Free Google Generative AI Leader Sample Questions

Correct answer: A — Veo, because it generates video sequences directly from descriptive text, useful for rough scene previsualization

• A. This model's core purpose is generating video sequences from text prompts, which is exactly what turning a script excerpt into rough scene footage requires. That makes it the model built for text-to-video previsualization work.
• B. This model generates still images rather than video, and it does not offer a mode that stitches stills into motion footage. Previsualization needs actual generated video, which is outside this model's core capability.
• C. This general reasoning and text model is not built to render animated video output from a script; its strength is multimodal understanding and conversation, not video generation. The studio needs a dedicated video-generation model for this task.
• D. This open model is meant for self-hosted fine-tuning of general or text models, not for training a scene-generation capability overnight, and it is not the model built for producing video. It would not deliver rough scene footage from a script.

Correct answer: A — Imagen, because it generates multiple still-image variations from text prompts, suited for rapid concept review

• A. This model's core strength is producing still-image variations directly from text prompts, which is exactly what a fast, image-only concept review needs before any video work starts. It is the model purpose-built for this kind of rapid still-image generation.
• B. This model's purpose is generating new video, and extracting still frames from a video pipeline is not the same as a model purpose-built for direct text-to-still-image generation. Using it here adds an unnecessary video-generation step for a task that only needs stills.
• C. This general reasoning model is not described as producing still images as an output of its multimodal reasoning; its strength is text and cross-format understanding, not image rendering. The agency needs a dedicated image-generation model, not a reasoning model repurposed for images.
• D. This open model is meant for self-hosted customization of general or text models, and it is not documented as including a built-in image-rendering capability. It does not fit a task that specifically requires generating varied still images from text.

Correct answer: A — Imagen fits, because the brief calls only for static images and Imagen specializes in generating still images

• A. The brief explicitly calls only for static images, and this model's specialty is generating still images directly from text, which is exactly what static banner ads require. There is no need to introduce a video-generation step at all.
• B. Generating a video first just to extract a still frame adds an unnecessary detour when the brief calls only for static images, and this model is not documented as producing sharper stills than a dedicated image-generation model. The simpler, direct path is a model built for still images.
• C. This model generates still images, not looping video clips, so this option mischaracterizes its output; regardless, the brief wants static banners, not looping animation. The description here does not match how either model actually works.
• D. Static banner ads are ordinary still images, not single-frame video files, so framing them as a video format misrepresents the deliverable. This model's core purpose is generating motion video, not the still images the brief requires.

Correct answer: A — Monitoring the deployed model's performance over time, detecting drift, and triggering retraining or updates as needed.

• A. Model management is the ongoing stage after deployment that tracks how a live model is performing, watches for drift in accuracy, and triggers retraining or an update when performance degrades.
• B. Loading raw data into central storage describes data ingestion, which happens near the beginning of the lifecycle, well before a model is trained, deployed, or managed.
• C. Cleaning and transforming raw records into a structured format describes data preparation, an earlier lifecycle stage that happens before training, not the post-deployment management stage.
• D. Selecting an algorithm and fitting its parameters describes the model training stage, which occurs before a model is ever deployed and therefore before management begins.

Correct answers: A, B, C — Generating a first draft of marketing copy from a short creative brief to speed up the content team's workflow.; Summarizing a long customer feedback dataset into a short list of the most common themes for a product team.; Analyzing structured sales data to surface trends and anomalies that inform a quarterly business review.

• A. Generating a first-draft of marketing copy from a brief is a genuine text generation use case, helping the content team start from something rather than a blank page.
• B. Condensing a large volume of customer feedback into the most common themes is a genuine summarization use case that gives a product team a quick, digestible overview.
• C. Surfacing trends and anomalies in structured sales data to inform a business review is a genuine data analysis use case that gen AI can support.
• D. Physically manufacturing hardware on a factory floor is a mechanical, physical-world process, not something a gen AI model performs; gen AI works with data and content, not physical assembly.
• E. No gen AI model can guarantee legal validity on its own or replace human legal review; contracts still require qualified human oversight, making this an unrealistic use case.
• F. Deleting a customer database is a destructive data-management action, not a gen AI capability, and it would not improve data analysis; it would remove the data needed for it.

Correct answer: A — The platforms layer, using an environment such as Vertex AI Studio to test and evaluate prompt variations before deployment.

• A. The platforms layer is correct because an environment such as Vertex AI Studio is built specifically for designing, testing, and evaluating prompt variations before anything reaches customers.
• B. Adding TPU capacity at the infrastructure layer increases raw compute throughput, but it does not give the team any way to compare prompt phrasings or measure output quality.
• C. Function calling at the agents layer lets the assistant retrieve external data at runtime, but it does not provide the prompt comparison and evaluation the team is asking for.
• D. Gemini for Workspace is an application-layer product for end users, not a tool for comparing and evaluating candidate prompts before deployment.

Correct answer: A — The engineer's tool represents the platforms layer, and the employee's tool represents the applications layer.

• A. This is correct because a prompt-testing tool for developers, such as Vertex AI Studio, sits at the platforms layer, while a daily writing tool embedded in an employee's document editor, such as Gemini for Workspace, sits at the applications layer.
• B. A prompt-testing tool used by developers is a platforms-layer tool, not a finished application, and an everyday document editor is a user-facing application, not an infrastructure resource.
• C. Both tools do rely on an underlying foundation model, but the layer they represent depends on their purpose, prompt testing versus daily end-user use, not merely on whether a model is involved.
• D. A prompt-testing environment is a development tool at the platforms layer rather than an autonomous tool-using agent, and a daily writing tool is a finished application, not a development platform.

Correct answer: A — Data that does not follow a predefined schema, such as free-form text documents, images, audio, and video files.

• A. Unstructured data lacks a predefined schema, so formats like free-form text, images, audio, and video fall into this category. Because it has no fixed fields, it typically needs preprocessing before a model can use it effectively.
• B. This describes structured data, which is organized into a consistent schema of rows and columns that database tools can query directly. That fixed organization is the opposite of the free-form nature of unstructured data.
• C. This describes labeled data, a concept about annotation rather than schema. Data of any structure can be labeled or unlabeled independently of whether it is structured or unstructured.
• D. This describes a model working with unlabeled data during unsupervised learning, which is a training approach rather than a definition of the data's structure. Unstructured data can be labeled or unlabeled just like structured data.

Correct answer: A — Standardize the file encoding and format across all exports so the tool can parse the combined shipment records correctly.

• A. Different character encodings across regional exports are a format inconsistency, and aligning them to a single standard encoding is what lets the tool parse the combined records without garbled text. This kind of format cleanup is a common prerequisite before combining multi-source data.
• B. Generative AI reporting tools do not require video input, and converting structured shipment records into video would not address an encoding mismatch at all. This suggestion misunderstands both the problem and the tool's actual input needs.
• C. Character encoding issues are unrelated to whether personal data such as customer names is present; the garbling comes from how text bytes are interpreted, not from the presence of specific fields. Removing names would not fix the encoding mismatch described.
• D. Generative AI tools do not automatically detect and correct encoding mismatches on their own, so merging the files unchanged would likely preserve or worsen the garbled text. The team needs to actively standardize the format before combining the exports.

Correct answer: A — Organizations gain more flexibility and reduced lock-in because they can choose among open models, open frameworks, and multiple deployment options.

• A. Correct. An open approach, spanning open models, open-source frameworks, and flexible deployment choices, reduces vendor lock-in and gives organizations more control over how they build and run generative AI.
• B. Incorrect. Openness increases tool choice rather than mandating a single framework, which contradicts the described benefit of flexibility.
• C. Incorrect. Choosing open models does not remove the need to evaluate security and compliance; those remain separate considerations regardless of licensing model.
• D. Incorrect. Google continues to offer support and enterprise tooling around open models like Gemma; adopting an open license does not eliminate vendor support.

Correct answer: A — Google Cloud maintains security certifications, compliance attestations, and data governance controls that support regulated, enterprise-scale AI deployments.

• A. Correct. Google Cloud backs its enterprise-readiness claims with security certifications, compliance attestations, and governance controls, which is the kind of formal assurance a government agency evaluating sensitive workloads would look for.
• B. Incorrect. Google Cloud publishes documented certifications and compliance materials rather than relying only on informal verbal assurances from sales staff.
• C. Incorrect. Google Cloud's enterprise-readiness framing, including security and compliance, is explicitly extended to its AI platform, not limited to non-AI services.
• D. Incorrect. Enterprise-readiness means meeting standard security requirements for AI workloads, not asking regulated customers to waive them.

Correct answers: A, B, C — Low-code and no-code tools that let non-specialists design and test prompts through a visual interface.; Pre-trained models that can be used directly for common tasks without requiring custom model training.; Ready-to-use APIs that add capabilities like translation or image analysis with a simple integration.

• A. Correct. Low-code and no-code tools let staff without deep technical backgrounds design and test prompts visually, which directly fits a small nonprofit team's situation.
• B. Correct. Pre-trained models can be used directly for common tasks, so the nonprofit does not need to invest time and expertise in training a model from scratch.
• C. Correct. Ready-to-use APIs let the organization add capabilities such as translation or image analysis through straightforward integration rather than building the capability in-house.
• D. Incorrect. No formal degree requirement exists to access Google Cloud's generative AI tools; the low-code and API-based offerings are designed to lower that exact barrier.
• E. Incorrect. Democratizing AI development means these tools are intended for organizations without a dedicated data science team, not restricted to those that already have one.
• F. Incorrect. Pre-trained models are meant to be used as-is without first building an equivalent custom model; requiring that would defeat their purpose.

Correct answer: A — Ground the agent's responses in the company's live tracking data through a connected data store, so answers are based on real records instead of guesses.

• A. Correct. Fabricated, fluent-sounding answers are a hallucination problem, and grounding the agent in the company's actual tracking data through a connected data store anchors its responses to real records instead of invented ones.
• B. Incorrect. The number of supported languages has no bearing on whether the underlying model fabricates facts; hallucination is not a language-count problem.
• C. Incorrect. Concurrency capacity affects how many conversations run in parallel, not the factual accuracy of any individual response.
• D. Incorrect. The communication channel, text or voice, does not change whether the underlying model is grounded in real data; switching channels alone would not fix fabricated answers.

Correct answer: A — Conversational Agents for automated self-service chat, combined with Agent Assist for real-time support on escalated human-handled cases.

• A. Correct. Conversational Agents is built to run automated conversational self-service, and Agent Assist is built to support a human agent in real time once a case escalates, together covering both stated needs.
• B. Incorrect. Vertex AI Search returns ranked results rather than holding an automated dialog, and Cloud Monitoring tracks infrastructure metrics rather than assisting a live agent.
• C. Incorrect. Conversational Insights analyzes past interactions rather than running a live self-service chat, and BigQuery is a data warehouse, not a real-time agent-assistance tool.
• D. Incorrect. Contact Center as a Service provides telephony infrastructure rather than an automated chat experience, and Cloud Storage holds files rather than assisting agents in real time.

Correct answer: A — Yes, because Gems are available on every Gemini plan, so the nonprofit can bundle its FAQ as knowledge into a free Gem.

• A. Correct. Gems are available on every Gemini plan, including the free tier, so the nonprofit can create a free Gem that uses its FAQ document as knowledge without paying for an upgrade.
• B. Incorrect. A paid subscription is not required to build a Gem; Gems are part of the free Gemini app experience, not an Advanced-only feature.
• C. Incorrect. Gem knowledge files are uploaded directly by the user creating the Gem and do not require an enterprise permissions-aware data store.
• D. Incorrect. Deploying a paid Workspace edition to every staff member is unnecessary; a Gem can be created directly in the free Gemini app.

Correct answer: A — Use Gemini's drafting assistance inside Google Docs to turn rough notes into a structured first draft for further editing.

• A. Correct. Gemini's drafting assistance inside Google Docs is built to expand rough notes into a structured first draft that the user can then continue editing.
• B. Incorrect. Configuring a multi-step approval workflow addresses process automation, not the manager's actual need to turn notes into a written charter draft.
• C. Incorrect. NotebookLM is oriented toward summarizing and answering questions about sources, and an audio summary is not the editable document draft the manager needs.
• D. Incorrect. Google Docs does have drafting assistance built in through Gemini, so this option incorrectly claims the capability does not exist there.

Correct answer: A — Gemini for Google Workspace, since it is embedded directly in Gmail and can generate responses using the company's own knowledge base.

• A. Correct. Gemini for Google Workspace is embedded directly in Gmail and can draft responses informed by the company's own connected knowledge and data, fitting the ticketing workflow described.
• B. Incorrect. The free consumer Gemini app has no built-in access to any company's private internal knowledge base; that kind of connection requires an enterprise or Workspace integration.
• C. Incorrect. NotebookLM organizes and answers questions from uploaded sources but does not send emails or integrate into a Gmail-based ticketing workflow on its own.
• D. Incorrect. Using the Agent Designer in isolation, without any Gmail or Workspace integration, would not deliver replies through the team's existing email-based workflow.

Correct answer: A — Prebuilt RAG grounding with Vertex AI Search, which retrieves and cites internal articles automatically.

• A. Prebuilt RAG grounding with Vertex AI Search retrieves relevant internal articles and lets the chatbot cite them automatically, without the team assembling their own retrieval pipeline.
• B. AutoML trains a new predictive model from labeled data, it does not retrieve or ground responses in existing knowledge base articles.
• C. Model Garden catalogs foundation models for discovery and deployment, it does not provide a ready-made retrieval pipeline over internal articles.
• D. The Agent Development Kit defines an agent's logic and behavior, but on its own it does not provide the managed retrieval service the team needs.

Correct answer: A — Grounding model responses in retrieved, relevant documents to reduce hallucination and improve accuracy.

• A. RAG retrieves relevant documents and provides them as context before generation, grounding the response and reducing the chance of confidently stated but inaccurate content.
• B. RAG retrieves existing content at query time rather than training a brand-new foundation model on an organization's proprietary documents.
• C. A traditional e-commerce search bar is a product catalog browsing feature, it is not what RAG is designed to solve for generative AI applications.
• D. RAG still relies on a foundation model to generate the final response, it augments that model with retrieved context rather than replacing it with static templates.

Correct answer: A — The Vertex AI RAG APIs, which expose modular retrieval, ranking, and generation steps for custom pipelines.

• A. The Vertex AI RAG APIs expose modular components for retrieval, ranking, and generation, giving the compliance team visibility and control over each step of the pipeline.
• B. Vertex AI Search grounding bundles retrieval into one managed step, which is convenient but offers less step-by-step configurability than the compliance team is asking for.
• C. AutoML is designed to train predictive models such as classifiers, it does not expose configurable retrieval and ranking steps for a RAG pipeline.
• D. Model Garden is a catalog for discovering and deploying foundation models, it does not provide configurable retrieval pipeline steps for RAG.

Correct answer: A — Document AI API to extract the fields, then a function that calls the payments API with the extracted values.

• A. Correct. Document AI API is designed to extract structured fields like vendor name and amount from a scanned invoice, and a function tool can then call the payments API with those extracted values to complete the task.
• B. Incorrect. Cloud Vision API performs general image analysis and cannot post data to an external payments API on its own, and skipping extraction leaves no structured data to post.
• C. Incorrect. Converting the invoice to speech and back to text adds an unnecessary audio round-trip and does not extract the vendor name or amount as structured fields.
• D. Incorrect. Cloud Video Intelligence API analyzes video footage over time; a scanned invoice is a static document, not a video, so this tool does not fit the extraction step.

Correct answer: A — The foundation model's knowledge is fixed at training time, so it needs a tool calling the weather API for today's live forecast.

• A. Correct. A foundation model's knowledge is fixed as of its training data, so it cannot know today's actual forecast; a tool that calls a live weather API is required to bring in current, real-time information.
• B. Incorrect. The model can generate general text about weather topics from its training; the real limitation is the lack of current, real-time forecast data, not an inability to write about weather at all.
• C. Incorrect. Translating the word weather is unrelated to the actual problem, which is the model's lack of access to live forecast data.
• D. Incorrect. Text formatting such as bolding is a presentation detail and is not the reason a live data tool is needed here.

Correct answer: A — Grounding connects a model's responses to verifiable external information sources so answers are anchored in real data rather than the model's memory alone.

• A. This is correct because grounding specifically means tying the model's output to external, verifiable sources of information rather than relying purely on patterns memorized during training.
• B. This is incorrect because permanently retraining the model's weights describes fine-tuning, a separate technique that changes model behavior rather than connecting it to external sources at answer time.
• C. This is incorrect because temperature is a generation setting that controls output randomness, and lowering it has nothing to do with connecting the model to external data sources.
• D. This is incorrect because writing longer prompts describes prompt engineering, which shapes instructions to the model but does not by itself connect it to any external, verifiable source.

Correct answers: A, B, C — Hallucination; Knowledge cutoff; Bias

• A. Hallucination is correct because it is a well-known foundation model limitation in which the model generates fluent but fabricated or inaccurate content.
• B. Knowledge cutoff is correct because it is a limitation caused by training data having a fixed end date, leaving the model unaware of anything after that point.
• C. Bias is correct because it is a limitation in which patterns present in the training data can cause the model to produce skewed or unbalanced outputs.
• D. Grounding is incorrect as a limitation because it is a Google Cloud-recommended technique used to address limitations, not a limitation itself.
• E. Fine-tuning is incorrect as a limitation because it is a technique used to adapt model behavior, not a shortcoming inherent to foundation models.
• F. Human in the loop is incorrect as a limitation because it is a recommended review practice used to catch problems, not a limitation of the model itself.

Correct answer: A — Grounding with Google Search, since the official library documentation is publicly hosted and indexed rather than stored in a private company data store.

• A. Since the documentation is public and already indexed on the open web, Grounding with Google Search lets the assistant reference it directly without the team building any data store.
• B. Pre-built RAG with Vertex AI Search requires ingesting content into a managed data store first, which adds setup work the team is explicitly trying to avoid by having no repository yet.
• C. Building a custom retrieval pipeline with the RAG APIs is more setup effort than needed when the target content is already public and searchable on the open web.
• D. First-party data must be content the team itself owns and authors; documentation published by an unrelated open-source project does not become the team's own enterprise data.

Correct answer: A — Pre-built RAG with Vertex AI Search for the private onboarding documents, combined with Grounding with Google Search for current public news.

• A. Pre-built RAG with Vertex AI Search handles the private onboarding documents with minimal setup, while Grounding with Google Search covers current public news, matching both needs without a custom pipeline.
• B. The RAG APIs demand building a custom retrieval pipeline, which the startup explicitly lacks resources for, and nightly fine-tuning cannot keep pace with breaking news.
• C. Grounding with Google Search only reaches public web content, so it cannot ground answers on the startup's own private onboarding documents, which are not publicly indexed.
• D. Fine-tuning on both data sets is resource-intensive, quickly goes stale for changing news, and does not solve the freshness or maintenance problems retrieval-based grounding avoids.

Correct answer: A — Switch to few-shot prompting by adding several correctly labeled document examples so the model can learn the company's specific category naming pattern.

• A. Adding several correctly labeled examples is few-shot prompting, and it directly teaches the model the company's specific category naming pattern that zero-shot instructions alone failed to convey.
• B. Splitting the task into an unrelated summary prompt and a separate category prompt does not teach the model the company's specific naming pattern the way labeled examples would.
• C. Adopting a persona like senior paralegal changes tone and framing but does not show the model the specific category names it needs to learn, so it would not fix a naming mismatch.
• D. Asking only for reasoning without ever naming categories would not resolve a category-naming problem, since the model still needs to learn the correct labels to apply.

Correct answer: C — One-shot prompting, which provides a single worked example so the model can replicate that example's style and structure across new entries.

• A. Zero-shot prompting would skip the worked example entirely and rely on instructions alone, which is not what happens when a single FAQ pair is deliberately provided.
• B. Defining a persona like product support writer describes role prompting, which is a separate technique from supplying one worked example to imitate.
• C. Supplying exactly one worked question-and-answer pair for the model to replicate is one-shot prompting, and it is well suited to setting a style and structure from a single demonstration.
• D. Few-shot prompting requires multiple worked examples, so a task built around a single question-and-answer pair does not match that description.

Correct answer: D — Zero-shot instructions are the better fit here, since a one-off task with no need for a repeatable format does not justify the time spent building an example.

• A. Few-shot prompting is meant for tasks needing a consistent pattern across many outputs, not for a single quick memo summary that will not be repeated.
• B. A single memo summary does not inherently need to be split into dependent steps; prompt chaining is reserved for workflows with genuinely separate, sequential stages.
• C. A worked example is not automatically better than clear instructions; for a single, non-repeated task, the time spent building a one-shot example rarely pays off.
• D. For a one-off task with no need to lock in a reusable format, zero-shot instructions are the more efficient choice, since building a worked example mainly pays off when the pattern will be reused.

Correct answer: A — Define clear objectives and measurable goals tied to business outcomes before the pilot begins

• A. This is correct because setting measurable objectives tied to business outcomes before launch gives the team a concrete way to judge the pilot's success afterward.
• B. This is incorrect because deciding success informally after the fact gives leadership no objective evidence and makes the outcome hard to defend or repeat.
• C. This is incorrect because counting submitted prompts measures usage volume, not whether the pilot achieved any meaningful business outcome.
• D. This is incorrect because deferring success criteria until a full company-wide rollout means the pilot itself is never actually evaluated.

Correct answers: A, B, C — Defining objectives and key results tied to business outcomes before the initiative launches; Tracking user adoption and usage rates among the employees the solution was built for; Measuring the time saved or cost reduced on the specific tasks the solution targets

• A. This is correct because objectives and key results set before launch give the team a concrete benchmark against which to measure the initiative's later impact.
• B. This is correct because adoption and usage rates show whether the intended employees are actually getting value from the solution.
• C. This is correct because time saved or cost reduced on the targeted tasks directly reflects the business value the initiative was meant to create.
• D. This is incorrect because a raw token count reflects how much text was produced, not whether that output created any business value.
• E. This is incorrect because a model's parameter count is a technical property of the model itself, unrelated to the initiative's real-world impact.
• F. This is incorrect because tracking competitor announcements measures market activity, not the impact this initiative had on the company's own outcomes.

Correct answer: A — Explainability, because it concerns the system's ability to give understandable reasons behind a specific decision.

• A. Explainability is specifically about giving understandable reasons for a model's individual decisions, which is exactly what the bank needs to answer the customer's question about the rejection.
• B. Scalability concerns processing volume and speed, not the ability to justify why a specific decision was reached, so it does not address the customer's question.
• C. Portability concerns deploying a model across environments, which has nothing to do with explaining an individual decision to an affected customer.
• D. Availability concerns system uptime, which is unrelated to the bank's ability to explain the reasoning behind a specific rejected application.

Correct answers: A, B, C — The training dataset over-represents content and behavior from that one user segment relative to others.; Historical engagement patterns used as training signals already reflected an existing imbalance among segments.; A feedback loop reinforces the model's early skewed recommendations by using its own past outputs as new training signal.

• A. Over-representing one segment's content in training data gives the model disproportionate exposure to that segment's patterns, a well-established cause of biased recommendations.
• B. Training signals drawn from historically imbalanced engagement patterns pass that imbalance directly into the model's learned behavior, contributing to skewed output.
• C. A feedback loop where the model's own skewed outputs become future training data compounds an initial imbalance over time, a recognized cause of amplified bias.
• D. Server location affects latency and data residency, not the statistical composition of the training data, so it does not explain a skew toward one segment.
• E. Interface color scheme affects visual appeal, not the underlying data or logic driving which content gets recommended, so it is not a cause of bias.
• F. Team staffing levels affect maintenance capacity, not the statistical patterns in the training data, so it does not explain the described skew.

Correct answer: A — Scope the service account's permissions to only the deployment actions the pipeline performs, and nothing beyond that task.

• A. Scoping a service account to exactly the deployment actions it performs limits what an attacker could do if that account were ever compromised, which is the least-privilege approach IAM supports for automation.
• B. Giving the pipeline the same broad permissions as a cloud administrator means a compromise of that automated account could affect far more than deployment, well beyond what the task requires.
• C. Using a personal login for an automated pipeline ties production deployments to an individual's credentials and removes the isolation a dedicated service account provides.
• D. Leaving permissions unrestricted avoids future access requests but also removes any limit on what the pipeline, or anyone who compromises it, can do to the production system.

Correct answer: A — Controlling which identities can perform which actions on specific resources within an organization's cloud environment.

• A. IAM's core function is defining which identities, whether users or service accounts, can perform which actions on which resources, giving organizations fine-grained control over access.
• B. Automatic encryption of stored data is a separate storage-level protection; it operates independently of the identity-and-permission model that IAM is responsible for.
• C. Summarizing compute and storage consumption is a billing and usage-reporting function, not the access-control role IAM plays.
• D. Scanning container images for vulnerabilities is a separate security scanning capability, not the identity and permission control that defines IAM's core function.