AI Fundamentals practice questions

Generative AI

Why wrong: Generative AI creates new content; it does not predict outcomes from labeled data.

Unsupervised learning

Why wrong: Unsupervised learning finds patterns without labels; it does not predict a specific outcome.

Reinforcement learning

Why wrong: Reinforcement learning learns via rewards from actions; not suited for static prediction from historical data.

The features used are irrelevant to lead conversion

Why wrong: If features were irrelevant, the model would never have performed well; the problem emerged over time.

The model is overfitting to noise in the training data

Why wrong: Overfitting would cause poor performance on new data, but the specific issue here is that old patterns are no longer valid.

The model is underfitting and needs more features

Why wrong: Underfitting would show poor performance on both training and new data; the issue is specifically with time relevance.

Intent detection

Why wrong: Intent detection is for identifying user goals in dialog, but routing often uses broader categories.

Entity extraction

Why wrong: Entity extraction identifies specific entities (names, dates), not the overall category.

Sentiment analysis

Why wrong: Sentiment analysis determines emotion, not the topic of the issue.

Concept drift

Why wrong: Concept drift affects performance over time, not a train-test split from the same time period.

Data leakage

Why wrong: Data leakage would cause overly optimistic performance, but not necessarily a large gap.

Underfitting

Why wrong: Underfitting would show low accuracy on both training and test sets.

F1 score

Why wrong: F1 is harmonic mean of precision and recall; precision alone is more direct for recommendation quality.

Recall

Why wrong: Recall measures how many relevant items were recommended, not the accuracy of recommendations.

Accuracy

Why wrong: Accuracy is overall correctness but not specific to recommended items; it can be misleading for imbalanced data.

Low model accuracy

Why wrong: Accuracy alone does not reveal fairness issues; the model could be accurate overall but biased.

Model overfitting

Why wrong: Overfitting is a technical issue, not specifically about demographic disparity.

Hallucination

Why wrong: Hallucination refers to generative AI producing false information; not applicable here.

Overfitting

Why wrong: Overfitting relates to training data, not generating new content.

Concept drift

Why wrong: Concept drift is about changes in data over time, not content generation errors.

Bias

Why wrong: Bias refers to systematic prejudice, not factual errors.

Object detection

Why wrong: Object detection locates and classifies objects but is broader; product recognition is more specific.

Image classification

Why wrong: Image classifies the whole image, but products need to be recognized individually.

Text extraction (OCR)

Why wrong: OCR extracts text, not product identity from images.

Unsupervised learning

Why wrong: Unsupervised learning finds patterns without labels; does not predict a specific value.

Generative AI

Why wrong: Generative AI creates new content, not predictions.

Reinforcement learning

Why wrong: Reinforcement learning learns via trial and error; not suited for static prediction.

Delete all records with missing values

Why wrong: Deleting may remove valuable data; imputation is often better.

Ignore the missing values and train the model as is

Why wrong: Most models cannot handle missing values; ignoring leads to errors.

Use a model that automatically handles missing data

Why wrong: While some models handle missing data, it's not a best practice; cleaning is recommended.

Accuracy

Why wrong: Accuracy may be high if billing is rare; misleading for imbalance.

F1 score

Why wrong: F1 balances precision and recall; but if false positives are the issue, precision is the direct metric.

Recall

Why wrong: Recall measures true positives among actual positives; does not address false positives.

Train a supervised model using assumptions as labels

Why wrong: Assumptions may introduce bias; unsupervised is safer.

Skip lead scoring until enough conversion data is collected

Why wrong: Business needs often require immediate solutions; unsupervised provides a starting point.

Use reinforcement learning to learn scoring from sales team feedback

Why wrong: Reinforcement learning requires an interactive environment; not practical for initial scoring.

F1 score

Why wrong: F1 is composite; the question asks for two metrics, and precision+recall are the components.

R-squared

Why wrong: For regression.

Mean squared error

Why wrong: For regression, not classification.

Train the model on the original dataset without changes

Why wrong: Model would be biased towards majority class.

Use accuracy as the primary metric

Why wrong: Accuracy is misleading for imbalanced data.

Increase the model's temperature setting to encourage creativity

Why wrong: Higher temperature increases randomness, potentially more hallucinations.

Disable all content filtering to avoid blocking useful output

Why wrong: Removing filters increases risk of harmful or inaccurate content.

Use the model's output directly without checking

Why wrong: Risks publishing incorrect content.

Unsupervised learning

Why wrong: Unsupervised learning does not use outcome labels and cannot directly predict conversion probability.

Deep learning

Why wrong: Deep learning is a subfield of machine learning; while it can be used, supervised learning is the broader correct category here.

Reinforcement learning

Why wrong: Reinforcement learning is used for sequential decision-making with rewards, not for static lead scoring.

Underfitting

Why wrong: Underfitting would cause poor performance even on training data, not just on new cases.

Data leakage

Why wrong: Data leakage would give the model unrealistic access to future information, often causing overly optimistic training performance, but not necessarily poor generalization.

Bias in training data

Why wrong: Bias in data might lead to systematic errors but not specifically the symptom of high training accuracy and low test accuracy.

Generative AI

Why wrong: Generative AI creates new content (e.g., text generation) rather than classifying intents.

Reinforcement learning

Why wrong: Reinforcement learning learns through trial and error in an environment, not for text classification.

Computer vision

Why wrong: Computer vision analyzes images, not text messages.

Precision

Why wrong: Precision would improve if fewer false positives occur, but the issue is missing true churners (false negatives).

Accuracy

Why wrong: Accuracy is already high but misleading because the model fails on the minority class.

F1 score

Why wrong: F1 is a harmonic mean of precision and recall; while it might improve, the direct issue is low recall.

Low recall for minority groups

Why wrong: Low recall would mean missing some minority applicants who should be approved, but the issue is higher denial rates across the board.

Hallucination in model predictions

Why wrong: Hallucination refers to generative AI producing false information, not relevant to loan approval models.

Overfitting to historical data

Why wrong: Overfitting might reduce generalization but doesn't inherently cause demographic bias unless the training data already contains bias.