AI Associate AI Fundamentals Practice Test 8 — 15 Questions

Question 1

A company’s AI model recommends products to customers. The team wants to measure how often the recommended products are actually purchased. Which metric is most appropriate?

Accepted Answer

Precision. Precision measures the proportion of recommended products that were actually purchased — i.e., of all the items the model flagged as 'recommend,' how many were true positives. That directly matches the business question of how often recommendations convert into purchases. Recall would instead measure how many of all actually purchased products were captured by the recommendations, which is a different question.

Answer

F1 score

Answer

Recall

Answer

Accuracy

Question 2

A financial services firm uses an AI model to approve loan applications. They discover that the model denies loans at a higher rate for a certain demographic group, even when financial indicators are similar. What is the primary ethical concern?

Accepted Answer

Bias in training data. The primary ethical concern is bias in training data, which leads to the model unfairly denying loans to a demographic group even when financial indicators are similar. This is a classic example of algorithmic bias, where historical data reflects past discrimination or systemic inequalities. The model learns these biases and perpetuates them, resulting in disparate impact. Addressing this requires careful data auditing and fairness-aware machine learning techniques.

Answer

Low model accuracy

Answer

Model overfitting

Answer

Hallucination

Question 3

A company uses generative AI to create personalized email content for each customer. They notice that occasionally the AI produces content that is factually incorrect. What is this phenomenon called?

Accepted Answer

Hallucination. Hallucination refers to the phenomenon where a generative AI model produces content that is factually incorrect, nonsensical, or not grounded in its training data. In the context of personalized email generation, the AI may fabricate details such as product features, prices, or customer information. This occurs because large language models predict the next word based on patterns, without an inherent understanding of truth. Thus, 'hallucination' is the correct term for factually incorrect AI-generated content.

Answer

Overfitting

Answer

Concept drift

Answer

Bias

Question 4

A CRM team wants to predict the expected revenue from each opportunity. The data includes opportunity amount, close date, stage, and historical win rates. Which type of AI is best suited?

Accepted Answer

Predictive AI. Predictive AI is correct because the task is to forecast a continuous numerical value (expected revenue) based on historical labeled data (opportunity amount, close date, stage, win rates). This is a classic regression problem, a core supervised learning technique within predictive AI. The model learns patterns from past opportunities to estimate future revenue, directly supporting the CRM team's goal.

Answer

Unsupervised learning

Answer

Generative AI

Answer

Reinforcement learning

Question 5

A data analyst is preparing data for a machine learning model. They notice that many records have missing values for the 'industry' field. What is the best first step?

Accepted Answer

Replace missing values with a placeholder or impute them. When a dataset has many missing values in a feature like 'industry,' the best first step is to handle them explicitly — either by imputing values (e.g., mode, mean, or model-based imputation) or by using a placeholder category like 'Unknown.' This preserves the maximum amount of data and avoids introducing bias from dropping records. Imputation or placeholder encoding is a standard data-preparation practice before training.

Answer

Delete all records with missing values

Answer

Ignore the missing values and train the model as is

Answer

Use a model that automatically handles missing data

Question 6

A company uses an NLP model to detect customer intent from chat messages. The model correctly identifies 'billing question' 90% of the time for actual billing questions, but also flags many non-billing messages as billing (false positives). Which metric should the team prioritize to reduce false alarms?

Accepted Answer

Precision. Precision is the metric that measures the proportion of positive identifications that are actually correct — i.e., of all messages flagged as 'billing question', how many truly are billing questions. Since the problem is false positives (non-billing messages incorrectly flagged as billing), precision directly captures this error type. High precision means fewer false alarms, which is exactly what the team wants to prioritize.

Answer

Accuracy

Answer

F1 score

Answer

Recall

Question 7

A sales director wants to implement lead scoring but has no historical data on which leads converted. What approach can the team use to start?

Accepted Answer

Use unsupervised learning to cluster leads into segments and score based on cluster characteristics. When no historical conversion labels exist, supervised learning cannot be used because there is no target variable to train on. Unsupervised learning — specifically clustering algorithms such as k-means or hierarchical clustering — can segment leads based on features like industry, engagement, and firmographics, and scores can then be assigned based on cluster characteristics. This provides a data-driven starting point without requiring labeled outcomes.

Answer

Train a supervised model using assumptions as labels

Answer

Skip lead scoring until enough conversion data is collected

Answer

Use reinforcement learning to learn scoring from sales team feedback

Question 8

A data scientist is building a churn prediction model. The dataset has 95% non-churn and 5% churn. Which THREE actions should the data scientist take to address the class imbalance?

Accepted Answer

Undersample the non-churn class. Option C (undersample the non-churn class) is correct because reducing the majority non-churn examples brings the class distribution closer to balanced, so the learning algorithm is no longer overwhelmed by the 95% majority and can better learn churn patterns. Option D (use precision, recall, or F1 score instead of accuracy) is correct because with a 95/5 split, a trivial model predicting 'non-churn' for every record achieves 95% accuracy while catching zero churners, so precision, recall, and F1 (and often AUC-PR) reflect real performance on the minority class. Option E (oversample the churn class) is correct because duplicating or synthetically generating minority churn examples (e.g., via SMOTE) increases the effective weight of the 5% class and reduces the bias toward the majority. Option A is not appropriate because training unchanged on a 95/5 dataset typically yields a model that ignores the minority class. Option B is not appropriate because accuracy is misleading under severe class imbalance and would reward the majority-class-only predictor.

Answer

Train the model on the original dataset without changes

Answer

Use accuracy as the primary metric

Question 9

A company uses a generative AI model to create marketing copy. They want to ensure the output is accurate and not misleading. Which TWO practices should they implement?

Accepted Answer

Always have a human review the generated content before publishing. Option C is correct because human-in-the-loop review is the most reliable safeguard against inaccurate or misleading generative output; a person can catch hallucinations, unsupported claims, and brand/compliance issues before publication. Option D is correct because grounding the model with a knowledge base of verified facts (retrieval-augmented generation, RAG) supplies authoritative context and citations, which measurably reduces hallucination and keeps the copy aligned with real product and legal facts. Option A is not appropriate because raising the temperature increases randomness and creativity, which raises the risk of fabricated or misleading statements. Option B is not appropriate because disabling content filtering removes safety guardrails and can allow harmful or non-compliant text through. Option E is not appropriate because publishing unchecked model output directly is exactly the practice that permits inaccuracies and misleading claims to reach customers.

Answer

Increase the model's temperature setting to encourage creativity

Answer

Disable all content filtering to avoid blocking useful output

Answer

Use the model's output directly without checking

Question 10

A sales operations manager wants to predict which leads are most likely to convert to deals. The CRM has historical data on thousands of leads with outcomes (converted or not). Which type of machine learning should they use?

Accepted Answer

Supervised learning. Supervised learning is correct because the CRM dataset contains labeled historical examples — each lead has a known outcome (converted or not) — which is the defining characteristic of supervised classification. The model learns a mapping from input features (lead attributes) to the binary target label and can then predict conversion probability for new leads. This is a classic binary classification problem solved with algorithms like logistic regression, gradient boosting, or random forests.

Answer

Unsupervised learning

Answer

Deep learning

Answer

Reinforcement learning

Question 11

A company uses an AI model to classify customer support cases into categories (billing, technical, general). The model performs well on training data but poorly on new cases. Which issue is MOST likely occurring?

Accepted Answer

Overfitting. Overfitting means the model memorized the training data and fails to generalize to new, unseen data. Underfitting would show poor performance on both training and test data. Data leakage occurs when future information leaks into training. Bias in data is a different issue.

Answer

Underfitting

Answer

Data leakage

Answer

Bias in training data

Question 12

A customer service team wants to automatically detect the intent of incoming chat messages (e.g., complaint, inquiry, purchase). Which AI technique is BEST suited for this task?

Accepted Answer

Natural language processing (NLP). Natural language processing (NLP) is specifically designed to analyze, understand, and classify human language. Intent detection is a classic NLP task, often implemented via text classification models (e.g., using transformers like BERT) that map chat messages to predefined categories such as complaint, inquiry, or purchase. Generative AI could produce text but is not optimized for classification; reinforcement learning learns from rewards, not labeled intents; computer vision handles images, not text.

Answer

Generative AI

Answer

Reinforcement learning

Answer

Computer vision

Question 13

A model predicts customer churn with 95% accuracy, but most customers who actually churn are not flagged by the model. Which metric should the team improve?

Accepted Answer

Recall. Recall measures the proportion of actual positive cases (customers who churned) that the model correctly identified. The scenario states that most actual churners are not flagged, which is precisely a low-recall problem. Improving recall will increase the number of true churners caught by the model.

Answer

Precision

Answer

Accuracy

Answer

F1 score

Question 14

A financial services firm uses an AI model to approve small business loans. The model denies loans at a much higher rate for businesses owned by minorities, even when financial indicators are similar. What is the MOST likely cause?

Accepted Answer

Bias in training data. When a model denies loans at a higher rate for minority-owned businesses despite similar financial indicators, the most likely cause is bias in the training data — historical lending data may reflect past discriminatory practices, and the model learns and perpetuates those patterns. This is a classic example of algorithmic bias stemming from biased training data.

Answer

Low recall for minority groups

Answer

Hallucination in model predictions

Answer

Overfitting to historical data

Question 15

What is the primary difference between narrow AI and general AI?

Accepted Answer

Narrow AI is designed for a specific task, while general AI can understand and reason across a wide range of tasks. Narrow AI (also called weak AI) is designed to perform a specific task or a limited set of tasks, such as image recognition or language translation. General AI (strong AI) would possess human-like general intelligence, able to understand, learn, and reason across a wide range of tasks. This distinction is the core definition tested.

Answer

Narrow AI is rule-based, while general AI uses machine learning

Answer

Narrow AI requires labeled data, while general AI can learn without data

Answer

Narrow AI can perform any intellectual task, while general AI is limited to one task