust, consistent safety alignment, especially against adversarial or nuanced harmful inputs.

Full explanation →
59
Multi-Selectmedium

Which TWO are key requirements for AI governance under the EU AI Act for high-risk AI systems? (Choose two.)

Select 2 answers
A.Regular performance benchmarks
B.Human oversight
C.Open-source licensing
D.Transparency and documentation
E.Mandatory use of cloud
AnswersB, D

Human oversight requires that natural persons monitor high-risk systems, intervene or halt operation, and review outputs. This satisfies the EU AI Act's governance requirement by ensuring meaningful human control over consequential automated decisions, mitigating risks to health, safety and fundamental rights.

Why this answer

Option B (Human oversight) is correct because the EU AI Act explicitly requires that high-risk AI systems be designed and developed with appropriate human oversight mechanisms, allowing humans to effectively monitor, intervene, and override the system to prevent or minimize risks to health, safety, and fundamental rights. Option D (Transparency and documentation) is correct because high-risk AI systems must be accompanied by technical documentation, instructions for use, and logging capabilities that ensure traceability and enable users and authorities to understand the system's functioning and compliance. Options A, C, and E are not key requirements under the Act: regular performance benchmarks are not a mandated governance requirement per se, open-source licensing is not a condition for high-risk compliance (though open-source models have some accommodations), and there is no mandatory use of cloud infrastructure for high-risk AI systems.

Exam trap

The AI0-001 exam often tests the distinction between general best practices (like performance benchmarks) and specific regulatory mandates (like human oversight and transparency), leading candidates to select familiar but non-required options such as regular performance benchmarks.

Full explanation →
60
MCQeasy

A company deploys a computer vision model for quality inspection on a manufacturing line. After deployment, the model's accuracy drops from 95% to 80% over two weeks. Which action is most likely to address this issue?

A.Retrain the model using recently collected production data.
B.Increase the confidence threshold for predictions.
C.Decrease the learning rate of the training algorithm.
D.Deploy an additional ensemble of models for redundancy.
AnswerA

A gradual accuracy decline over two weeks on a production line indicates data drift, such as new lighting, camera angles or component variants. Retraining on recently collected production data aligns the model with the current input distribution, directly restoring the lost accuracy.

Why this answer

The accuracy drop over two weeks indicates data drift or concept drift, where the production data distribution changes over time. Retraining the model with recently collected production data realigns it with the current data distribution, directly addressing the drift. Option B (increasing confidence threshold) may reduce false positives but does not fix the underlying drift and could lower recall.

Option C (decreasing learning rate) is irrelevant for inference; it only affects training and cannot be applied post-deployment to fix drift. Option D (deploying an ensemble) adds computational overhead and does not resolve drift; it might even mask the issue without correcting it.

Full explanation →
61
MCQeasy

A data scientist is choosing a hardware accelerator for training a large transformer model. Which of the following is specifically designed for deep learning workloads and offers the highest throughput for matrix multiplications?

A.TPU
B.GPU
C.NPU
D.CPU
AnswerA

TPUs are Google's custom ASICs built specifically for tensor computations, delivering the highest throughput for matrix multiplications in deep learning.

Why this answer

The TPU (Tensor Processing Unit) is an application-specific integrated circuit (ASIC) designed by Google specifically to accelerate deep learning workloads. Its systolic array architecture is optimized for the matrix multiplications and convolutions that dominate transformer model training, delivering the highest throughput among the listed options for these operations.

Exam trap

This question tests the distinction between hardware designed for training versus inference. The trap is that candidates may choose GPU because it is the most common deep learning accelerator, overlooking that TPU is purpose-built for the highest matrix multiplication throughput in training workloads.

How to eliminate wrong answers

Option B (GPU) is wrong because while GPUs are widely used for deep learning and offer high parallelism, they are general-purpose processors originally designed for graphics rendering, not specifically optimized for the dense matrix operations in transformer training. Option C (NPU) is wrong because Neural Processing Units are typically designed for low-power inference on edge devices, not for high-throughput training of large models. Option D (CPU) is wrong because CPUs are general-purpose processors optimized for sequential tasks and low-latency operations, lacking the massive parallel compute units and specialized matrix multiplication hardware needed for efficient transformer training.

Full explanation →
62
MCQmedium

A machine learning engineer wants to prevent data poisoning during the training of a model. Which practice is MOST effective for ensuring the integrity of the training data?

A.Differential privacy
B.Secure data pipelines
C.Red teaming the model
D.Output filtering
AnswerB

Securing data pipelines directly enforces integrity across ingestion, transformation and storage, blocking tampering or injection before poisoned samples reach training. This satisfies the stem's data-poisoning constraint by applying authentication, encryption and validation controls at each transfer stage, so unauthorised modification is prevented rather than merely detected after the model has already learned corrupted patterns.

Why this answer

Secure data pipelines enforce integrity controls — provenance tracking, access control, encryption, and validation — across the entire data ingestion and preprocessing flow, which directly prevents adversaries from injecting poisoned samples. Because poisoning attacks occur before or during training, protecting the pipeline is the most effective defense. Differential privacy, red teaming, and output filtering address different threat surfaces.

Exam trap

AI0-001 often tests the distinction between preventive controls (secure pipelines) and detective/mitigative controls (red teaming, output filtering) — candidates pick differential privacy because it sounds security-related but addresses privacy, not integrity.

How to eliminate wrong answers

Option A is wrong because differential privacy protects individual privacy in outputs by adding noise; it does not prevent malicious data from being injected during training. Option C is wrong because red teaming evaluates model behavior post-training and is a detection/assessment activity, not a preventive integrity control on training data. Option D is wrong because output filtering operates at inference time to block harmful outputs; it cannot stop poisoned data from corrupting the model during training.

Full explanation →
63
MCQhard

A financial institution is deploying an AI system to approve personal loans. To comply with the EU AI Act's high-risk AI requirements, the bank must ensure meaningful human oversight. Which implementation BEST satisfies this requirement?

A.Require a human to review and approve every loan decision before it becomes final
B.Use a separate AI model to audit the primary AI's decisions weekly
C.Allow applicants to appeal AI decisions through a customer service process
D.Provide a dashboard showing the AI's confidence score for each application
AnswerA

Requiring a human to review and approve every loan decision before it becomes final ensures meaningful human oversight, satisfying the EU AI Act's high-risk requirement. Human-in-the-loop approval prevents fully automated decisions, unlike post-hoc monitoring or logging alone.

Why this answer

The EU AI Act's high-risk requirements (Article 14) mandate that humans can effectively oversee AI systems, including the ability to intervene, override, or halt decisions. Requiring a human to review and approve every loan decision before it becomes final embeds a genuine human-in-the-loop control at the decision point, which is the strongest form of meaningful oversight for a high-risk credit-scoring use case. This ensures no automated output becomes binding without human judgment, directly satisfying the oversight obligation.

Exam trap

AI0-001 often tests the misconception that any human touchpoint (appeals, dashboards, audits) counts as 'meaningful human oversight' when the Act specifically requires the ability to intervene in or override the decision before it takes effect.

How to eliminate wrong answers

Option B is wrong because using a second AI model to audit decisions is still automated oversight with no human in the loop, and weekly review is retrospective rather than preventive. Option C is wrong because an appeal process is a post-hoc redress mechanism, not meaningful human oversight of the AI system itself, and the decision has already been enforced. Option D is wrong because displaying a confidence score merely informs a human without granting them authority or a mechanism to intervene, override, or stop the decision.

Full explanation →
64
Multi-Selecteasy

A data analyst needs to select two appropriate unsupervised learning techniques for clustering unlabeled data. (Choose two.)

Select 2 answers
A.Linear regression
B.Support vector machine
C.Hierarchical clustering
D.Decision tree
E.K-means
AnswersC, E

Hierarchical clustering builds a nested dendrogram by iteratively merging or splitting clusters, requiring no labels. It satisfies the unlabelled-data constraint, and the dendrogram lets analysts choose cluster granularity after the fact rather than fixing k in advance.

Why this answer

Hierarchical clustering (C) is correct because it is an unsupervised technique that groups unlabeled data by iteratively merging or splitting clusters based on a distance metric, producing a dendrogram without requiring predefined labels. K-means (E) is also correct because it is an unsupervised partitioning algorithm that assigns unlabeled points to k clusters by minimizing within-cluster variance. Linear regression (A) is a supervised method that predicts a continuous target from labeled data, so it is not a clustering technique.

Support vector machine (B) is primarily a supervised classifier (and can be used for regression), not an unsupervised clustering method. Decision tree (D) is a supervised model for classification or regression on labeled data, so it does not fit the unlabeled clustering scenario.

Exam trap

The AI0-001 exam often tests the distinction between supervised and unsupervised learning by including familiar algorithms like linear regression or decision trees as distractors, leading candidates to mistake them for clustering techniques due to their popularity in data analysis contexts.

Full explanation →
65
MCQeasy

A retail company wants to add natural language search to its product catalog. The team plans to convert product descriptions and customer queries into embeddings so that semantically similar items surface even when the wording differs. They need an embedding model that maps text into a dense vector space where cosine similarity reflects meaning. Which type of model should they use?

A.An automatic speech recognition model
B.A named entity recognition model
C.A text embedding model
D.A sentiment analysis model
AnswerC

Text embedding models are trained so that semantically similar sentences land close together in a dense vector space, which is exactly what semantic product search needs. Encoding both catalog descriptions and incoming queries with the same model lets cosine similarity rank items by meaning rather than exact keyword overlap, so a query about a waterproof jacket can match related listings.

Why this answer

Semantic search depends on representing both queries and documents as vectors whose proximity encodes meaning. A text embedding model is trained precisely for that, producing dense representations where cosine similarity correlates with semantic relatedness. Using one model for both catalog text and queries keeps them in the same space, enabling relevant matches despite different phrasing.

Exam trap

The trap here is confusing any NLP model with an embedding model, when only models trained for representation learning yield the dense vectors similarity search requires.

Full explanation →
66
MCQhard

A healthcare analytics team deploys a federated learning system across three hospitals to train a diagnostic model without centralizing patient records. A security researcher demonstrates that the shared gradient updates can still be inverted to reconstruct individual patient images. Which additional protection should the team implement on the client updates before aggregation?

A.Require mutual TLS between the aggregation server and each hospital client.
B.Add local differential privacy by clipping and noising each hospital's gradient update before transmission.
C.Increase the number of federated rounds so that gradients converge more slowly.
D.Apply secure aggregation with pairwise masking so the server only sees the summed update.
AnswerB

Local differential privacy perturbs each client's update at the source, so even a curious aggregator or an attacker who intercepts updates cannot invert them to recover patient images. Because noise is added before the update leaves the hospital, the raw gradient never exists in a recoverable form outside the client, directly countering the demonstrated reconstruction attack.

Why this answer

The demonstrated attack reconstructs patient images from shared gradient updates, so the fix must alter the gradients before they leave each hospital. Local differential privacy, applied by clipping and noising each client update at the source, ensures no recoverable raw gradient is ever transmitted. Transport encryption, secure aggregation, and additional rounds leave the underlying gradient content exploitable by inversion techniques.

Exam trap

The trap here is believing that secure aggregation alone hides individual updates, when the aggregate in a small cohort can still be inverted to reconstruct patient data.

Full explanation →
67
MCQhard

A data scientist is building a model to predict the likelihood of a patient having a rare disease. The dataset is highly imbalanced, with only 2% of patients having the disease. The data scientist trains a logistic regression model and achieves 98% accuracy, but the model predicts 'no disease' for all patients. Which evaluation metric should the data scientist use to better assess the model's performance?

A.Mean Squared Error (MSE)
B.Precision-Recall AUC (Area Under the Curve)
C.Accuracy
D.R-squared (R²)
AnswerB

Precision-Recall AUC is particularly useful for imbalanced datasets because it focuses on the positive class. It plots precision against recall at various thresholds, providing a comprehensive view of the model's ability to identify the rare disease without being overwhelmed by the large number of true negatives. This metric helps assess how well the model distinguishes the minority class.

Why this answer

Precision-Recall AUC is designed for imbalanced classification problems, focusing on the performance of the positive class. It provides a more informative picture than accuracy, which can be misleadingly high when the negative class dominates. This metric helps the data scientist understand the trade-off between precision and recall for the rare disease detection.

Exam trap

The trap here is relying on accuracy as the primary metric for imbalanced datasets, which can hide poor performance on the minority class.

Full explanation →
68
Multi-Selectmedium

Which THREE practices are recommended for versioning machine learning models in a production environment?

Select 3 answers
A.Use a model registry like MLflow or DVC.
B.Store model metadata such as hyperparameters and training data hash.
C.Automate model deployment based on version tags.
D.Use Git to version model binaries.
E.Keep only the latest model to save storage.
AnswersA, B, C

A model registry such as MLflow or DVC provides centralised, immutable version tracking, linking each model artefact to its training data, parameters and metrics. This satisfies the production requirement for reproducible lineage and rollback, letting teams promote or revert specific model versions without ambiguity.

Why this answer

Option A is correct because a dedicated model registry such as MLflow or DVC is purpose-built for tracking model artifacts, versions, and lifecycle stages, which is the recommended practice for production ML versioning. Option B is correct because storing metadata like hyperparameters and the training data hash ties each model version to its exact training configuration and dataset, enabling reproducibility and auditability. Option C is correct because automating deployment based on version tags ensures that only approved, traceable model versions reach production and keeps deployment consistent with the registry.

Option D is not recommended because Git is designed for source code and text, not large binary model files, which bloat repositories and lack proper artifact lineage. Option E is wrong because discarding older models destroys rollback capability, reproducibility, and compliance auditing, and storage savings do not justify that risk.

Exam trap

CompTIA often tests the misconception that Git is suitable for versioning all artifacts, including large binary model files, when in fact Git's architecture is optimized for text diffs and cannot efficiently manage model binaries in a production ML pipeline.

Full explanation →
69
MCQeasy

An ML engineer wants to deploy a model as a REST API that can scale to handle thousands of inference requests per second. Which serving approach is most appropriate?

A.Export the model to ONNX format and use a batch processing pipeline
B.Use gRPC streaming for all inference requests
C.Run the model directly on the client device
D.Deploy the model as a REST API endpoint using a containerized inference server
AnswerD

A containerised inference server exposes the model as a REST endpoint and supports horizontal scaling across replicas behind a load balancer, meeting the thousands-of-requests-per-second requirement. Containers also package dependencies consistently, enabling elastic autoscaling of inference capacity.

Why this answer

Deploying the model as a REST API endpoint using a containerized inference server (e.g., TensorFlow Serving, TorchServe, or NVIDIA Triton Inference Server) is the most appropriate approach for handling thousands of inference requests per second. These servers are designed for high-throughput, low-latency serving, support horizontal scaling via load balancers, and provide built-in batching and model versioning. REST APIs are stateless and can be easily integrated with existing web infrastructure, making them ideal for production-scale inference.

Exam trap

The AI0-001 exam often tests the distinction between serving infrastructure (REST API with containerized server) and data processing pipelines (batch) or communication protocols (gRPC), leading candidates to confuse a transport mechanism or batch method with a scalable serving architecture.

How to eliminate wrong answers

Option A is wrong because exporting to ONNX and using a batch processing pipeline is designed for offline/batch inference, not for real-time REST API serving with thousands of requests per second; batch pipelines introduce latency and are not suitable for synchronous, low-latency inference. Option B is wrong because gRPC streaming is a communication protocol that can be used for inference, but it is not a serving approach itself; moreover, gRPC streaming is typically used for bidirectional or long-lived streams, not for high-volume stateless REST API requests, and it adds complexity without inherent scalability benefits over REST for this use case. Option C is wrong because running the model directly on the client device (edge inference) offloads computation from the server but does not provide a centralized REST API; it also introduces challenges with model updates, device heterogeneity, and security, and is not a server-side serving approach.

Full explanation →
70
MCQeasy

A data scientist wants to protect the privacy of individuals whose data is used to train a model, even if the model is compromised. Which technique ensures that the model does not memorize sensitive information?

A.Federated learning
B.Homomorphic encryption
C.Differential privacy
D.Data anonymization
AnswerC

Differential privacy injects calibrated noise into training or query outputs, bounding any single individual's influence so the model cannot memorise their record. This satisfies the requirement that privacy survives model compromise, unlike encryption or anonymisation, which protect data at rest rather than learned parameters.

Why this answer

Differential privacy (C) is the correct technique because it adds calibrated noise to the training data or model updates, ensuring that the model's outputs do not reveal whether any specific individual's data was included. This guarantees that even if an attacker gains full access to the model, they cannot extract sensitive information about any single record, as the noise bounds the influence of any one data point.

Exam trap

CompTIA often tests the misconception that data anonymization (D) is sufficient for model privacy, but candidates must recognize that anonymization does not protect against model inversion or membership inference attacks, whereas differential privacy provides a formal mathematical guarantee.

How to eliminate wrong answers

Option A is wrong because federated learning distributes training across devices but does not inherently prevent memorization; the model can still leak sensitive data if the aggregation or updates are not privacy-preserving. Option B is wrong because homomorphic encryption allows computation on encrypted data but protects data in transit or at rest, not the model's internal memorization of training examples. Option D is wrong because data anonymization removes direct identifiers but is vulnerable to re-identification attacks via auxiliary information, and does not prevent the model from memorizing patterns that can be linked back to individuals.

Full explanation →
71
MCQmedium

A machine learning team is training a large transformer model on a text corpus. They need to reduce training time while maintaining model accuracy. Which hardware configuration would be MOST effective for this task?

A.Use a high-core-count CPU with large RAM
B.Use a cluster of GPUs with data parallelism
C.Use a single GPU with model parallelism
D.Use a single TPU with model parallelism
AnswerB

Data parallelism distributes each batch across many GPUs, each holding a full model replica and synchronising gradients, which cuts wall-clock training time substantially. This scales effectively for large transformer models while preserving accuracy through equivalent gradient updates.

Why this answer

Training large transformer models is computationally intensive, and data parallelism across a cluster of GPUs allows the model to process multiple batches simultaneously, dramatically reducing training time. Each GPU holds a full copy of the model and processes a different subset of the data, with gradients synchronized across devices. This approach scales well and maintains accuracy as long as the effective batch size is tuned appropriately.

Exam trap

AI0-001 often tests the misconception that a single powerful device (TPU or GPU) with model parallelism is better than a multi-GPU cluster; candidates may pick option D because TPUs are marketed for transformers, but the question emphasizes reducing training time, which favors data parallelism across many GPUs.

How to eliminate wrong answers

Option A is wrong because high-core-count CPUs with large RAM are far slower than GPUs for the matrix multiplications and attention operations in transformers; CPUs lack the massive parallelism and high memory bandwidth needed for efficient deep learning. Option C is wrong because a single GPU with model parallelism splits the model across layers within one device, which does not reduce training time as effectively as scaling across multiple GPUs; it also introduces communication overhead and is limited by the single GPU's memory and compute. Option D is wrong because a single TPU with model parallelism, while powerful, is less effective than a cluster of GPUs for reducing training time on large transformers; TPUs are optimized for specific workloads and may require code changes, and a single device cannot match the throughput of a multi-GPU cluster.

Full explanation →
72
MCQeasy

An AI system for fraud detection shows a gradual decline in precision over several weeks, though recall remains stable. Which type of model drift is most likely occurring?

A.Data drift
B.Covariate shift
C.Label drift
D.Concept drift
AnswerD

Precision falling while recall stays stable indicates the relationship between input features and the fraud label has shifted, so previously correct positive predictions increasingly become false positives. That change in the underlying target concept, rather than input distribution, is concept drift.

Why this answer

Concept drift occurs when the statistical relationship between input features and the target variable changes over time, causing the model's decision boundary to become less accurate. In this scenario, precision is declining while recall remains stable, indicating that the model is producing more false positives even though it still catches the same proportion of true positives. This is a classic sign of concept drift, where the underlying definition of fraud has shifted, not the data distribution itself.

Exam trap

The CompTIA AI exam often tests the distinction between data drift and concept drift by presenting a scenario where only one performance metric changes, tempting candidates to incorrectly choose data drift because they associate any performance decline with input data changes, rather than recognizing that a stable recall with dropping precision points to a shift in the underlying concept.

How to eliminate wrong answers

Option A is wrong because data drift refers to changes in the distribution of input features, which would typically affect both precision and recall or cause a shift in all performance metrics, not a selective decline in precision alone. Option B is wrong because covariate shift is a specific type of data drift where the distribution of input features changes while the conditional distribution P(y|x) remains the same; here, the conditional relationship is changing, as evidenced by the precision drop. Option C is wrong because label drift involves changes in the distribution of the target labels (e.g., the overall fraud rate), which would affect recall and precision together, not precision in isolation with stable recall.

Full explanation →
73
Multi-Selectmedium

A financial institution wants to use AI for loan approvals and must comply with fair lending laws. Which TWO practices should the institution adopt to mitigate bias and ensure compliance?

Select 2 answers
A.Remove all features except credit score to avoid bias
B.Use a black-box model without explainability to protect intellectual property
C.Use only demographic features to ensure equal treatment
D.Apply fairness-aware machine learning techniques during model training
E.Conduct disparate impact analysis on model outcomes
AnswersD, E

Fairness-aware training constrains the model's objective function to reduce disparate impact across protected groups, directly addressing the fair lending requirement. By penalising biased outcomes during fitting rather than only auditing afterwards, it satisfies the stem's compliance constraint at the point where bias enters the model.

Why this answer

Option D is correct because fairness-aware machine learning techniques (e.g., pre-processing, in-processing, or post-processing bias mitigation methods such as reweighting, adversarial debiasing, or equalized odds) directly address bias during model training, which is essential for fair lending compliance. Option E is correct because disparate impact analysis quantifies whether model outcomes disproportionately disadvantage protected groups, a key requirement under fair lending laws like the Equal Credit Opportunity Act (ECOA) and Fair Housing Act. Option A is incorrect because removing all features except credit score does not eliminate bias—credit scores themselves can reflect historical discrimination—and it may violate fair lending rules by ignoring relevant factors.

Option B is incorrect because black-box models without explainability hinder regulatory audits and adverse action explanations required by laws such as ECOA and the Fair Credit Reporting Act (FCRA). Option C is incorrect because using only demographic features would be both discriminatory and nonsensical for loan approval decisions.

Exam trap

AI0-001 often tests the misconception that removing sensitive features ('fairness through unawareness') eliminates bias, when proxy variables and historical bias in remaining features preserve discrimination.

Full explanation →
74
MCQmedium

A deep learning model for sentiment analysis uses a softmax output layer. The hidden layers currently use tanh activation. Which activation function should replace tanh to mitigate vanishing gradients in deeper networks?

A.Sigmoid
B.Softmax
C.ReLU
D.Linear
AnswerC

ReLU outputs the input directly for positive values, giving a derivative of one and avoiding the saturation that drives tanh's gradient toward zero in deep stacks. This preserves gradient magnitude during backpropagation, directly mitigating the vanishing gradient problem in the deeper network.

Why this answer

ReLU (Rectified Linear Unit) is correct because it outputs zero for negative inputs and a positive linear slope for positive inputs, which avoids the saturation problem of tanh. In deeper networks, tanh gradients can vanish as activations approach ±1, slowing or halting learning. ReLU's non-saturating nature keeps gradients flowing for positive inputs, mitigating the vanishing gradient problem.

Exam trap

Candidates often mistakenly believe that any non-linear activation works equally well in deep networks, but the trap is that they may choose sigmoid because it is non-linear, ignoring its saturation-induced vanishing gradient problem in deeper architectures.

How to eliminate wrong answers

Option A is wrong because sigmoid also saturates at 0 and 1, causing vanishing gradients in deep networks, similar to tanh. Option B is wrong because softmax is typically used in the output layer for multi-class classification, not in hidden layers, and it saturates across classes, exacerbating vanishing gradients. Option D is wrong because a linear activation would collapse the network into a single linear transformation, removing the non-linearity needed to learn complex patterns in sentiment analysis.

Full explanation →
75
MCQhard

A team is training a deep learning model for image classification. The training loss decreases rapidly but validation loss starts increasing after a few epochs. Which regularization technique should be applied to mitigate this issue?

A.Data augmentation
B.L2 regularization
C.Early stopping
D.Dropout
AnswerC

Rising validation loss alongside falling training loss signals overfitting. Early stopping halts training at the epoch where validation loss is minimal, restoring the best generalising weights. This directly mitigates the divergence described, unlike dropout or weight decay, which alter the architecture or loss function.

Why this answer

Early stopping halts training when validation loss starts increasing, preventing overfitting. Option A (data augmentation) is wrong because it increases data diversity but does not stop training when validation loss increases. Option B (L2 regularization) is wrong because it penalizes large weights but does not directly address the issue of validation loss increasing.

Option D (dropout) is wrong because while it helps generalize by randomly dropping neurons, it does not stop training when overfitting occurs.

Full explanation →

Page 1 of 13

Page 2

All pages

12345678910111213

Practice AI0-001 by domain

Target a specific domain to shore up weak areas.

Implementing AI SolutionsAI Governance and EthicsAI Concepts and TechniquesAI Concepts and FoundationsAI SecurityAI Infrastructure and TechnologiesAI Models and Data EngineeringMachine Learning and Deep LearningAI Security, Ethics and GovernanceAI Implementation and Operations
See all domains with question counts →