IBM Assessment: Foundations of AI Practice Exam 2025: Latest Questions

A customer support team wants an AI system that can answer common questions using a predefined set of support articles and must provide a short explanation with the source passage. Which approach best fits this requirement?

Which statement best describes the difference between Artificial Intelligence (AI) and Machine Learning (ML)?

A team is building a model to predict whether an email is spam (spam vs not spam). What type of machine learning problem is this?

An AI product team wants to reduce the risk of exposing customer personal data when developing and testing models. Which practice is most appropriate?

A retailer trained a demand-forecasting model on last year's data. After a major change in customer behavior, forecast accuracy drops significantly. What is the most likely issue and a reasonable first step?

A team reports 98% accuracy for a model that detects fraudulent transactions, but fraud is extremely rare. Which additional evaluation approach is most appropriate to confirm the model’s usefulness?

A developer is integrating an external foundation model API into an app. The app occasionally receives transient errors and timeouts. Which implementation choice is the best practice for improving reliability?

A bank wants to use AI to assist loan approvals. Regulators require the bank to provide understandable reasons for adverse decisions and to document model behavior. Which approach best meets this requirement?

A team builds a chatbot that can call internal tools (e.g., create support tickets, issue refunds). During testing, the chatbot follows a user prompt that tricks it into issuing an unauthorized refund. Which control is most effective to mitigate this class of risk?

An HR team trains a resume-screening model using historical hiring decisions. After deployment, audits show the model disadvantages a protected group even when qualifications are similar. What is the most appropriate next step?

A customer support team wants to automatically group incoming emails into categories (billing, technical issue, cancellation) but has very few labeled examples. Which approach is the best starting point to discover natural groupings in the data?

In a supervised learning project, a data scientist evaluates a classifier and sees high training accuracy but significantly lower validation accuracy. What is the most likely issue?

A product team is deciding whether a task should use AI or simple business rules. Which situation is the best fit for a traditional rules-based approach rather than AI?

A retailer trained a model to predict churn. After deployment, the model’s performance degrades because customer behavior changes over time. Which concept best describes this issue?

A team building an AI-powered document search wants to compare semantic similarity between queries and documents. Which representation is most appropriate for enabling semantic similarity comparisons?

A developer integrates a hosted foundation model into an app. Users report that the model sometimes includes personal data that appeared in earlier prompts within the same session. Which best practice should the developer apply to reduce this risk?

A team evaluates a binary classifier for disease detection. False negatives are much more costly than false positives. Which metric is most appropriate to prioritize during model selection?

An organization wants to deploy an AI assistant that answers questions using internal policies. They need to reduce hallucinations and ensure answers are grounded in approved documents. Which architecture pattern best addresses this?

A bank uses an AI model to help decide loan approvals. Regulators require the bank to provide understandable reasons for adverse decisions to customers. Which approach best supports this requirement?

A team notices that their language model evaluation results are unexpectedly high. Investigation reveals that test set examples were accidentally included in the training data pipeline. What is the correct diagnosis and immediate corrective action?

A support team wants to automatically categorize incoming customer emails into topics (e.g., billing, technical issue, cancellation) to route them to the right queue. They have a labeled dataset of past emails with the correct topic. Which machine learning approach best fits this task?

A product owner asks for a quick demo that answers questions over internal policy documents. The team wants to minimize custom model training and prefers a pattern that can cite relevant passages from the source documents. Which solution approach is most appropriate?

A team builds a binary classifier for a rare disease screening tool. Only 1% of patients in the dataset actually have the disease. The model reports 99% accuracy, but clinicians suspect it is not useful. What is the most likely issue and best next step?

A model used for loan pre-qualification shows a significantly higher rejection rate for one demographic group. The team wants a practical first step to investigate and reduce potential unfairness while keeping the process auditable. What should they do first?

A team builds a chatbot that uses an LLM to generate answers. During testing, the bot sometimes fabricates policy details that are not in the source material. The team has already added document retrieval, but the issue persists. Which additional control is most effective to reduce the risk of these hallucinations in production?