SaveMyCert
Log in
5 of 5 free questions left today·for 10 a day
GenAI Leader · Domain 1

Fundamentals of gen AI practice questions

Fundamentals of gen AI is worth 30% of the GenAI Leader exam — the 2nd-heaviest of the 4 domains. Core generative AI concepts and use cases, data types, the gen AI landscape layers, and Google’s foundation models. Official (approximate) weighting ~30%. 6 fully worked examples are further down this page, answers included.

Exam weight
30%
the 2nd-heaviest of the 4 domains
Questions
139
across 4 topics
Free, no account
5/day
a free account allows 10 a day
Explanations
Every option
right and wrong

Build a practice session

5 free questions left today.

Domains

How many?

5 free questions left today.

Mode

Ready when you are

10 fresh questions drawn across 1 of 4 domains, in Learn mode.

Review what you missed

Every question you answer incorrectly, and every question you flag while practising, is saved here automatically. Finish a session and you can come back to re-drill just those.

6 sample Fundamentals of gen AI questions, fully explained

Questions from the GenAI Leader bank mapped to domain 1, with the answer key and the reasoning behind every option. None of them repeat the examples on the main GenAI Leader practice page.

Question 1Fundamentals of gen AI

A company evaluating gen AI for drafting replies and answering staff questions asks what a large language model (LLM) is. Which statement best describes it?

Choose one.

  • a
    A foundation model trained on very large volumes of text that can understand and generate language Correct

    An LLM is a language-specialised foundation model: trained on huge text corpora, it predicts likely next tokens, which lets it answer, draft, translate and write code.

  • b
    A diffusion model trained on image and caption pairs that generates pictures from a language prompt

    That describes an image model such as Imagen. It accepts a text prompt but its output is images, not language.

  • c
    A supervised classifier trained on labeled language data that sorts documents into fixed categories

    That is traditional NLP classification: it labels text but cannot generate new language or answer open questions.

  • d
    A rules-based language system that parses sentences with grammar rules written by linguists

    Rule-based NLP predates machine learning; it is not trained on data and is not a foundation model.

The concept

A large language model is a foundation model specialised in language. Trained on enormous text datasets, it predicts the most likely next words, which lets it understand and generate human language for tasks such as drafting, Q&A, translation and code. It builds on natural language processing (NLP), the field of AI concerned with language.

Why that’s the answer

Drafting replies and answering questions both need a model that both understands and generates language, which is what an LLM is. A text-to-image diffusion model takes language in but produces images. A supervised text classifier understands text only enough to assign a label. A rule-based parser is pre-ML NLP with no pretraining.

How to reason it out
  1. Note the outputs needed: generated text and answers.
  2. Recall that an LLM is a foundation model trained on large text corpora to understand and generate language.
  3. Reject options that only classify text, only produce images, or are not trained at all.

Exam tip: LLM = language-specialised foundation model that both understands and generates text.

Generative AI Concepts: Foundation Models, LLMs, and Business Use Cases — the lesson that teaches this.

Question 2Fundamentals of gen AI

An insurer wants one model that takes a photo of vehicle damage, the customer's recorded voice description and the typed claim form, and reasons across all three in a single request. Which model capability is the insurer relying on?

Choose one.

  • a
    Multilingual support

    Multilingual support means handling several human languages. The requirement here is several data types, not several languages.

  • b
    A large context window

    Context window is how much input fits in one request. It does not by itself let a model interpret images or audio.

  • c
    Prompt tuning

    Prompt tuning adapts a model's behaviour for a task; it does not add the ability to process new data types.

  • d
    Multimodality Correct

    A multimodal foundation model, such as Gemini, accepts and combines several data types (images, audio, text) in one model.

The concept

A multimodal foundation model understands and combines more than one data type, such as text, images, audio and video, in a single model. Gemini is natively multimodal.

Why that’s the answer

The deciding detail is that the inputs are of different types: an image, audio and text. Handling several modalities in one request is multimodality. Multilingual support is the tempting near-twin, but it concerns languages, not data types. A large context window helps with long inputs, and prompt tuning shapes behaviour; neither lets a text-only model see a photo or hear a recording.

How to reason it out
  1. List the input types: photo, audio, typed text.
  2. Recognise that multiple data types in one model means multimodality.
  3. Separate it from multilingual (languages) and context window (input size).

Exam tip: Several data types in one model = multimodal; several languages = multilingual.

Generative AI Concepts: Foundation Models, LLMs, and Business Use Cases — the lesson that teaches this.

Question 3Fundamentals of gen AI

An image model starts from random noise and refines it over many small steps into a picture that matches a text prompt, having learned during training to undo progressively added noise. Which type of generative model is this?

Choose one.

  • a
    Diffusion model Correct

    Diffusion models learn to reverse a gradual noising process and generate images by denoising step by step; Imagen is built on this approach.

  • b
    Generative adversarial network

    GANs also generate images, but by pitting a generator against a discriminator in one pass, not by iterative denoising.

  • c
    Large language model

    LLMs generate text by predicting the next token; they do not build images by removing noise.

  • d
    Variational autoencoder

    VAEs generate by decoding a sample from a learned latent space, not by many steps of reversing added noise.

The concept

A diffusion model is trained by watching images get progressively degraded into noise and learning to run that process backwards. At generation time it starts from noise and denoises step by step toward an image that matches the prompt. It is the technique behind modern image (and video) generation such as Imagen.

Why that’s the answer

Learning to undo added noise, and generating by iterative denoising, is the defining mechanism of diffusion models. GANs and VAEs are other real image-generation approaches, which makes them tempting, but neither works by many steps of noise removal. LLMs generate language, not images.

How to reason it out
  1. Find the mechanism in the stem: training to remove noise, generating by stepwise denoising.
  2. Map that mechanism to diffusion models.
  3. Rule out other generative families by how they generate: adversarial training (GAN), latent decoding (VAE), next-token prediction (LLM).

Exam tip: Noise in, step-by-step denoising out = diffusion model (Imagen, Veo).

Generative AI Concepts: Foundation Models, LLMs, and Business Use Cases — the lesson that teaches this.

Question 4Fundamentals of gen AI

A team lead hears colleagues mixing up prompt engineering and prompt tuning. Which statement correctly distinguishes them?

Choose one.

  • a
    Engineering rewrites the instructions a person sends; tuning learns soft-prompt parameters while model weights stay frozen. Correct

    Prompt engineering is a human crafting the prompt text; prompt tuning learns a small set of prompt parameters from examples and leaves the base model unchanged.

  • b
    Engineering learns soft-prompt parameters while weights stay frozen; tuning rewrites the instructions a person sends.

    This swaps the two definitions: the learned parameters belong to prompt tuning, and writing instructions is prompt engineering.

  • c
    Engineering rewrites the instructions a person sends; tuning retrains the model's own weights on labeled examples.

    Retraining the model's weights describes fine-tuning. Prompt tuning keeps the weights frozen.

  • d
    Engineering adds retrieved company documents to each prompt; tuning learns soft-prompt parameters with frozen weights.

    Adding retrieved documents to prompts is grounding / retrieval-augmented generation, not the definition of prompt engineering.

The concept

Prompt engineering is the human practice of writing better instructions (task, context, examples, format). Prompt tuning is a parameter-efficient technique that learns an optimised 'soft prompt' from examples while the foundation model's weights stay frozen. Fine-tuning goes further and updates the model's weights.

Why that’s the answer

The correct pairing is human-written instructions for prompt engineering and learned prompt parameters with frozen weights for prompt tuning. Swapping them is the most common confusion. Describing tuning as retraining all the weights confuses it with fine-tuning. Describing engineering as adding retrieved documents confuses it with grounding, which is a separate technique.

How to reason it out
  1. Ask who or what changes: a person edits text (engineering) or a system learns parameters (tuning).
  2. Check the base model: prompt tuning keeps weights frozen; fine-tuning changes them.
  3. Separate grounding (adding retrieved data) from both.

Exam tip: Engineering = people write the prompt; prompt tuning = the system learns a soft prompt, weights frozen; fine-tuning = weights change.

Generative AI Concepts: Foundation Models, LLMs, and Business Use Cases — the lesson that teaches this.

Question 5Fundamentals of gen AI

A lender trains a model on ten years of past loan applications, each record marked with whether the borrower repaid, so it can predict repayment for new applicants. Which machine learning approach is this?

Choose one.

  • a
    Supervised learning Correct

    Each training example carries the known outcome (repaid or not), so the model learns a mapping from inputs to a labeled output.

  • b
    Unsupervised learning

    Unsupervised learning works on unlabeled data to find groupings. Here every record already has the answer attached.

  • c
    Reinforcement learning

    Reinforcement learning learns from rewards for actions taken in an environment, not from a historical table of labeled outcomes.

  • d
    Self-supervised learning

    Self-supervised learning creates its own training signal from unlabeled data (for example, predicting hidden words). These records carry an explicit, human-recorded outcome label.

The concept

Supervised learning trains on labeled data: examples paired with the correct answer. Typical business uses are credit risk, fraud detection, spam filtering and forecasting.

Why that’s the answer

The repayment outcome on every record is a label, and the goal is to predict that label for new cases, which is supervised learning. Unsupervised learning would have no outcome column. Reinforcement learning needs actions and rewards. Self-supervised learning, used to pretrain LLMs, derives its signal from the raw data itself rather than from an explicit outcome label.

How to reason it out
  1. Check the data: does every example carry a known correct answer?
  2. Check the goal: predict that answer for new inputs.
  3. Labeled examples plus prediction of the label = supervised learning.

Exam tip: Known answers in the training data = supervised learning.

Generative AI Concepts: Foundation Models, LLMs, and Business Use Cases — the lesson that teaches this.

Question 6Fundamentals of gen AI

A retailer wants to discover natural groups of shoppers from purchase histories. Nobody has defined the groups in advance, and no record says which group a shopper belongs to. Which machine learning approach fits?

Choose one.

  • a
    Supervised learning

    Supervised learning needs records labeled with the correct group, which the retailer does not have.

  • b
    Reinforcement learning

    Reinforcement learning optimises actions through rewards; there is no action or reward here, only data to organise.

  • c
    Transfer learning

    Transfer learning reuses a model trained on one task for another task; it does not by itself find groups in unlabeled data.

  • d
    Unsupervised learning Correct

    Unsupervised learning finds structure, such as clusters of similar shoppers, in data that has no labels.

The concept

Unsupervised learning trains on unlabeled data and discovers hidden structure on its own: customer segmentation (clustering), anomaly detection and market-basket analysis are the classic business examples.

Why that’s the answer

No predefined groups and no labels means there is nothing to supervise against; the model must find the segments itself, which is unsupervised learning. Supervised learning would need labeled segments. Reinforcement learning needs actions and rewards. Transfer learning is about reusing a pretrained model, not about the absence of labels.

How to reason it out
  1. Check for labels: there are none.
  2. Check the goal: find groupings, not predict a known answer.
  3. Unlabeled data plus pattern discovery = unsupervised learning.

Exam tip: No labels, find the groups = unsupervised learning.

Generative AI Concepts: Foundation Models, LLMs, and Business Use Cases — the lesson that teaches this.

What GenAI Leader domain 1 tests, topic by topic

The official exam guide breaks Fundamentals of gen AI into 4 topics. The question bank follows the same split, so a weak topic shows up as a cluster of misses you can go back and read.

Published GenAI Leader practice questions per topic in Fundamentals of gen AI
TopicWhat it coversQuestions
Describe core generative AI (gen AI) concepts and use casesOfficial Gen AI Leader objective. Defining core concepts (AI, NLP, machine learning, generative AI, foundation and multimodal foundation models, diffusion models, prompt tuning, prompt engineering, large language models); describing ML approaches (supervised, unsupervised, reinforcement); the ML lifecycle stages (data ingestion, preparation, training, deployment, management) and Google Cloud tools for each; choosing a foundation model for a use case (modality, context window, security, availability, cost, performance, fine-tuning); business use cases where gen AI can create, summarize, discover, and automate.37
Describe how various data types are used in gen AI and the business implicationsOfficial Gen AI Leader objective. Explaining the characteristics and importance of data quality and data accessibility (completeness, consistency, relevance, availability, cost, format); differences between structured and unstructured data with real-world examples; differences between labeled and unlabeled data.34
Identify the core layers of the gen AI landscape and the business implicationsOfficial Gen AI Leader objective. The core layers of the gen AI landscape and their business implications: infrastructure, models, platforms, agents, and applications.34
Identify the use cases and strengths of Google's foundation modelsOfficial Gen AI Leader objective. Use cases and strengths of Google's foundation models: Gemini, Gemma, Imagen, and Veo.34
Total139

Revise Fundamentals of gen AI before you drill it

Other GenAI Leader domains

Fundamentals of gen AI: your questions

Fundamentals of gen AI is domain 1 of the GenAI Leader exam guide and carries 30% of the scored content — the 2nd-heaviest of the 4 domains. On a 55-question paper that works out to roughly 17 questions, though Google Cloud does not publish an exact per-domain count and individual exam forms vary.

Source

The domain weight and topic list on this page come from the official GenAI Leader exam guide.