Google Cloud Generative AI Leader Practice Exam

• A. It stores documents and returns exact matches to queries
• B. It is trained on one labeled dataset for a single task
• C. It follows hand-written rules to produce each output
• D. It is trained on broad data and adapted to many tasks Correct answer

Option D is correct because a foundation model is trained on broad, large-scale data and can then be adapted (prompted, tuned) to many downstream tasks. Options B and C describe single-task supervised models and rule-based systems, and option A describes a search or retrieval system, not a generative foundation model.

• A. Reinforcement learning
• B. Transfer learning
• C. Unsupervised learning Correct answer
• D. Supervised learning

Option C is correct because unsupervised learning finds structure, such as customer clusters, in data that has no predefined labels. Supervised learning (D) needs labeled examples, reinforcement learning (A) learns from rewards, and transfer learning (B) reuses a pretrained model rather than defining a grouping approach.

• A. Text summarization Correct answer
• B. Sentiment scoring
• C. Code completion
• D. Language translation

Option A is correct because reducing a long transcript to a short case note is text summarization. Sentiment scoring (B) rates tone, code completion (C) suggests code, and translation (D) converts between languages, none of which condense content.

• A. Raise the temperature so answers vary more
• B. Ground answers in the firm's case database Correct answer
• C. Switch to a model with a larger context window
• D. Fine-tune the model on recent client emails

Option B is correct because grounding ties the model's answers to a trusted source, here the firm's case database, which reduces fabricated citations (hallucinations). Raising temperature (A) increases randomness, a larger context window (C) does not supply verified facts, and fine-tuning on client emails (D) adds irrelevant data.

• A. Relational data
• B. Time-series data
• C. Unstructured data Correct answer
• D. Structured data

Option C is correct because text such as reviews, emails and chat logs has no fixed schema or columns, which defines unstructured data. Structured and relational data (D, A) fit rows and columns, and time-series data (B) is ordered numeric measurements over time.

• A. Tensor Processing Units (TPUs) Correct answer
• B. Axion Arm-based processors
• C. Titan security chips
• D. Pixel Tensor G-series processors

Option A is correct because Tensor Processing Units (TPUs) are Google-designed accelerators built for training and serving AI models in Google Cloud. Axion is an Arm-based general-purpose CPU, Titan chips are hardware security chips, and Pixel Tensor processors are for Pixel phones.

• A. Veo Correct answer
• B. Imagen
• C. Gemma
• D. Lyria

Option A is correct because Veo is Google's video generation model that creates video clips from text descriptions. Imagen generates images, Gemma is a family of open text models, and Lyria generates music.

• A. Model Garden Correct answer
• B. Customer Engagement Suite
• C. Gemini Enterprise
• D. Gemini Code Assist

Option A is correct because Model Garden is the Gemini Enterprise Agent Platform catalog where teams discover, test and deploy Google, open and third-party models. The other options are an agent suite, an enterprise assistant and a coding assistant, not a model catalog.

• A. Gemini for Google Workspace Correct answer
• B. Customer Engagement Suite
• C. Model Garden
• D. Gemini Code Assist

Option A is correct because Gemini for Google Workspace brings generative AI into Gmail, Docs and other Workspace apps for drafting and summarizing. Customer Engagement Suite targets customer service, Model Garden is a model catalog, and Gemini Code Assist helps developers write code.

• A. Gemini Enterprise Correct answer
• B. Google AI Studio
• C. Gemini for Google Workspace
• D. Gemini Code Assist

Option A is correct because Gemini Enterprise provides one assistant that searches across Workspace, Microsoft 365 and other business apps and lets employees build agents without code. Google AI Studio is a developer prototyping tool, Gemini for Workspace is limited to Workspace apps, and Gemini Code Assist is for coding.

• A. Training data bias Correct answer
• B. Context window overflow
• C. Output token limit
• D. Knowledge cutoff

Option A is correct because training data bias arises when the data a model learned from over-represents some regions or groups, so its output is less accurate for the under-represented ones. A context window overflow, an output token limit and a knowledge cutoff concern input size, response length and recency, not uneven coverage.

• A. Shortening the system instructions
• B. Raising the model's temperature setting
• C. Switching to a larger context window
• D. Human review of high-stakes outputs Correct answer

Option D is correct because a person checking high-stakes output before release is the control that catches fluent but wrong answers (hallucinations). Shortening instructions, raising temperature and a larger context window do not verify accuracy, and a higher temperature makes errors more likely.

• A. Fine-tune the model on general medical textbooks
• B. Add a longer expert role description to the prompt
• C. Answer only from retrieved formulary passages Correct answer
• D. Switch to a model with a larger context window

Option C is correct because grounding the answer in retrieved formulary passages (retrieval-augmented generation) restricts the model to approved source text, which reduces invented doses. Fine-tuning on general textbooks, a longer role description or a larger context window do not tie answers to the formulary.

• A. Extend the context window
• B. Increase the maximum output tokens
• C. Lower the temperature value Correct answer
• D. Add a longer role description

Option C is correct because a lower temperature makes token selection less random, so repeated runs give more consistent wording. The context window size, the maximum output tokens and a longer role description do not control randomness.

• A. High inference latency
• B. Overreliance on outputs Correct answer
• C. Insufficient context length
• D. Stale training knowledge

Option B is correct because overreliance is the risk that people accept model output without checking it. Latency, insufficient context length and stale training knowledge are different limitations that do not describe staff skipping verification.

• A. Consolidate all company data into one repository
• B. Define the business problem and target outcomes Correct answer
• C. Select the largest foundation model on offer
• D. Give every employee access to a chat assistant

Option B is correct because a gen AI initiative should start from a clearly defined business problem and target outcomes, which then drive the choice of model, vendor and data. Options A, C and D are premature or unfocused steps: consolidating all data, picking the largest model, or giving everyone a chatbot do not tie the effort to a measurable goal.

• A. The model ranked highly on a public benchmark
• B. Time to publish listings fell by half Correct answer
• C. Weekly prompt volume doubled over the pilot
• D. The tool ran on the newest Gemini version

Option B is correct because executives judge value by business outcomes, and halving the time to publish listings is a direct, measurable efficiency gain. Benchmark rank, prompt volume and the model version used are technical or activity measures that do not show business impact.

• A. Move the workload onto dedicated TPUs
• B. Switch to a larger foundation model
• C. Appoint an executive sponsor to align teams Correct answer
• D. Add more examples to the prompt template

Option C is correct because the stall comes from organizational misalignment, and an executive sponsor can align priorities, budget and ownership across departments. Dedicated TPUs, a larger model or more prompt examples are technical changes that do not resolve disagreement between teams.

• A. Switch the tool to a newer Gemini model first
• B. Plan integration, training and support for scale Correct answer
• C. Fine-tune a custom model on all call transcripts
• D. Start a second use case in another department

Option B is correct because scaling from 30 to 3,000 users requires planning for integration with existing systems, training and ongoing support. Switching models, fine-tuning on all transcripts or starting another use case do not prepare the proven pilot for wide rollout.

• A. Move the assistant to a larger Gemini model
• B. Limit access to a smaller group of testers
• C. Involve staff early and train them to use it Correct answer
• D. Publish usage dashboards for team leads

Option C is correct because fear of replacement is a change-management issue, best addressed by involving employees early and training them to work with the assistant. A larger model, restricting access or publishing usage dashboards do not address the concern and may increase distrust.