50 IBM A1000-108 - Assessment: Foundations of AI and Machine Learning Practice Questions: Question Bank 2025

A hospital wants an AI system that helps radiologists by highlighting potentially abnormal regions in X-ray images, but the final diagnosis must remain with the clinician. Which description best matches this approach?

A team is building a model to predict whether a customer will churn (yes/no). Which type of machine learning problem is this?

During data preparation, a dataset contains missing values in a numeric column used for model training. Which approach is generally an appropriate first step to consider?

A bank deploys a loan approval model and wants to ensure decisions can be explained to customers and regulators. Which responsible AI principle is most directly addressed?

A retailer has 1 million product reviews and wants to automatically tag reviews as positive, negative, or neutral. They have a subset of reviews already labeled by humans. Which approach is most suitable?

A team reports 98% accuracy for a fraud detection model, but fraud is extremely rare (0.5% of transactions). Which evaluation metric would be most helpful to understand model usefulness on the minority class?

A model performs very well on the training dataset but significantly worse on new, unseen data. Which situation is most likely occurring, and what is a common remedy?

A dataset includes a 'customer_id' column that uniquely identifies each person. The team includes it as a feature and sees a large performance boost during validation, but performance collapses after deployment. What is the most likely issue?

A company deploys a model to screen job applicants. After launch, audits reveal that qualified candidates from a protected group are rejected at a higher rate than others, even when they have similar qualifications. Which action best addresses this issue in a responsible AI program?

A data scientist accidentally trained and tuned a model using the test set multiple times to choose hyperparameters, then reported the final score as the expected real-world performance. What is the best corrective approach to obtain a reliable performance estimate?

A team is building an AI system to triage customer support emails into categories (billing, technical issue, account access). They have 5,000 previously labeled emails. Which approach is most appropriate to start with?

A dataset contains customer ages with some missing values. Which imputation strategy is generally most appropriate for a foundational baseline model when age is roughly symmetric and has few outliers?

Which statement best describes why a separate validation set is used during model development?

A bank trains a loan-approval model. During deployment, approval rates for a protected group drop significantly compared to training-time evaluation. What is the best first step to investigate the issue?

A retail team evaluates a binary classifier for predicting whether a customer will churn. Only 2% of customers churn. Which metric is typically more informative than accuracy for this imbalanced dataset?

A data scientist standardizes features (zero mean, unit variance) before training a model. What is the correct best practice to avoid data leakage?

A company is deploying a conversational assistant. Stakeholders want transparency about how the assistant behaves and its limitations. Which artifact most directly supports this goal as a Responsible AI practice?

A team trains a model on customer data where the target variable is influenced by past business decisions (e.g., who was offered discounts previously). What risk does this introduce, and what is a reasonable mitigation?

A company wants to use a pre-trained language model for internal document Q&A. Some documents contain sensitive personal information. Which combination best addresses privacy while still enabling useful retrieval-based responses?

A model shows excellent performance on the training set but substantially worse performance on the validation set. The features are many and highly correlated, and the model is complex. Which action is most likely to improve generalization?

A team is new to AI and asks for a simple definition. Which statement best describes the relationship between AI, machine learning (ML), and deep learning?

A data scientist builds a supervised model to predict loan default. Which dataset element is the model learning to predict?

A dataset contains a numeric column "age" with some missing values. Which approach is generally a reasonable baseline for handling missing ages before training a model?

A retailer trains a model to classify customer support tickets, but performance drops sharply after deployment. Investigation shows product categories were renamed in the source system. What is the most likely issue?

A team wants to reduce the risk of overfitting in a classification model. Which evaluation practice best supports this goal?

A project involves combining customer records from multiple sources. Some customers appear multiple times with slight variations ("Jon Smith" vs "Jonathan Smith"). What data preparation step best addresses this issue?

A team is building a model for hiring recommendations. Which action most directly supports the principle of transparency/explainability to affected users?

A model shows 95% accuracy on a dataset where only 5% of cases are positive (e.g., fraud). However, it rarely detects actual fraud. What metric is typically more informative for this situation?

A team performs feature scaling (e.g., standardization) before training. After deployment, predictions are inconsistent even though the model file is correct. What is the most likely root cause?

A healthcare organization trains a model to predict readmission risk. They want to minimize privacy risk while still enabling model improvement over time. Which approach best aligns with responsible AI and privacy-by-design?

A retailer wants to predict next week’s demand for each product using historical sales and promotions. The target value is a numeric quantity. Which machine learning task best fits this requirement?

A dataset contains customer ages, but some records have missing values. Which approach is generally a reasonable first step before model training when ages are missing at random?

A team is asked to explain to non-technical stakeholders what 'overfitting' means. Which description is most accurate?

An organization wants to ensure its AI system decisions can be reviewed and traced back to the data and model version used. Which principle or practice best supports this requirement?

A binary classifier shows 95% accuracy. However, only 2% of customers actually churn, and the model rarely predicts churn. Which evaluation metric is most useful to understand how well the model identifies churners?

A data scientist trains a model using a single table that accidentally includes a column called 'refunded_amount' that is only populated after a purchase is completed. The model’s performance looks unrealistically high. What is the most likely issue?

A company wants to publish an internal dataset for analytics while reducing the risk of identifying individuals. Which technique is most aligned with this goal?

A team is building a multi-class image classifier. During training, validation loss decreases at first but then starts increasing while training loss keeps decreasing. Which action is a common and appropriate response?

A bank uses an AI model for loan approvals. An internal audit finds that applicants from a protected group are approved at a significantly lower rate than others, even after controlling for income. What is the most appropriate next step in a responsible AI workflow?

A team must design a training/validation/test split for time-series forecasting (weekly demand). The dataset spans three years. Which splitting approach best avoids information leakage and provides a realistic estimate of future performance?

A retail team wants to use AI to group customers into segments based on purchase behavior, without having predefined labels. Which type of machine learning task best fits this requirement?

A project team is building a model to predict loan default. They notice that the dataset contains duplicate customer records with conflicting labels (some duplicates marked default, others non-default). What is the BEST next step before training?

In the context of responsible AI, which practice most directly supports the principle of accountability for an AI system’s outcomes?

A manufacturing company wants to detect faulty products on an assembly line using images. Only 2% of products are defective, and missing a defect is costly. Which metric is typically MOST important to prioritize?

A data scientist is preparing a dataset with a 'Country' column containing 100+ unique values for a classification model. Which encoding approach is typically the BEST starting point for many linear models?

A team reports excellent performance during training but significantly worse results on unseen data. Which issue is MOST likely, and what is the best conceptual remedy?

A company is deploying an AI model that recommends employee promotions. Which action BEST reduces the risk of unfair bias affecting protected groups?

During data preparation, a team standardizes numeric features (mean=0, std=1) before splitting into train and test sets. After deployment, performance is unstable. What is the MOST likely problem with their approach?

A team is building a churn model and discovers that the feature set includes a field named 'CancellationProcessedDate' which is populated only after a customer cancels. The model shows extremely high AUC. What is the BEST explanation and action?

An organization must provide understandable reasons for automated credit decisions to comply with internal governance. They are choosing between a complex black-box model and a simpler interpretable model with slightly lower accuracy. What is the BEST decision approach?