AI Associate AI Fundamentals Practice Test 11 — 15 Questions

Question 1

A data scientist notices that a sentiment analysis model performs well on general product reviews but fails to correctly classify negative sentiment in industry-specific jargon (e.g., 'the API is flaky'). The most likely cause is:

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

The training data lacked examples of industry-specific language. The model was trained on general product reviews and therefore never learned the vocabulary and phrasing of industry-specific jargon like 'the API is flaky.' Sentiment models rely on learned lexical and contextual patterns; without domain examples, terms that carry negative sentiment in a technical context are treated as neutral or unknown. This is a classic domain-shift / out-of-distribution vocabulary problem.

Answer

The model is overfitted to training data

Answer

The model architecture is too simple

Answer

Sentiment analysis cannot handle jargon

Question 2

What does the term 'hallucination' refer to in the context of generative AI?

Accepted Answer

The model produces outputs that are not grounded in the training data or reality. In generative AI, hallucination refers to the model producing fluent, confident output that is not grounded in its training data or factual reality — for example, inventing citations, statistics, or events. It is a byproduct of probabilistic text generation, not a training bug or a keyword dependency. This is the standard definition tested on AI fundamentals exams.

Answer

The model only works when the user prompts it with specific keywords

Answer

The model has a bug in the training algorithm

Answer

The model generates creative but accurate content

Question 3

A company is deploying an AI model to automatically classify customer emails into categories (Complaint, Inquiry, Feedback). They have 10,000 labeled emails. Which TWO actions are essential to ensure the model's accuracy? (Select TWO.)

Accepted Answer

Split the data into training and test sets. Option B is correct because splitting the 10,000 labeled emails into training and test (or validation) sets is essential to evaluate the model on data it has not seen, which detects overfitting and gives an honest estimate of generalization accuracy. Option E is correct because the three target categories (Complaint, Inquiry, Feedback) must each be adequately represented in the training data; a skewed class distribution biases the classifier toward majority classes and degrades per-class accuracy, so balanced representation (or appropriate class weighting/resampling) is needed. Option A is not correct because driving training accuracy to 100% typically means overfitting, which harms test accuracy rather than ensuring it. Option C is not correct because discarding 9,000 labeled examples and training on only the most recent 1,000 reduces data volume and can introduce recency bias, weakening the model. Option D is not correct because removing all emails with spelling errors shrinks and skews the dataset and removes real-world noise the model must handle; text normalization or spell correction is preferable to deletion.

Answer

Train the model until it achieves 100% accuracy on the training set

Answer

Use only the most recent 1,000 emails for training to reflect current trends

Answer

Remove all emails that contain spelling errors

Question 4

A financial services firm wants to use AI to detect fraudulent transactions. They have a dataset with 1% fraudulent and 99% legitimate transactions. Which THREE actions should they take to address class imbalance? (Select THREE.)

Accepted Answer

Use F1 score as the evaluation metric. Option C is correct because F1 score balances precision and recall, making it far more informative than accuracy when the positive (fraud) class is only 1% of the data; a model predicting all legitimate transactions would score 99% accuracy but near-zero F1. Option D is correct because assigning higher class weights to the minority class during training penalizes misclassification of fraudulent transactions more heavily, forcing the model to learn the minority pattern without altering the dataset. Option E is correct because oversampling the minority class (e.g., SMOTE or random duplication) or undersampling the majority class rebalances the class distribution so the learning algorithm sees a more representative mix of fraud and legitimate cases. Option A is not appropriate because a larger but even more imbalanced dataset typically worsens the imbalance problem and does not by itself correct skewed class priors. Option B is not appropriate because accuracy is misleading on imbalanced data, as a trivial majority-class classifier achieves 99% accuracy while detecting no fraud.

Answer

Use a larger dataset even if it is more imbalanced

Answer

Use accuracy as the primary evaluation metric

Question 5

A retail company wants to use AI to predict next month's sales for each product category. They have five years of monthly sales data. Which THREE factors are most critical for the accuracy of the predictive model? (Select THREE.)

Accepted Answer

Including seasonality and trend components from historical data. Option B is correct because retail sales data almost always exhibits recurring seasonal peaks (e.g., holidays, back-to-school) and long-term trends, so explicitly modeling seasonality and trend components (as in SARIMA, Prophet, or seasonal decomposition) materially improves forecast accuracy. Option D is correct because missing values, outliers, and inconsistent records in five years of monthly data bias parameter estimation and can cause the model to learn spurious patterns, so cleaning and imputing the training data is a prerequisite for reliable predictions. Option E is correct because sales are driven by exogenous events such as holiday calendars, promotions, and markdowns; incorporating these as external regressors (e.g., in Prophet or SARIMAX) captures demand shifts that pure historical autocorrelation cannot explain. Option A does not belong because algorithm popularity is irrelevant to accuracy; the right choice depends on data characteristics, and a trendy algorithm can underperform a simple, well-tuned one. Option C does not belong because deep learning is not inherently more accurate for a small tabular dataset of five years of monthly sales, and simpler models often generalize better with limited data.

Answer

Choosing the most popular AI algorithm on the market

Answer

Using a deep learning model instead of a simpler model

Question 6

A marketing team wants to use AI to predict which leads are most likely to convert. The CRM contains historical lead data with conversion outcomes. Which type of machine learning should be used?

Accepted Answer

Supervised learning. Supervised learning uses labeled data (historical outcomes) to predict future outcomes, making it ideal for lead scoring.

Answer

Unsupervised learning

Answer

Generative AI

Answer

Reinforcement learning

Question 7

A company uses an AI model to classify customer support cases into categories. The model often misclassifies cases from a specific region, leading to longer resolution times. What is the most likely cause?

Accepted Answer

The training data lacks diversity for that region. When a model performs well overall but consistently misclassifies cases from a specific region, the most likely cause is that the training data lacks sufficient representative examples from that region. The model has not learned the patterns, language, or characteristics unique to that region's support cases, leading to systematic errors. This is a data diversity/representation problem, not a model architecture issue.

Answer

The model is underfitted

Answer

The model uses too many features

Answer

The model is overfitted

Question 8

A sales manager wants to predict which deals are likely to close this quarter. The CRM has rich historical data on won/lost opportunities, deal amount, and sales stage. Which AI approach is best suited for this task?

Accepted Answer

Predictive AI for opportunity scoring. Predictive AI is designed to analyze historical data and forecast outcomes, making it ideal for scoring opportunities based on patterns from past won/lost deals. It uses supervised learning models trained on labeled data (e.g., deal amount, sales stage, historical outcomes) to output a probability of closing. This directly addresses the sales manager's need to prioritize deals likely to close this quarter.

Answer

Generative AI to create new sales content

Answer

Unsupervised learning to cluster opportunities

Answer

Reinforcement learning for sales strategy

Question 9

A data scientist trains a model to predict customer churn. The model achieves 98% accuracy on training data but only 72% on test data. What issue is most likely occurring?

Accepted Answer

Overfitting. A large gap between training performance (98%) and test performance (72%) is the classic signature of overfitting: the model memorized training noise and patterns that do not generalize. The model has high variance, fitting the training set too closely while failing on unseen data. Remedies include regularization, dropout, pruning, more training data, or simpler model architectures.

Answer

Data leakage

Answer

Underfitting

Answer

Bias in the training data

Question 10

A customer service team wants to automatically route incoming emails to the appropriate department based on content. Which NLP capability is essential for this task?

Accepted Answer

Text classification. Text classification categorizes text into predefined classes, enabling automated routing.

Answer

Sentiment analysis

Answer

Entity extraction

Answer

Intent detection

Question 11

An e-commerce company uses AI to provide product recommendations. The model suggests popular items but fails to personalize for individual users. Which type of learning could improve personalization?

Accepted Answer

Supervised learning. Supervised learning is correct because personalization is fundamentally a prediction problem: given a user's past behavior (features), predict what they will engage with or buy (label). The model is trained on labeled historical data — clicks, purchases, ratings — mapping user/item features to known outcomes, which is exactly the supervised paradigm. Popularity-based recommendations ignore user-specific signals; supervised models learn per-user preferences from those labeled interactions, enabling individualized recommendations.

Answer

Generative AI

Answer

Unsupervised learning

Answer

Reinforcement learning

Question 12

A model predicts customer lifetime value with high precision but low recall on high-value customers. What is the business impact?

Accepted Answer

Many false negatives for high-value customers. Low recall means many high-value customers are missed, leading to lost opportunities to retain them.

Answer

The model is overfitted

Answer

Many false positives for high-value customers

Answer

High accuracy on low-value customers

Question 13

A financial services firm uses an AI model to approve loan applications. They discover the model denies loans at a higher rate for a protected demographic. What is the most likely root cause?

Accepted Answer

The training data contains historical bias. Historical bias in training data can cause models to learn and perpetuate discrimination.

Answer

The model is overfitted

Answer

The model uses too few features

Answer

The model has low precision

Question 14

A generative AI chatbot sometimes produces factually incorrect responses about a company's products. What is this phenomenon called?

Accepted Answer

Hallucination. Hallucination is the correct term for a generative AI model producing fluent but factually incorrect or fabricated output, such as inventing product features or specifications that do not exist. It occurs because LLMs generate statistically plausible token sequences rather than retrieving verified facts, so they can confidently state falsehoods. This is a well-known limitation of generative models and a key concern for enterprise deployments.

Answer

Data leakage

Answer

Overfitting

Answer

Bias

Question 15

A company uses computer vision to scan receipts for expense reporting. The model performs well on high-resolution scans but poorly on blurry photos. Which improvement is most effective?

Accepted Answer

Add blurry images to the training data. Adding blurry images to the training data directly addresses the distribution mismatch between the training set (high-resolution scans) and the real-world inference inputs (blurry photos). This is data augmentation / domain adaptation, and it teaches the model to generalize to the degraded inputs it will actually see.

Answer

Decrease the learning rate

Answer

Use a larger batch size

Answer

Increase the model's number of layers