AI0-001 Practice Test 35 — 15 Questions

Question 1

In the AI project lifecycle, which phase involves splitting the dataset into training, validation, and test sets while ensuring no data leakage?

Accepted Answer

Data preparation. Splitting the dataset into training, validation, and test sets is a core data preparation step that must be performed before any model training begins. This phase ensures that data leakage is prevented by keeping the test set completely isolated until final evaluation, which is critical for obtaining an unbiased estimate of model performance. In the AI project lifecycle, data preparation encompasses cleaning, transforming, and partitioning the data, making option A the correct phase.

Answer

Problem definition

Answer

Data acquisition

Answer

Model evaluation

Question 2

A logistics company wants to detect packages damaged in transit by analyzing photos taken at warehouse checkpoints. The team has only about 200 labeled examples of damaged packages but tens of thousands of photos of undamaged ones. They need a working classifier quickly and cannot collect more damaged-package images in the near term. Which approach should the team use?

Accepted Answer

Fine-tune a pretrained vision model on the labeled images and address the class imbalance with techniques such as class weighting or data augmentation.. With only a few hundred damaged-package examples against many undamaged ones, transfer learning is the practical path: fine-tuning a pretrained vision model leverages general visual features and works with small labeled sets, while class weighting or augmentation handles the imbalance. Training from scratch overfits, waiting for more data violates the timeline, and rule-based comparison cannot reliably distinguish damage from normal variation.

Answer

Collect at least 50,000 additional damaged-package images before training any model.

Answer

Train a convolutional neural network from scratch on the 200 damaged-package images until training accuracy reaches 100 percent.

Answer

Deploy a rule-based image comparison that flags any photo differing from the warehouse's average undamaged package appearance.

Question 3

A logistics company runs a vision model on edge devices in warehouses to detect damaged packages on conveyor belts. The model must classify each package within 40 milliseconds, and network connectivity to the cloud is unreliable. During a pilot, engineers notice that accuracy on the edge devices is several points lower than the accuracy measured during cloud-based evaluation on the same test images. Which cause is MOST likely?

Accepted Answer

Quantization applied to compress the model for the edge device reduced numerical precision in the weights and activations.. When the same test images yield lower accuracy on the edge than in the cloud, the difference must come from something that changes the computation itself. Quantization to a lower-precision numeric format alters weights and activations, which is the standard cause of a small accuracy regression in compressed edge models. Camera mismatch would affect both environments equally, and hardware speed or batch size do not change the mathematical output for a given image.

Answer

The edge devices have less RAM and slower CPUs than the cloud inference servers used during evaluation.

Answer

The test images were captured with a different camera model than the ones mounted on the conveyor belts.

Answer

The cloud evaluation used a batch size of 32 while the edge device processes one image at a time.

Question 4

A media company serves a generative AI assistant to customers through an API. After an update to the system prompt, users begin reporting that the assistant produces responses outside the company's approved tone and occasionally reveals parts of its internal instructions. The operations team must add safeguards that reduce these behaviors in production. (Choose two.)

Accepted Answer

Add adversarial red-team testing of prompt injection and instruction-extraction attempts, and use the findings to harden the system prompt and input handling.. The complaints describe two distinct failure modes: off-policy tone and disclosure of internal instructions. A response moderation layer enforces policy on what actually leaves the system, while adversarial red-team testing hardens the prompt and input handling against extraction attempts. Together they cover both detection and root-cause reduction. Temperature, token limits, and response caching affect variability, length, and cost, none of which constrain the model's willingness to violate policy or leak its instructions.

Answer

Raise the maximum token limit for responses so the assistant has more room to explain itself clearly.

Answer

Cache frequent prompts and their responses to reduce load on the inference endpoint and improve consistency.

Answer

Increase the model's temperature setting so responses vary more and are less likely to follow a fixed undesirable pattern.

Question 5

An AI application needs to generate structured JSON output from an LLM. The development team wants to ensure the output always conforms to a specific schema. Which prompt engineering technique is MOST suitable?

Accepted Answer

System prompt with JSON schema and a 'respond only with valid JSON' instruction. Providing the JSON schema directly in the system prompt, combined with an explicit instruction to respond only with valid JSON, is the most direct and reliable way to constrain an LLM's output format. This technique leverages the model's instruction-following capability and schema awareness without requiring examples or retraining, ensuring strict adherence to the desired structure.

Answer

Few-shot examples showing correct JSON

Answer

Chain-of-thought prompting

Answer

Fine-tuning the model on JSON datasets

Question 6

A marketing team wants to use an AI system to generate personalized advertisements that include realistic images of celebrities endorsing products. The legal team raises concerns about compliance with the EU AI Act. Which action should the team take to comply with the Act's transparency requirements?

Accepted Answer

Clearly disclose that the images are AI-generated and not real. Under the EU AI Act, deepfake content must be disclosed as artificially generated or manipulated. Generating realistic images of celebrities endorsing products without disclosure would violate this transparency requirement. The team must clearly label the images as AI-generated, regardless of licensing or platform policies.

Answer

Use only images of celebrities who are deceased, as the Act only applies to living individuals

Answer

Limit the advertisements to social media platforms that have their own AI disclosure policies

Answer

Obtain a license from each celebrity before generating their image

Question 7

A hospital plans to deploy an AI system that analyzes patient data to predict the likelihood of hospital readmission. The system will be used to allocate post-discharge care resources. The hospital's ethics committee wants to ensure compliance with the EU AI Act's requirements for high-risk AI systems. Which practice is MOST critical for meeting the Act's human oversight requirements?

Accepted Answer

Ensuring that a qualified clinician can review and override the AI's recommendations before care resources are allocated. For high-risk AI systems, the EU AI Act mandates human oversight to ensure that humans can effectively monitor and intervene in the system's operation. In a hospital readmission prediction tool used for care allocation, a clinician must be able to review and override the AI's recommendations, preserving human judgment in decisions that affect patient health.

Answer

Requiring that the AI system's predictions are always followed to maintain consistency in care allocation

Answer

Publishing the AI system's algorithm in a peer-reviewed journal before deployment

Answer

Deploying the AI system only after it achieves 100% accuracy on historical readmission data

Question 8

A company wants to automatically group customer support tickets into categories (e.g., billing, technical, account) without pre-labeled data. Which machine learning approach should they use?

Accepted Answer

Unsupervised clustering using K-means. The company has no pre-labeled data, which means supervised learning (which requires labeled examples) is not feasible. Unsupervised clustering, such as K-means, groups data points into clusters based on feature similarity without needing any labels, making it ideal for automatically discovering categories like billing, technical, or account from raw ticket text.

Answer

Supervised classification with logistic regression

Answer

Semi-supervised learning with a small labeled set

Answer

Reinforcement learning with a reward function

Question 9

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

Accepted Answer

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

Answer

The system violates individual fairness

Answer

The system suffers from selection bias

Answer

The system is not calibrated

Question 10

An organization wants to classify support tickets into categories (billing, technical, etc.). Which type of machine learning is most suitable?

Accepted Answer

Supervised learning. Supervised learning is the correct choice because the organization has labeled historical support tickets (e.g., 'billing' or 'technical') and wants to train a model to map new tickets to these predefined categories. This is a classic classification task, where the algorithm learns from input-output pairs to predict the correct label for unseen data.

Answer

Unsupervised learning

Answer

Reinforcement learning

Answer

Regression

Question 11

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?

Accepted Answer

ReAct pattern with tool descriptions and function calling. 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.

Answer

Chain-of-thought prompting only

Answer

Zero-shot prompting with JSON mode

Answer

Simple prompt with no tool descriptions

Question 12

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?

Accepted Answer

API key authentication. 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.

Answer

Rate limiting

Answer

Input sanitization

Answer

IP whitelisting

Question 13

A healthcare AI system misdiagnosed patients due to adversarial inputs. What security measure should be prioritized?

Accepted Answer

Implement adversarial training. (Implement adversarial training) is correct because adversarial training makes the model robust to input manipulation. Option A (Encrypt all patient data) protects data privacy but not model integrity. Option B (Use stronger authentication) is for access control. Option C (Regular software updates) is general maintenance and does not specifically address adversarial inputs.

Answer

Encrypt all patient data

Answer

Use stronger authentication

Answer

Regular software updates

Question 14

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?

Accepted Answer

Data drift. 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.

Answer

Concept drift

Answer

Bias-variance tradeoff

Answer

Overfitting

Question 15

A financial institution uses a machine learning model to approve loan applications. The model was trained on historical data that inadvertently encoded a bias against applicants from certain zip codes, leading to discriminatory lending practices. A recent audit reveals that the model's decisions are unfair, and regulators require the bank to remediate the bias without significantly reducing overall approval accuracy. The data science team has access to the training data, the model, and a set of fairness metrics. They also have a small, unbiased validation set. Which course of action should the team take to satisfy regulatory requirements?

Accepted Answer

Implement adversarial debiasing using the unbiased validation set to enforce fairness constraints. Adversarial debiasing directly addresses the bias encoded in the model by training a predictor and an adversary simultaneously. The adversary tries to predict the protected attribute (e.g., zip code) from the model's predictions, while the predictor is penalized for allowing such inference, enforcing fairness constraints. Using the unbiased validation set ensures the debiasing process is guided by ground truth labels that are free from historical bias, allowing the model to retain high accuracy while reducing discrimination.

Answer

Remove the zip code feature from the model and retrain

Answer

Increase the weight of samples from disadvantaged zip codes in the training data

Answer

Retrain the model using only the unbiased validation set