Fundamentals of gen AI — Gen AI Leader Practice Quiz

Question 1

A retail company deploys a generative AI assistant trained largely on historical customer support transcripts from one region. After launch, leaders notice the assistant frequently uses idioms and cultural references unfamiliar to customers in other regions, causing confusion. Which responsible AI concern does this situation most directly illustrate?

Accepted Answer

Bias, because skewed training data caused outputs that do not serve the broader customer base fairly

Answer

Hallucination, because the model invented references that do not exist

Answer

Privacy, because customer transcripts were used in training

Answer

Security, because the model was manipulated by malicious input

Question 2

A retail company deploys a customer-service chatbot built on a large language model. During testing, an employee asks about the return policy, and the chatbot confidently states that customers have 90 days to return items — but the actual company policy is 30 days, and this number appears nowhere in the model's training data or prompt. Which gen AI concept does this behavior best illustrate?

Accepted Answer

Hallucination — the model generated fluent but factually incorrect information not grounded in any source

Answer

A context window limitation that truncated the policy document

Answer

Embedding drift caused by outdated vector representations

Answer

Prompt injection introduced by the testing employee

Question 3

A product team is building a customer-support assistant on a large language model. During long conversations, the assistant starts to 'forget' details the customer mentioned near the beginning of the chat. An engineer explains this is due to a limit on how much text the model can consider at once. Which core gen AI concept is the engineer describing?

Accepted Answer

The context window — the maximum amount of input and output text (measured in tokens) the model can process in a single interaction

Answer

An embedding — a numeric vector representation of text used to measure semantic similarity

Answer

A hallucination — content the model generates that is fluent but factually incorrect

Answer

A prompt template — a reusable structure for formatting instructions sent to the model

Question 4

A product analyst pastes a very long transcript into a gen AI assistant and asks for a summary. The tool responds but appears to have ignored the earliest parts of the transcript, summarizing only the later sections. What is the most likely cause?

Accepted Answer

The input exceeded the model's context window, so the earliest content was pushed out and not processed.

Answer

The model hallucinated because the transcript contained factual errors.

Answer

The embeddings for the early text were corrupted during upload.

Answer

The model applied bias mitigation that removed the opening sections.

Question 5

A product team is explaining to executives how their new gen AI feature can group customer feedback by meaning rather than exact keywords. They want to describe the underlying concept that lets the system understand that 'the app keeps crashing' and 'it freezes constantly' are semantically similar. Which core generative AI concept are they describing?

Accepted Answer

Embeddings, which represent text as numeric vectors so that similar meanings are positioned close together in a shared space

Answer

Tokens, which split text into smaller units that the model processes one at a time

Answer

Context windows, which define how much text the model can consider at once

Answer

Hallucinations, which occur when the model produces plausible but incorrect output

Question 6

A financial services company wants to build an internal tool that lets employees search a large knowledge base using natural language. When someone asks 'How do I reset a locked customer account?', the tool should return documents that are conceptually related even if they use different wording, such as 'unlock a client profile.' Which generative AI concept most directly enables this semantic matching between the query and the documents?

Accepted Answer

Embeddings that represent text as numerical vectors capturing semantic meaning

Answer

A larger context window that lets the model process more tokens at once

Answer

Prompt engineering that rephrases the user's question multiple times

Answer

Reducing hallucinations by lowering the model's temperature setting

Question 7

A retail company wants a single AI system that its teams can reuse for several distinct tasks: drafting product descriptions, answering customer emails, and generating internal training summaries. Leadership is deciding whether to build separate specialized models for each task or adopt one large model that can be adapted to all of them. Which characteristic of foundation models best supports choosing the single-model approach?

Accepted Answer

Foundation models are trained on broad, large-scale data and can be adapted to many downstream tasks through prompting or fine-tuning.

Answer

Foundation models guarantee factually accurate outputs, eliminating the need for human review across tasks.

Answer

Foundation models are purpose-built for one narrow task and must be retrained from scratch for each new use case.

Answer

Foundation models only process numerical data, making them ideal for structured retail inventory tasks.

Question 8

A software company wants to build several internal tools: a document summarizer, a code-comment generator, and a customer-email drafter. Their data science lead suggests starting from a single large pre-trained model rather than building three separate systems from scratch. What key property of foundation models makes this approach practical?

Accepted Answer

Foundation models are trained on broad data at scale and can be adapted to many different downstream tasks

Answer

Foundation models are trained on a specific labeled dataset for one narrow task and cannot be reused elsewhere

Answer

Foundation models eliminate the need for prompts because they infer intent automatically without any input

Answer

Foundation models guarantee factually accurate outputs, removing the need for human review across all tasks

Question 9

A product manager is explaining to executives why the company's new generative AI assistant, built on a foundation model, can handle a wide range of tasks it was never explicitly programmed for. Which characteristic of foundation models best explains this broad capability?

Accepted Answer

They are trained on massive, broad datasets in a self-supervised way, allowing them to generalize across many downstream tasks.

Answer

They are hand-coded with explicit rules for each individual business task they need to perform.

Answer

They are trained on a small, narrowly labeled dataset specific to a single use case.

Answer

They store every possible answer in a lookup database and retrieve exact matches at runtime.

Question 10

A retail company's leadership team is evaluating new technology and asks their innovation lead to explain what fundamentally distinguishes generative AI from the analytics tools they already use. Which description best captures the defining capability of generative AI?

Accepted Answer

It creates new, original content such as text, images, or code based on patterns learned from training data

Answer

It classifies existing records into predefined categories with high accuracy

Answer

It stores and retrieves structured business data faster than a traditional database

Answer

It automates repetitive rule-based tasks by following explicitly coded if-then logic

Question 11

A product manager at a media company is explaining to executives how the new generative AI writing assistant fundamentally works, compared to the company's older spam-detection system. Which statement best captures the defining behavior of the generative AI tool?

Accepted Answer

It learns statistical patterns from large amounts of text and produces new, original content in response to a prompt.

Answer

It assigns each incoming item to one of a fixed set of predefined categories based on labeled training examples.

Answer

It stores every document it has seen and retrieves the closest exact match when asked a question.

Answer

It calculates a single numeric probability score to predict whether a future event will occur.

Question 12

A retail operations manager wants a system that predicts next quarter's inventory demand for each store based on years of historical sales, seasonality, and pricing data, outputting a numerical demand figure. A separate team wants a tool to automatically draft product descriptions for new items. Which statement correctly distinguishes the appropriate approach for each need?

Accepted Answer

The demand forecast is best served by traditional/predictive ML, while the product descriptions are a fit for generative AI.

Answer

Both needs should use generative AI because it is the most advanced form of machine learning available.

Answer

The demand forecast should use generative AI, while the product descriptions should use traditional/predictive ML.

Answer

Both needs should use traditional/predictive ML because neither involves conversation.

Question 13

A bank's data science team already uses a model that scores incoming credit card transactions and flags each one as 'fraudulent' or 'legitimate' based on historical labeled data. A business leader now asks whether generative AI would be a better fit for this exact task. What is the most accurate response?

Accepted Answer

The current fraud-flagging task is a classification problem best served by traditional predictive ML, since it assigns transactions to predefined categories rather than creating new content.

Answer

Generative AI is always superior because it is a newer technology and can replace any predictive model with better accuracy.

Answer

Generative AI should be used because it does not require any training data, unlike predictive ML.

Answer

The task should switch to generative AI because generative models produce category labels more efficiently than classifiers.

Question 14

A retail operations manager wants to reduce stockouts by predicting how many units of each product the store will sell next month, based on years of historical sales data, seasonality, and promotions. A colleague suggests using a generative AI chatbot instead. Which statement best explains the most appropriate approach?

Accepted Answer

Traditional predictive machine learning is the better fit, because the goal is to forecast a numeric quantity from structured historical patterns.

Answer

Generative AI is the better fit, because it can create new sales data that never existed to fill in the forecast.

Answer

Generative AI is the better fit, because forecasting is a content-creation task like writing product descriptions.

Answer

Either approach is equivalent, since both generative and predictive models are trained on data and output numbers.

Question 15

A retail company's data science team already uses a machine learning model to predict which customers are likely to churn based on historical purchase data. The marketing director now wants a new capability: automatically producing original, on-brand promotional email copy for each product launch. Which statement best explains why a generative AI solution—rather than the existing predictive ML approach—is appropriate for this new task?

Accepted Answer

Generative AI creates new content such as text, whereas the predictive model outputs classifications or scores from patterns in existing data.

Answer

Generative AI is always more accurate than predictive ML, so it should replace the churn model as well.

Answer

Predictive ML cannot process text data at all, so any text-related task must use generative AI.

Answer

Generative AI does not require any training data, making it faster to deploy for the marketing use case.

Question 16

A retail analytics team is debating two proposed projects. Project 1 will analyze historical sales data to estimate the probability that a given customer will churn next month. Project 2 will automatically draft personalized product description paragraphs for thousands of new catalog items. Which statement correctly characterizes these projects with respect to generative AI?

Accepted Answer

Project 1 is a good fit for traditional predictive ML, while Project 2 is a good fit for generative AI.

Answer

Both projects require generative AI because both produce a business output.

Answer

Project 1 requires generative AI because probabilities are newly created content, while Project 2 is predictive.

Answer

Neither project can use AI because both rely on historical data rather than real-time input.

Question 17

A marketing manager asks a general-purpose large language model to write a company history section for a website. The model produces a fluent, confident paragraph that includes a specific founding date, an award, and a founder quote — none of which the company ever provided or that actually exist. Which core generative AI concept best explains why this happened?

Accepted Answer

The model hallucinated, generating plausible-sounding but factually unsupported content because it predicts likely text rather than retrieving verified facts

Answer

The model exceeded its context window and truncated the correct information

Answer

The model's embeddings were misaligned, causing it to misclassify the input category

Answer

The temperature setting was too low, forcing the model to repeat memorized training data verbatim

Question 18

A product manager asks a general-purpose LLM to describe the technical specifications of a newly released internal device. The model produces a fluent, confident answer listing battery life, weight, and processor details — but none of these facts exist in any source, and the device was never in the model's training data. What is this behavior called?

Accepted Answer

A hallucination, where the model generates plausible-sounding but fabricated information

Answer

A context window overflow that dropped the correct answer

Answer

A tokenization error that corrupted the model's numeric output

Answer

An embedding mismatch between the query and the stored knowledge base

Question 19

A law firm uses a large language model to draft internal case summaries. During review, an associate discovers the model produced a summary citing a court case that does not actually exist, though it was worded with complete confidence. The managing partner asks the AI leader to explain what happened and how to reduce the chance of this recurring. What is the best explanation and mitigation?

Accepted Answer

The model hallucinated by generating plausible-sounding but fabricated content; grounding responses in verified source documents and requiring human verification reduces the risk.

Answer

The model ran out of tokens in its context window, so it invented a citation; increasing the maximum output length will eliminate the problem.

Answer

The model's training data was biased against the firm; retraining it on the firm's own cases will guarantee only real citations appear.

Answer

The model experienced a prompt injection attack; adding stricter input filters will prevent any future fabricated citations.

Question 20

A product manager at a software company is briefing her team on the technology behind a new writing assistant. She wants to explain what specifically makes a large language model (LLM) distinct from other AI systems the company has used. Which statement best describes the defining characteristic of an LLM?

Accepted Answer

It is trained on massive amounts of text data and predicts the next token to generate human-like language across many tasks

Answer

It is a rules-based system where engineers manually define every grammar and vocabulary rule the model must follow

Answer

It is a model designed exclusively to classify structured numeric data into predefined categories

Answer

It is a database that stores and retrieves exact answers written in advance by human experts

More Gen AI Leader practice