Google Cloud Generative AI Leader: sample questions with answers
20 sample questions, each with the correct answer and an explanation. They come from the question bank and follow the exam format.
Questions are written from samples prepared by specialists in the field and follow the official exam format.
1. A gen AI system recommends which suppliers to drop. Who should be accountable for the final decisions?
- A) No one in particular, because the whole process is automated
- B) Named people who review the recommendations and own the outcome
- C) The model itself, because it produced the recommendation
- D) The cloud provider, because it hosts and operates the model
Show the answer and explanation
Correct answer: B
Named people who review the recommendations own the outcome.
2. A manufacturer wants a gen AI assistant that answers maintenance questions from 20 years of manuals, but the manuals are scanned PDFs in three versions with no owner. What should come first?
- A) Build the assistant now and let the model work out which manual version is correct
- B) Assign data ownership, select the current versions and make the documents searchable before building the assistant
- C) Fine-tune a model on all three versions so it learns every answer
- D) Skip the manuals and rely on the model's general knowledge of machines
Show the answer and explanation
Correct answer: B
First assign an owner, pick the current versions and make the documents searchable.
3. A large company has dozens of teams building gen AI tools independently, with no common rules. What should leadership put in place first?
- A) A governance framework with policies, risk review, approved tools and owners
- B) A rule that internal tools can launch without any review at all
- C) A ban on all teams except IT building any gen AI tools from now on
- D) A single large model that every team must use for every use case
Show the answer and explanation
Correct answer: A
Set an AI governance framework: policies, risk review, approved tools and owners.
4. What makes a gen AI pilot most useful for deciding whether to scale?
- A) Clear success criteria, a defined user group, a baseline and a time frame
- B) As many features as possible in the first release to impress users
- C) No measurement at all, so that users feel free to experiment openly
- D) The hardest problem in the company, to prove what the technology can do
Show the answer and explanation
Correct answer: A
A useful pilot has success criteria, a user group, a baseline and a time frame.
5. A CV-screening model rates candidates from certain universities higher because past hires came mostly from them. What does this show?
- A) The model's temperature is set too low
- B) The model has a knowledge cutoff
- C) The model is hallucinating facts about the universities
- D) The model has learned historical bias from its training data
Show the answer and explanation
Correct answer: D
The model learned historical bias from its training data.
6. A team wants a model to answer questions about a 1,500-page technical manual in a single request. Which model characteristic matters most?
- A) Whether it uses supervised or unsupervised training
- B) A long context window measured in tokens
- C) A high temperature setting
- D) The number of image styles it supports
Show the answer and explanation
Correct answer: B
A long context window, measured in tokens, matters most.
7. What is a prompt in the context of a generative model?
- A) The hardware accelerators that run the model during inference
- B) The input instructions and context given to the model to guide its output
- C) The output text, image or code that the model returns to the user
- D) The large dataset used to pre-train the model before its release
Show the answer and explanation
Correct answer: B
A prompt is the input instructions and context that guide the output.
8. A regulated bank must keep all data and model weights inside its own environment and audit the model fully. Which option suits it best?
- A) An open model such as Gemma deployed in the bank's controlled environment
- B) A consumer chatbot app used through a browser
- C) Any managed API, since all APIs store data in the customer's data centre
- D) A model chosen only by benchmark ranking
Show the answer and explanation
Correct answer: A
Full control of weights and environment favours an open model such as Gemma.
9. Which set of components does Google describe as the core of an AI agent?
- A) A database for storage, a dashboard for users and a tool for reporting results
- B) A training dataset, a team of human labellers and a dedicated GPU cluster
- C) A model for reasoning, tools to act on systems, and orchestration of the steps
- D) An image generator, a digital watermark and a public web search engine
Show the answer and explanation
Correct answer: C
An agent has a reasoning model, tools and orchestration.
10. What distinguishes an AI agent from a simple chatbot that answers questions?
- A) An agent always uses a larger and more expensive model than a chatbot does
- B) An agent cannot access external data, so it relies only on its training
- C) An agent only produces images and video, while a chatbot produces text
- D) An agent can plan steps toward a goal and use tools or APIs to take actions
Show the answer and explanation
Correct answer: D
An agent plans steps and uses tools or APIs to act.
11. A retailer's existing system predicts next week's demand for each product, and a new tool writes product descriptions from a few keywords. How are the two best classified?
- A) The first is predictive machine learning; the second is generative AI
- B) Both are generative AI because both produce an output
- C) Both are traditional machine learning because both learn from data
- D) The first is rule-based AI; the second is reinforcement learning
Show the answer and explanation
Correct answer: A
Demand forecasting is predictive ML; writing new descriptions is generative AI.
12. An agent needs a standard way to connect to the company's databases and internal tools as data sources. Which open protocol is commonly used for this?
- A) Agent2Agent protocol (A2A)
- B) Model Context Protocol (MCP)
- C) Top-p (nucleus) sampling
- D) Secure AI Framework (SAIF)
Show the answer and explanation
Correct answer: B
MCP is the standard for connecting an agent to tools and data.
13. A hospital needs to extract text and fields from scanned referral forms quickly, without building a model. Which option fits best?
- A) Gemma fine-tuned on a small set of medical images
- B) A new extraction model trained from scratch on TPUs
- C) Document AI, a pre-trained document processing service
- D) Veo, used to turn the scanned forms into a video summary
Show the answer and explanation
Correct answer: C
Document AI is a ready-made document extraction service.
14. An engineering team building its own RAG pipeline needs to store and search embeddings at large scale. Which Google Cloud capabilities support this?
- A) Google Slides, which stores embeddings in presentations
- B) SynthID, which tags each embedding with a watermark
- C) Cloud Billing reports, which index usage by project
- D) Vector Search, or vector search in BigQuery and AlloyDB
Show the answer and explanation
Correct answer: D
Embeddings are stored and searched in Vector Search or BigQuery and AlloyDB.
15. A team wants the model to classify customer emails into its own five categories with a consistent format. The first attempt, with instructions only, is inconsistent. What should it try next?
- A) Removing the category names to let the model decide
- B) Raising the temperature for more varied categories
- C) Zero-shot prompting with shorter, fewer instructions
- D) Few-shot prompting: add a few labelled example emails
Show the answer and explanation
Correct answer: D
Few-shot prompting: add a few labelled examples.
16. A team has refined prompts for weeks but the model still lacks knowledge of the firm's internal procedures. What is the best next step?
- A) Raise the maximum output tokens for longer answers
- B) Keep rewriting the prompt with more detailed adjectives
- C) Ground the model in the procedure documents through RAG
- D) Switch to a larger model trained on more public data
Show the answer and explanation
Correct answer: C
Prompts add no knowledge; ground the model in the documents with RAG.
17. In which case is fine-tuning more appropriate than prompt engineering or RAG?
- A) When the team has no example data at all and needs results within the same day
- B) When the model must reliably follow a specialised style or format that prompts cannot hold
- C) When answers must reflect information that changes every hour across many systems
- D) When the only goal is to shorten answers, which prompts and settings cannot do
Show the answer and explanation
Correct answer: B
Fine-tune for a specialised style or format that prompts cannot hold.
18. How does retrieval-augmented generation (RAG) improve answers?
- A) It compresses the model so it runs faster and can read more documents in each request
- B) It retrieves relevant passages from a knowledge source at query time and adds them to the prompt
- C) It raises the model's temperature so that answers draw on a wider range of knowledge
- D) It retrains the model's weights on all new company documents every night in a batch job
Show the answer and explanation
Correct answer: B
RAG retrieves relevant passages at query time and adds them to the prompt.
19. A finance analyst asks the model a multi-step question about loan repayments and often gets wrong totals. Which prompting technique is most likely to help?
- A) Role prompting, asking the model to act as a famous economist
- B) Asking for a shorter answer with only the final total shown
- C) Chain-of-thought prompting, asking for step-by-step reasoning
- D) Zero-shot prompting with no context and no instructions
Show the answer and explanation
Correct answer: C
Chain-of-thought: ask the model to reason step by step.
20. A legal team needs consistent, factual clause extraction, while the marketing team wants varied slogan ideas. How should temperature be set?
- A) High for the legal team, low for the marketing team
- B) Low for both, because temperature only affects speed
- C) Low for the legal team, higher for the marketing team
- D) High for both, to maximise quality
Show the answer and explanation
Correct answer: C
Low temperature for consistent facts, higher for varied ideas.