Overview
Fundamentals of gen AI is the conceptual base of the Generative AI Leader exam. It has four official objectives. The first covers core terms such as AI, machine learning, large language models, foundation models and prompting, along with supervised, unsupervised and reinforcement learning, the ML lifecycle, and how to choose a foundation model. The second looks at data: quality, accessibility, structured versus unstructured, and labeled versus unlabeled. The third describes the gen AI landscape as five layers: infrastructure, models, platforms, agents and applications. The fourth asks you to match Google's foundation models (Gemini, Gemma, Imagen and Veo) to use cases.
At about 30%, this is the second-largest section. The exam is written for business and technology leaders, not engineers, so questions test whether you can explain a concept and apply it to a business decision rather than build or train a model. As of October 2026 the exam is current, with a guide refreshed for Google's new product names.
How it's tested
Section 1 questions are short business scenarios written in plain language. A team has a goal, some data and a constraint, and you choose the concept, learning approach, landscape layer or Google model that matches. Expect definitions to be tested through application rather than recall, so read for the clue words.
-
Name the concept
A description of a capability or process is given and you identify the term, such as foundation model, LLM, fine-tuning or reinforcement learning.
-
Judge the data
A data set is described and you decide whether it is structured or unstructured, labeled or unlabeled, and whether its quality supports the use case.
-
Pick the model or layer
A content need such as text, image, video or on-device use is stated, and you select Gemini, Gemma, Imagen or Veo, or place a product in its layer.
A marketing team wants short product videos generated from text descriptions. Gemini is multimodal and can understand video, which makes it tempting. Imagen generates still images, so it does not produce clips. Veo is Google's video generation model, built to create video from text or image prompts. The output modality is the deciding clue, so Veo is the expected answer.
Objectives 4
Official objectives from the Google Cloud outline, in outline order.
1.1 Describe core generative AI (gen AI) concepts and use casesAI, ML, LLM and foundation model terms, learning approaches, the ML lifecycle and model choice.
- What it means
- This objective covers the basic vocabulary: artificial intelligence, machine learning, generative AI, large language models, foundation models, multimodal models, prompts and tokens. It adds the three ML approaches (supervised, unsupervised and reinforcement learning), the stages of the ML lifecycle with the Google Cloud tools for each, choosing a foundation model by modality, cost, latency and customization needs, and business use cases built around generating, summarizing, discovering and automating.
- What to understand
- Generative AI is a subset of ML that creates new content, while predictive ML classifies or forecasts. A foundation model is pre-trained on broad data and adapted to many tasks; an LLM is a foundation model focused on language. Supervised learning uses labeled examples, unsupervised finds patterns in unlabeled data, and reinforcement learning improves through rewards. The lifecycle runs from data ingestion and preparation through training, deployment and monitoring, with Agent Platform (formerly Vertex AI) covering most stages.
- How it shows up
- A goal or process is described and you name the concept, learning approach, lifecycle stage or model selection factor that fits.
- Common confusion
- Treating every AI system as generative; a fraud classifier is predictive ML, not gen AI.
1.2 Describe how various data types are used in gen AI and the business implicationsData quality and accessibility, structured versus unstructured, labeled versus unlabeled.
- What it means
- This objective covers why data quality and accessibility decide whether a gen AI project succeeds, how structured data such as tables and transaction records differs from unstructured data such as emails, PDFs, images and audio, and how labeled data differs from unlabeled data.
- What to understand
- Quality covers accuracy, completeness, consistency, relevance and timeliness; poor data leads to poor output no matter how capable the model is. Accessibility means the data can be found, permissioned and connected, which is often the practical blocker. Most enterprise knowledge is unstructured, which is why gen AI is valuable for search and summarization. Labels cost time and money to create, so a use case that needs labeled data has a higher preparation cost than one that can use unlabeled text.
- How it shows up
- A data set is described and you classify it or judge whether its quality and accessibility support the stated business goal.
- Common confusion
- Assuming a spreadsheet of free-text comments is structured because it sits in rows; the content itself is unstructured.
1.3 Identify the core layers of the gen AI landscape and the business implicationsInfrastructure, models, platforms, agents and applications.
- What it means
- This objective asks you to place gen AI components into five layers: infrastructure (compute such as TPUs and GPUs, storage and networking), models (foundation models like Gemini), platforms (tools to access, customize and deploy models, such as Agent Platform), agents (systems that reason and use tools to complete tasks), and applications (the products end users touch, such as the Gemini app or Gemini for Google Workspace).
- What to understand
- Each layer builds on the one below it, and an organization can enter at different layers depending on skills and control needs. Buying an application gives fast value with little control; building on the platform layer gives more customization but needs more expertise; owning infrastructure decisions is rarely needed for most businesses. Agents sit between models and applications because they combine a model with tools, memory and orchestration.
- How it shows up
- A product or decision is described and you identify its layer, or choose which layer a business should adopt given its skills and need for control.
- Common confusion
- Confusing the model layer with the platform layer; Gemini is a model, Agent Platform is the platform used to work with it.
1.4 Identify the use cases and strengths of Google's foundation modelsGemini, Gemma, Imagen and Veo.
- What it means
- This objective covers matching Google's model families to tasks. Gemini is the natively multimodal flagship family for reasoning over text, code, images, audio and video. Gemma is a family of lightweight open models that organizations can download, run and customize on their own infrastructure. Imagen generates and edits images from text. Veo generates video from text or image prompts.
- What to understand
- Choose by output modality and deployment need. Multimodal understanding, long context and complex reasoning point to Gemini. A need for open weights, on-premises or edge deployment, or full control over tuning points to Gemma. Marketing visuals and product images point to Imagen, and short video clips point to Veo. Many of these models are also reachable through Model Garden in Agent Platform.
- How it shows up
- A content or deployment requirement is given, and you pick the Google model family that meets it.
- Common confusion
- Choosing Gemini when the clue is open weights or self-hosting, which points to Gemma.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Foundation model
A large model pre-trained on broad data and adaptable to many tasks.
- Remember
- One model, many downstream uses.
- Tested as
- Concept questions.
- Trap
- Equating it only with chatbots.
-
Large language model (LLM)
A foundation model specialized in understanding and generating language.
- Remember
- LLMs predict the next token.
- Tested as
- Concept questions.
- Trap
- Assuming LLMs look facts up.
-
Multimodal model
A model that accepts or produces several data types.
- Remember
- Gemini is natively multimodal.
- Tested as
- Model choice questions.
- Trap
- Picking a text-only option for image input.
-
Supervised learning
Training on labeled input-output examples.
- Remember
- Needs labels.
- Tested as
- Learning-approach questions.
- Trap
- Using it when no labels exist.
-
Unsupervised learning
Finding structure in unlabeled data, such as clusters.
- Remember
- No labels required.
- Tested as
- Learning-approach questions.
- Trap
- Confusing clustering with classification.
-
Reinforcement learning
Learning by trial and reward signals.
- Remember
- Optimizes actions toward a goal.
- Tested as
- Learning-approach questions.
- Trap
- Mixing it up with supervised learning.
-
ML lifecycle
Data ingestion, preparation, training, deployment and monitoring.
- Remember
- Monitoring continues after launch.
- Tested as
- Process questions.
- Trap
- Treating deployment as the final step.
-
Data quality
Accuracy, completeness, consistency, relevance and timeliness of data.
- Remember
- Bad data gives bad output.
- Tested as
- Readiness questions.
- Trap
- Blaming the model for a data problem.
-
Structured vs. unstructured data
Tables and fields versus documents, images and audio.
- Remember
- Most enterprise knowledge is unstructured.
- Tested as
- Data questions.
- Trap
- Judging by container rather than content.
-
Gen AI landscape layers
Infrastructure, models, platforms, agents, applications.
- Remember
- Higher layers build on lower ones.
- Tested as
- Landscape questions.
- Trap
- Calling a model a platform.
-
Gemini
Google's multimodal flagship model family.
- Remember
- Reasoning across text, code, image, audio and video.
- Tested as
- Model questions.
- Trap
- Using it when open weights are required.
-
Gemma
Lightweight open models from Google.
- Remember
- Self-host and customize freely.
- Tested as
- Model questions.
- Trap
- Thinking it is a closed API model.
-
Imagen and Veo
Imagen generates images; Veo generates video.
- Remember
- Match the output modality.
- Tested as
- Model questions.
- Trap
- Swapping the two.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Supervised vs. unsupervised vs. reinforcement learning
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Supervised | Learns from labeled examples | Requires labels | Choosing it with unlabeled data |
| Unsupervised | Finds patterns in unlabeled data | No target label | Expecting named categories |
| Reinforcement | Learns from rewards over actions | Feedback loop, not labels | Confusing it with supervised |
Gemini vs. Gemma vs. Imagen vs. Veo
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Gemini | Multimodal flagship models | Broad reasoning, managed access | Picked when self-hosting is needed |
| Gemma | Lightweight open models | Open weights, run anywhere | Assumed to be closed |
| Imagen | Image generation and editing | Produces still images | Chosen for video |
| Veo | Video generation | Produces video clips | Chosen for images |
Structured vs. unstructured data
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Structured | Fixed fields such as tables and records | Easy to query | Assuming gen AI only uses it |
| Unstructured | Documents, emails, images, audio | Needs interpretation | Labeling it structured because it is stored in rows |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Wrong output modality
- What candidates pick
- Imagen for a request to generate short video clips.
- Why it looks right
- Imagen is a well-known generation model.
- Why it's wrong
- Imagen creates still images; Veo is the Google model built to generate video.
Clue to watch for The deliverable is video.
-
Closed model for self-hosting
- What candidates pick
- Gemini for a team that must run the model on its own servers.
- Why it looks right
- Gemini is the flagship.
- Why it's wrong
- Gemma provides open weights that can be downloaded and run on the organization's own infrastructure.
Clue to watch for On-premises, edge or open weights.
-
Labels that do not exist
- What candidates pick
- Supervised learning for grouping customers when no categories have been defined.
- Why it looks right
- Supervised learning is the most familiar approach.
- Why it's wrong
- Without labels, unsupervised learning such as clustering is the fit for discovering groups.
Clue to watch for No predefined categories or labels.
-
Model fixes bad data
- What candidates pick
- Switch to a larger model to fix inaccurate answers from outdated records.
- Why it looks right
- Bigger models seem smarter.
- Why it's wrong
- Output quality is bounded by data quality; stale or incomplete records must be fixed first.
Clue to watch for The data is described as old, incomplete or inconsistent.
-
Layer confusion
- What candidates pick
- Call Agent Platform a foundation model.
- Why it looks right
- It is closely tied to Gemini.
- Why it's wrong
- Agent Platform sits in the platform layer, giving access to and tooling around models such as Gemini.
Clue to watch for The question asks which layer a product belongs to.
-
Every AI is gen AI
- What candidates pick
- Label a demand forecasting model as generative AI.
- Why it looks right
- It uses machine learning.
- Why it's wrong
- Forecasting predicts a value from data; generative AI creates new content such as text, images or code.
Clue to watch for The output is a prediction, score or class.
Answer framework
For any Section 1 scenario, work through these steps to reach the expected answer.
- Decide whether the task creates new content or predicts, classifies or groups existing data, which separates gen AI from other ML.
- If it is about learning approach, check for labels and feedback: labeled examples mean supervised, no labels mean unsupervised, rewards mean reinforcement.
- Assess the data: is it structured or unstructured, labeled or unlabeled, and is it accurate, current, complete and accessible enough for the goal?
- Place the product or decision in the right landscape layer, from infrastructure up through models, platforms and agents to applications.
- When a model is needed, match the output modality and deployment need: Gemini for multimodal reasoning, Gemma for open self-hosted use, Imagen for images, Veo for video.
Practice question
Want more like this? The GenAI Leader practice test has questions for every domain.
Checklist
0 of 11
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Start with vocabulary, because this section is about precise meaning. Write one-line definitions for AI, ML, gen AI, LLM, foundation model, multimodal, prompt and token, then test yourself by explaining each to a non-technical colleague. Next, learn the three ML approaches through examples: labeled churn data, unlabeled customer segments and a reward-driven recommendation loop.
Then study data. Collect sample data from your own work and sort it into structured or unstructured and labeled or unlabeled, and list what would make it unfit for a gen AI project. Draw the five-layer landscape and place Google products on it, from TPUs up to the Gemini app.
Finish with the model families. Make a four-row table for Gemini, Gemma, Imagen and Veo with output type, deployment style and a typical business use. Google's free Generative AI Leader learning path on Google Skills covers this material at the right depth. Once the vocabulary is solid, move to Section 2, which applies these ideas to Google Cloud's offerings, and use practice questions to confirm you read scenarios for the modality and data clues.
FAQ
How much of the GCP Gen AI Leader exam is Fundamentals of gen AI?
Fundamentals of gen AI is about 30% of the exam, the second-largest section after Google Cloud's gen AI offerings.
Do I need to know how to code or train models for this section?
No. The exam targets business and technology leaders. You need to explain concepts and apply them to decisions, not write code or train models.
What is the difference between a foundation model and an LLM?
A foundation model is any large pre-trained model adaptable to many tasks, across modalities. An LLM is a foundation model focused on language.
What are the five layers of the gen AI landscape?
Infrastructure, models, platforms, agents and applications. Each layer builds on the one beneath it, and organizations can adopt at whichever layer matches their skills and need for control.
When should I choose Gemma over Gemini?
Choose Gemma when the scenario needs open weights, self-hosting, edge or on-premises deployment, or full control over tuning. Choose Gemini for managed, multimodal reasoning at scale.
Why does data quality show up in a gen AI exam?
Model output depends on the data used to train, tune or ground it. Leaders must judge whether data is accurate, current, complete and accessible before committing to a use case.
Is the Generative AI Leader exam still current?
Yes. As of October 2026 it is active, and Google updated the exam guide to reflect new product names such as Agent Platform and Gemini Enterprise.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 1 knowledge
You have covered core gen AI terms, data types, the landscape layers and Google's model families. Check whether you can apply them to business scenarios under exam conditions.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the GenAI Leader AI tutor Get any concept explained your way