AI0-001 Practice Test 49 — 15 Questions

Question 1

An AI model for detecting fraudulent transactions has high precision but low recall. Which business impact is most likely?

Accepted Answer

Many fraudulent transactions go undetected. High precision means that when the model flags a transaction as fraudulent, it is very likely correct. However, low recall indicates that the model misses a significant proportion of actual fraudulent transactions. Therefore, the most likely business impact is that many fraudulent transactions go undetected, leading to financial losses.

Answer

The model has no impact on fraud detection

Answer

The model detects all fraudulent transactions

Answer

Many legitimate transactions are flagged as fraud

Question 2

A company is building an AI-based resume screening tool. They want to ensure the system is secure against data poisoning attacks during the training phase. Which THREE of the following are appropriate defensive measures?

Accepted Answer

Use robust statistical methods (e.g., trimmed mean) that are less sensitive to outliers. Option B is correct because robust statistical methods such as trimmed mean, median, or RANSAC reduce the influence of maliciously injected outlier samples during training, directly mitigating data poisoning. Option C is correct because validating and cleaning training data to detect and remove anomalies, label inconsistencies, and outliers prevents poisoned samples from entering the training set in the first place. Option D is correct because restricting training data to trusted, verified providers reduces the attack surface by ensuring provenance and integrity of the data supply chain, which is a key defense against poisoning. Option A is not appropriate here because input sanitization at inference time addresses runtime adversarial inputs (e.g., evasion or prompt injection), not training-phase poisoning. Option E is not appropriate because differential privacy protects against privacy leakage of individual training records and does not by itself defend against data poisoning attacks.

Answer

Apply input sanitization to inference-time queries

Answer

Implement differential privacy during model training

Question 3

A large enterprise is developing an internal LLM-powered assistant that can access the internet and execute code. To mitigate risks from excessive agency (e.g., the model performing unauthorized actions), which THREE security measures should be implemented?

Accepted Answer

Require human-in-the-loop approval for code execution and write operations. Option B is correct because requiring human-in-the-loop approval for code execution and write operations directly constrains excessive agency by ensuring a human authorizes high-impact actions before the model can perform them. Option C is correct because least-privilege API tokens limit the blast radius of any tool or external access the model invokes, so even a misused token can only perform the minimum permitted operations. Option D is correct because input validation and sanitization reduce prompt-injection vectors that could otherwise hijack the model into issuing unauthorized tool calls or code, which is a primary enabler of excessive agency. Option A is not among the marked answers because monitoring anomalous input patterns is detective and does not by itself prevent unauthorized model actions. Option E is not among the marked answers because output filtering addresses data leakage in responses rather than constraining the model's ability to take unauthorized actions.

Answer

Deploy monitoring for anomalous input patterns

Answer

Apply output filtering to block sensitive data in responses

Question 4

An organization wants to integrate an AI-powered summarization feature into their existing web application. The AI service will be called via API. Which factor is MOST important to consider for cost management?

Accepted Answer

Token pricing of the AI model. Token pricing directly determines the cost of each API call because AI models charge based on the number of input and output tokens processed. Since the summarization feature will make frequent API calls, token pricing is the primary cost driver. Other factors like authentication, rate limits, and latency affect security, throughput, and performance, but not the direct cost per request.

Answer

Authentication method (API key vs. OAuth)

Answer

Rate limits per minute

Answer

Network latency to the API endpoint

Question 5

A data scientist discovers that a model trained to predict loan defaults is denying loans at a higher rate for a particular demographic group. Which type of bias is MOST likely present?

Accepted Answer

Historical bias. Historical bias occurs when the training data reflects past societal inequalities, leading the model to learn and perpetuate those patterns. In this case, if historical loan data shows higher denial rates for a demographic group due to past discriminatory practices, the model will replicate that bias in its predictions. This is the most likely cause because the model is not inherently biased but inherits bias from the data it was trained on.

Answer

Confirmation bias

Answer

Selection bias

Answer

Algorithmic bias

Question 6

A DevOps team is deploying a machine learning model using a CI/CD pipeline. They want to ensure the model is reproducible and traceable. Which TWO practices should they implement?

Accepted Answer

Version the training dataset and code using Git and DVC.. Option A is correct because versioning both the training dataset and the code with Git and DVC (Data Version Control) captures the exact data and source revisions used, which is essential for reproducing and tracing a model. Option E is correct because packaging the model in a Docker container with a fixed (pinned) base image locks down the OS libraries, dependencies, and runtime environment, ensuring the model behaves identically across environments and builds. Option B is not appropriate because manual deployment after approval is error-prone and not automated or traceable, undermining CI/CD reproducibility. Option C is wrong because storing only the final model artifact in a shared drive loses the training data, code, and environment context needed for reproducibility. Option D is wrong because a spreadsheet is a manual, non-versioned record that cannot reliably or automatically trace model versions.

Answer

Manually deploy the model to production after approval.

Answer

Store only the final model artifact in a shared drive.

Answer

Use a spreadsheet to record model version numbers.

Question 7

A data scientist is using PyTorch to train a custom NLP model. The training is slow on a single GPU. They want to speed up training by using multiple GPUs on a single machine. Which PyTorch feature should they use?

Accepted Answer

torch.nn.DataParallel. torch.nn.DataParallel is PyTorch's built-in module for single-machine, multi-GPU training. It wraps the model and automatically splits the input batch across available GPUs, replicating the model on each device and gathering outputs on the primary GPU — requiring only a one-line code change (wrapping the model).

Answer

TorchScript tracing

Answer

torch.optim.SGD

Answer

PyTorch Lightning's zero_grad function

Question 8

A bank uses an AI system for credit scoring. To meet fairness requirements, they want to ensure the model predicts similar outcomes for individuals who are similar with respect to the target variable, regardless of protected attributes. Which fairness metric addresses this?

Accepted Answer

Individual fairness. Individual fairness requires that similar individuals receive similar predictions — that is, the model treats people who are alike with respect to the target variable in the same way, regardless of protected attributes. This matches the bank's requirement to predict similar outcomes for similar individuals irrespective of protected characteristics.

Answer

Demographic parity

Answer

Calibration

Answer

Equalised odds

Question 9

Refer to the exhibit. What is the most likely issue and what action should be taken?

Accepted Answer

Overfitting; apply early stopping around epoch 15. The training loss continues to decrease while the validation loss starts to increase after approximately epoch 15, which is a classic sign of overfitting. The model is memorizing the training data rather than generalizing, so applying early stopping around epoch 15 would prevent further divergence and preserve the best validation performance.

Answer

Learning rate is too low; increase it

Answer

Underfitting; increase model complexity

Answer

Data imbalance; use class weights

Question 10

Under the EU AI Act, an AI system that uses subliminal techniques to materially distort a person's behaviour, causing psychological or physical harm, would be classified under which risk tier?

Accepted Answer

Unacceptable risk. Under the EU AI Act, AI systems that deploy subliminal, purposefully manipulative, or deceptive techniques to materially distort a person's behaviour and cause significant harm are explicitly banned. These practices are considered a clear threat to fundamental rights and are therefore classified as unacceptable risk. The Act's Article 5 prohibits such systems, placing them in the highest risk tier with outright bans.

Answer

Limited risk

Answer

High risk

Answer

Minimal risk

Question 11

A team fine-tunes a 7B parameter LLM using LoRA on a custom instruction dataset. After training, they observe that the model's outputs are only marginally different from the base model. Which is the MOST likely cause?

Accepted Answer

The LoRA rank was set too low (e.g., r=1), limiting the adapter's capacity to learn the task. LoRA (Low-Rank Adaptation) injects trainable low-rank matrices into the model's attention layers. The rank r determines the dimension of these matrices and thus the adapter's capacity to capture task-specific patterns. With r=1, the update matrices are extremely low-rank, severely restricting the number of parameters that can be tuned and limiting the model's ability to learn complex instruction-following behavior. As a result, the fine-tuned model's outputs remain very close to the base model, as observed.

Answer

The dataset contained too many examples, overfitting the adapter

Answer

The base model was too small to benefit from fine-tuning

Answer

The learning rate was too high, causing the model to diverge

Question 12

A company is building a recommendation system for an e-commerce site. They have historical user-item interaction data. Which approach is most appropriate?

Accepted Answer

Train a collaborative filtering model on user-item interactions. Collaborative filtering is the standard approach for recommendation systems when historical user-item interaction data (e.g., ratings, purchases, clicks) is available. It leverages patterns across users and items—such as 'users who liked X also liked Y'—to generate personalized recommendations without requiring explicit item features. Training a collaborative filtering model directly on this interaction matrix captures latent preferences and produces relevant suggestions, making it the most appropriate choice for the described scenario.

Answer

Use a large language model to generate random product suggestions

Answer

Use a pre-trained image classification model to recommend visually similar products

Answer

Deploy a rule-based system that always recommends best-selling items

Question 13

An AI model achieves high accuracy on training data but performs poorly on new test data. The data scientist suspects the model has memorized noise. Which technique directly adds a penalty term to the loss function to address this?

Accepted Answer

L2 regularization. L2 regularization (also known as weight decay) directly adds a penalty term proportional to the squared magnitude of the model's weights to the loss function. This discourages the model from fitting the noise in the training data by keeping weights small, thereby reducing overfitting and improving generalization to new test data.

Answer

Batch normalization

Answer

Data augmentation

Answer

Dropout

Question 14

A company is developing an AI policy. Which of the following should be included to ensure accountability for AI-driven decisions?

Accepted Answer

Designated roles for human oversight and decision authority. Accountability for AI-driven decisions requires clear assignment of human roles with oversight and decision authority. This ensures that there is a responsible party who can review, override, or be held liable for the AI's outputs, which is a core principle of AI governance frameworks such as the NIST AI Risk Management Framework.

Answer

A set of acceptable use cases for the AI

Answer

A description of the model architecture used

Answer

A list of approved training data sources

Question 15

A machine learning engineer is deploying a real-time anomaly detection system for manufacturing sensor data. The system must process thousands of readings per second with minimal latency. Which deployment architecture is BEST suited?

Accepted Answer

AI microservices with an async processing queue and streaming responses. Real-time anomaly detection on high-throughput sensor streams requires low-latency, scalable, event-driven processing. An AI microservices architecture with an async processing queue and streaming responses decouples ingestion from inference, allowing horizontal scaling of model-serving instances and back-pressure handling. This design keeps latency low while processing thousands of readings per second, unlike batch or monolithic approaches.

Answer

Batch processing using Apache Spark jobs triggered hourly

Answer

Serverless functions deployed on a CDN

Answer

A monolithic web application with a relational database