AI0-001 Practice Test 42 — 15 Questions

Question 1

An LLM-powered application occasionally generates factual-sounding but incorrect information. Users rely on this output for decision-making. Which risk does this primarily represent?

Accepted Answer

Hallucinations and over-reliance. The scenario describes an LLM generating plausible but incorrect information (hallucination) and users relying on it for decisions (over-reliance). This directly matches the combined risk of hallucinations and over-reliance, as the model's confident but false outputs can lead to poor decision-making without proper verification.

Answer

Sensitive information disclosure

Answer

Model denial of service

Answer

Prompt injection

Question 2

A company is fine-tuning a large language model using PEFT (Parameter-Efficient Fine-Tuning) to reduce GPU memory usage. They have limited hardware and need to fine-tune a 70B parameter model on a single GPU with 24 GB VRAM. Which technique is MOST suitable?

Accepted Answer

QLoRA (Quantization-aware LoRA) with 4-bit quantization. QLoRA combines quantization (4-bit) and LoRA to fine-tune very large models on limited hardware, achieving significant memory reduction while maintaining performance.

Answer

Full fine-tuning with gradient checkpointing

Answer

Instruction tuning with a smaller 7B model

Answer

LoRA (Low-Rank Adaptation) alone

Question 3

A team is designing a RAG system for a large collection of PDFs. They need to choose document chunking strategies. Which TWO strategies are considered best practices? (Choose two.)

Accepted Answer

Semantic chunking (e.g., sentence or paragraph boundaries). Semantic chunking (A) is a best practice because splitting text at natural sentence or paragraph boundaries preserves coherent, self-contained units of meaning, which improves embedding quality and retrieval relevance in a RAG pipeline. Hierarchical chunking (C) is also a best practice because it captures the document's section and subsection structure, allowing retrieval at multiple granularities (e.g., retrieving a subsection but supplying its parent section as context) and better handling long, structured PDFs. Fixed-size chunking with no overlap (B) is not recommended here because it can cut sentences or ideas mid-thought and provides no overlap to preserve context across boundaries. A single chunk per document (D) is unsuitable because large PDFs would exceed embedding model token limits and dilute semantic focus, hurting retrieval precision. Random character-length chunks (E) are arbitrary and break semantic and structural coherence, making retrieval unreliable.

Answer

Fixed-size chunking with no overlap

Answer

Single chunk per document

Answer

Random character-length chunks

Question 4

A company uses an LLM to generate code. They want to ensure that the model does not accidentally output sensitive internal logic. Which practice should they implement?

Accepted Answer

Output filtering to block sensitive patterns. Output filtering is the correct practice because it directly inspects the model's generated text for patterns that match sensitive internal logic (e.g., API keys, source code snippets, or proprietary algorithms) and blocks or redacts them before the output is returned to the user. This is a reactive security control that operates at the application layer, ensuring that even if the LLM inadvertently generates sensitive content, it is never exposed. Rate limiting, red teaming, and federated learning address different concerns (availability, vulnerability discovery, and data privacy during training, respectively) and do not prevent the accidental leakage of internal logic in real-time outputs.

Answer

Rate limiting API calls

Answer

Red teaming the model

Answer

Federated learning

Question 5

A company is building an AI-powered document intelligence system to extract key fields from scanned invoices. The data contains 95% of invoices from one vendor and 5% from others. During model training, the F1 score is 0.95 on the overall test set, but the performance on the minority vendor invoices is very poor. What is the MOST likely cause?

Accepted Answer

The dataset is imbalanced, and the model is biased toward the majority class. The F1 score of 0.95 on the overall test set is misleading because 95% of invoices come from one vendor, so the model can achieve high overall accuracy by performing well on the majority class while failing on the minority class. This is a classic class imbalance problem where the model is biased toward the majority class. The poor performance on minority vendor invoices confirms that the model has not learned to generalize across vendors.

Answer

The model is overfitting on the minority class

Answer

The data has a train/test leakage problem

Answer

The feature extraction is incorrect for the minority vendor invoices

Question 6

A data science team needs to implement privacy-preserving ML for a healthcare model. They require that individual patient records cannot be distinguished in the training output. Which technique should be applied?

Accepted Answer

Differential privacy. Differential privacy is the correct technique because it adds calibrated noise to the training process or query outputs, ensuring that the inclusion or exclusion of any single patient record does not significantly affect the model's output. This provides a formal mathematical guarantee that individual records cannot be distinguished, which directly meets the requirement for privacy-preserving ML in healthcare.

Answer

Homomorphic encryption

Answer

Model pruning

Answer

Federated learning

Question 7

A company wants to use AI to automatically detect anomalies in server log data. The data is time-series and labeled with 'normal' and 'anomaly' for the past year. Which TWO techniques are appropriate for this use case?

Accepted Answer

Use a time-series anomaly detection model (e.g., Isolation Forest with sliding windows). Option B is correct because the data is time-series log data, and techniques like Isolation Forest applied over sliding windows (or similar time-series anomaly detectors) are designed to capture temporal patterns and flag deviations from normal behavior without requiring the labels, which suits anomaly detection on sequential log streams. Option C is correct because the dataset is labeled with 'normal' and 'anomaly' for a full year, so a supervised classifier such as XGBoost can be trained on extracted features (e.g., counts, rates, error codes, latency statistics) to directly learn the mapping from features to the anomaly label. Option A is not appropriate because converting logs to graph screenshots and using a CNN image classifier discards the underlying time-series structure and numeric log semantics, making it an indirect and lossy approach. Option D is wrong because code generation models fix code rather than detect anomalies in log data, which is the stated goal. Option E is wrong because a recommendation system based on user activity logs addresses personalization, not anomaly detection in server logs.

Answer

Train an image classification model (CNN) on screenshots of log graphs

Answer

Use a code generation model to fix the anomalies automatically

Answer

Build a recommendation system based on user activity logs

Question 8

During a security review, an auditor finds that an LLM application can call external functions (e.g., send emails, update databases) based on user prompts. Which risk is MOST concerning?

Accepted Answer

Excessive agency. Excessive agency (OWASP LLM Top 10 LLM08) describes a system where an LLM is granted more permissions, functionality, or autonomy than necessary — here, the ability to send emails and update databases based on user prompts. The most concerning risk is that a prompt injection or hallucination can trigger real-world side effects (data exfiltration, unauthorized transactions) because the model has the agency to act. The root problem is the excessive capability granted to the model, not the injection itself.

Answer

Prompt injection

Answer

Model denial of service

Answer

Hallucinations producing dangerous advice

Question 9

A team is evaluating an LLM-based chatbot that frequently hallucinates when answering questions about internal policies. Which testing approach would MOST effectively quantify this issue?

Accepted Answer

Evaluation frameworks for LLM output quality. Evaluation frameworks for LLM output quality, such as those using metrics like faithfulness, factuality, or ROUGE/BLEU scores, are specifically designed to detect and quantify hallucinations by comparing generated responses against a ground-truth knowledge base. This directly measures the rate at which the chatbot fabricates or misstates internal policy details, providing a quantitative baseline for improvement.

Answer

Integration tests for API calls

Answer

Unit tests for the data pipeline

Answer

Regression testing of model accuracy over time

Question 10

A healthcare AI system that diagnoses medical images must provide explanations for its predictions to comply with regulatory requirements. Which technique should the team implement?

Accepted Answer

Apply model interpretability methods such as SHAP or LIME.. SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are established model interpretability techniques that provide per-prediction explanations, which are essential for regulatory compliance in healthcare AI. These methods generate feature attribution scores or local surrogate models to explain why a specific diagnosis was made, meeting transparency requirements without sacrificing model performance.

Answer

Reduce the model's accuracy to make it simpler.

Answer

Only deploy rule-based systems.

Answer

Use a more complex deep learning model.

Question 11

An AI team is concerned about their model leaking sensitive information from its training data when queried. Which privacy-preserving technique adds noise to the training process to limit what can be inferred about any individual record?

Accepted Answer

Differential privacy. Differential privacy (A) is the correct answer because it directly addresses the concern of leaking sensitive information from training data by adding calibrated noise to the training process or query responses. This noise ensures that the output of the model does not significantly change whether any single individual's record is included or excluded, thereby limiting what can be inferred about any specific record. The technique is formalized through a privacy budget (ε, epsilon) that quantifies the privacy guarantee, making it the standard approach for privacy-preserving machine learning.

Answer

Homomorphic encryption

Answer

Data sanitization

Answer

Federated learning

Question 12

A team is training a deep learning model for image classification. They observe that training accuracy is high but validation accuracy is low, indicating overfitting. Which TWO techniques should they apply to reduce overfitting? (Select TWO)

Accepted Answer

Use L2 regularization. Option A, using L2 regularization, is correct because adding a weight-decay penalty (e.g., lambda * sum of squared weights) to the loss function constrains the magnitude of the model's weights, discouraging the network from fitting noise in the training set and thereby reducing overfitting. Option C, adding dropout layers, is correct because dropout randomly deactivates a fraction of neurons (e.g., p=0.5) during each training step, preventing units from co-adapting and forcing the network to learn more robust, generalizable features. Option B, increasing the learning rate, is not appropriate because a larger learning rate typically causes unstable or divergent training rather than reducing overfitting. Option D, increasing the number of layers, would raise model capacity and generally worsen overfitting. Option E, reducing training data size, would make overfitting more severe by giving the model even less data to generalize from.

Answer

Increase learning rate

Answer

Increase the number of layers

Answer

Reduce training data size

Question 13

A team trained a ResNet-50 model with the configuration shown. The high training accuracy and lower validation accuracy suggest overfitting. Which change to the training configuration is MOST likely to reduce overfitting?

Accepted Answer

Add dropout layers after convolutional layers.. Adding dropout layers after convolutional layers is a regularization technique that randomly drops a fraction of neurons during training, which forces the network to learn more robust features and reduces overfitting. This directly addresses the symptom of high training accuracy with lower validation accuracy by preventing the model from relying too heavily on specific neurons.

Answer

Reduce number of epochs to 5.

Answer

Increase batch size to 64.

Answer

Increase learning rate to 0.01.

Question 14

A company is developing an AI-powered recruitment tool. To prevent bias and ensure fairness, they want to audit the model's training data and outputs. Which TWO practices should they implement as part of secure AI development?

Accepted Answer

Threat modeling using STRIDE for AI-specific threats. Option B is correct because threat modeling with STRIDE helps identify AI-specific security and fairness risks (e.g., tampering with training data, information disclosure, or elevation of privilege in the ML pipeline) before they manifest, directly supporting a secure and auditable AI development process. Option D is correct because implementing access controls on the training dataset enforces least privilege and prevents unauthorized modification or exfiltration of data, which is essential for maintaining data integrity and enabling trustworthy bias audits. Options A and E are incorrect because model parallelism and larger batch sizes are performance/scalability tuning techniques that do not address fairness, bias auditing, or security. Option C is incorrect because increasing the learning rate is a hyperparameter change that affects convergence and training dynamics, not the governance or security posture of the AI system.

Answer

Enabling model parallelism

Answer

Increasing the model's learning rate

Answer

Using a larger batch size

Question 15

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

Accepted Answer

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

Answer

Hugging Face Transformers

Answer

scikit-learn

Answer

Keras