25 free Google Generative AI Leader practice questions from the CertStash pack of 99, each with the correct answer and a full explanation. The rule is the same for every exam: ten free on the product page, 25 free here, all 99 in the pack, and each step includes the one before it.
Get all 99 questions (US$39) · Download the first 10 as a PDF · Read the study guide
Question 1
What is the definition of generative AI?
Show answer and explanation
Correct answer: B. A type of artificial intelligence that can create new content and ideas, including text, images, music, and code.
Generative AI is a type of artificial intelligence that creates new content and ideas, such as text, images, music, audio, video, and code. Rather than only classifying or predicting values from existing data, a generative model learns patterns from large amounts of training data and uses them to produce original outputs in response to a prompt. This ability to generate novel material is what separates generative AI from the broader fields of machine learning and deep learning, and from traditional predictive techniques such as regression.
Why the other options are wrong
- A. Fitting a line to observed data describes linear regression, a predictive technique that does not create new content.
- C. Interconnected nodes inspired by the human brain describe an artificial neural network, not generative AI.
- D. Autonomous learning and improvement with neural networks describes machine learning and deep learning in general, not the content-creating capability of generative AI.
Question 2
A company wants to use an AI agent to automate some tasks. They want everyone to understand the different functions of an AI agent.
What is the function of an AI agent in the context of gen AI?
Show answer and explanation
Correct answer: D. To analyze situations, use multiple tools, and make informed decisions without requiring constant human input.
An AI agent uses a model's reasoning ability to pursue a goal on its own. It observes and analyzes the situation, decides which steps to take, calls tools such as APIs, data stores, or other services, and acts on the results without needing a person to direct every step. This combination of reasoning, tool use, and autonomous decision making is what lets a company hand off multi-step tasks to an agent instead of simply sending prompts to a model and acting on the replies manually.
Why the other options are wrong
- A. Providing compute for training and serving models is the role of the infrastructure layer, such as TPUs and GPUs.
- B. Storing and managing datasets is the job of data storage and data management services.
- C. A user-friendly interface is a front end for a model and does not reason, use tools, or act autonomously.
Question 3
What is the definition of an AI agent?
Show answer and explanation
Correct answer: A. An application that learns how to achieve a goal based on inputs and tools available to it.
Google defines an AI agent as an application that tries to achieve a goal by observing the world and acting upon it using the tools it has at its disposal. The agent takes in information, reasons about what to do next, uses tools such as APIs, functions, and data stores to gather data or take actions, and adjusts its approach based on the results. This goal-directed, tool-using behavior is what distinguishes an agent from a simple chat interface or a fixed piece of application logic.
Why the other options are wrong
- B. A human who oversees a gen AI application is a human-in-the-loop role, not an AI agent.
- C. A user interface is only the means of interacting with an application and does not pursue goals or use tools.
- D. An agent is dynamic and goal-driven, so a static component that only displays information does not fit the definition.
Question 4
A company is developing an AI character for a video game. The AI character needs to learn how to navigate a complex environment and make decisions to achieve certain objectives within the game.
When the AI takes actions that lead to positive outcomes like finding a reward or overcoming an obstacle, it receives a positive score.
When it takes actions that lead to negative outcomes, like hitting a wall or losing progress, it receives a negative score. Through this process of trial and error, the AI gradually improves the character’s ability to play the game effectively.
What machine learning should the company use?
Show answer and explanation
Correct answer: B. Reinforcement learning
Reinforcement learning trains an agent by letting it interact with an environment and rewarding or penalizing its actions. The agent receives positive feedback for desirable outcomes, such as finding a reward or overcoming an obstacle, and negative feedback for undesirable ones, such as hitting a wall. Through repeated trial and error it learns a strategy that maximizes its total reward over time. A game character learning to navigate a complex environment and reach objectives through scored actions is a textbook reinforcement learning scenario.
Why the other options are wrong
- A. Unsupervised learning finds patterns in unlabeled data and does not use rewards or penalties for actions.
- C. Supervised learning needs labeled examples of correct answers rather than learning from trial-and-error feedback.
- D. Deep learning describes the use of multi-layer neural networks, not the reward-based learning approach the scenario describes.
Question 5
A research team has collected a large dataset of sensor readings from various industrial machines. This dataset includes measurements like temperature, pressure, vibration levels and electrical current recorded at regular intervals. The team has not yet assigned any labels or categories to these readings and wants to identify potential anomalies, malfunctions or natural groupings of machine behavior based on the sensor data alone.
What type of machine learning should they use?
Show answer and explanation
Correct answer: D. Unsupervised learning
Unsupervised learning works with unlabeled data and discovers structure in it without being told the correct answers. Clustering can reveal natural groupings of machine behavior, and anomaly detection can flag readings that differ from normal patterns, which may indicate malfunctions. Because the research team has not assigned any labels or categories to the sensor readings and wants insights from the data alone, unsupervised learning is the appropriate approach.
Why the other options are wrong
- A. Deep learning is a family of neural network techniques, not the learning approach defined by the absence of labels.
- B. Reinforcement learning trains an agent through rewards for actions in an environment, which does not apply to analyzing static sensor data.
- C. Supervised learning requires labeled data, and the team has not labeled any of the readings.
Question 6
A highly regulated financial institution wants to use Gemini as the core decision engine for a loan approval system that will deterministically approve or reject loan applications based on a strict set of predefined criteria.
Why is this an inappropriate use case for Gemini?
Show answer and explanation
Correct answer: B. Gemini’s designed for flexible content generation and inference, not rigid rule-based decisions.
Gemini is a generative model that produces probabilistic outputs, which makes it strong at flexible tasks such as summarizing, reasoning over varied inputs, and generating content. A loan approval engine that must deterministically approve or reject applications against a strict set of predefined criteria needs the same input to produce the same result every time, with fully traceable logic. That requirement is better met by a rule-based system, so using Gemini as the core decision engine is a poor fit for the task.
Why the other options are wrong
- A. Cost and complexity are not the reason; the mismatch is between probabilistic generation and rigid rule-based decisions.
- C. Gemini can process structured and numerical data, so this is not the limitation that makes the use case inappropriate.
- D. Gemini can be connected to enterprise data sources through Google Cloud services, so integration is not the issue.
Question 7
What is an example of unsupervised machine learning?
Show answer and explanation
Correct answer: C. Analyzing customer purchase patterns to identify natural groupings.
Analyzing customer purchase patterns to identify natural groupings is a clustering task, a classic form of unsupervised learning. The data has no predefined labels for customer segments, so the algorithm discovers similarities and groups customers on its own. Businesses often use these segments for targeted marketing or product recommendations. The other scenarios all rely on known outcomes in historical data, such as past renewal status, labeled image categories, or recorded sales figures, which makes them supervised learning tasks.
Why the other options are wrong
- A. Predicting renewal from past renewal status uses labeled outcomes, which makes it supervised classification.
- B. Training on images with labeled categories is supervised learning by definition.
- D. Forecasting sales from historical figures and marketing spend learns from known target values, which is supervised regression.
Question 8
A user asks a generative AI model about the scientific accuracy of a popular science fiction movie. The model confidently states that humans can indeed travel faster than light, referencing specific but entirely fictional theories and providing made-up explanations of how this is achieved according to the movie's "established science." The model presents this information as factual, without indicating that it originates from a fictional work.
What type of model limitation is this?
Show answer and explanation
Correct answer: B. Hallucination
A hallucination occurs when a generative model produces output that is false or fabricated but presents it confidently as fact. Here the model invents scientific theories, supplies made-up explanations for faster-than-light travel, and fails to signal that the material comes from a fictional work. Models generate statistically likely text rather than verifying truth, so they can blend fiction with reality. Grounding responses in trusted sources and having people review critical outputs are common ways to reduce hallucinations.
Why the other options are wrong
- A. Bias is a systematic skew or unfair preference in outputs, not the invention of false facts.
- C. Knowledge cutoff concerns missing information about events after the training data ends, which is not the issue here.
- D. Data dependency describes problems caused by the quality or coverage of training data, while this case is about fabricating content and presenting it as fact.
Question 9
What does a diffusion model do?
Show answer and explanation
Correct answer: A. Generates high-quality content by refining noise into structured data.
A diffusion model is a generative model that creates content by starting from random noise and refining it step by step into structured output. During training, the model learns how noise is gradually added to real data and how to reverse that process. At generation time it applies the learned denoising steps to turn noise into a coherent result, such as a high-quality image. Diffusion models are widely used for image generation and are also applied to video and audio.
Why the other options are wrong
- B. Storing and managing structured data is the role of databases and data platforms, not a generative model.
- C. Analyzing data to predict trends describes predictive analytics, not content generation from noise.
- D. Optimizing business processes and resource allocation is an operations task, not what a diffusion model does.
Question 10
An organization is collecting data to train a generative AI model for customer service. They want to ensure security throughout the ML lifecycle.
What is a critical consideration at this stage?
Show answer and explanation
Correct answer: D. Implementing access controls and protecting sensitive information within the training data.
During data collection and preparation, the main security concern is protecting the training data itself. Customer service data often contains personal or confidential information, so the organization should restrict who can access the datasets, protect sensitive fields through methods such as encryption, masking, or de-identification, and follow its privacy and compliance obligations. Securing data at this early stage of the ML lifecycle prevents leaks and keeps sensitive information from being learned and later exposed by the model.
Why the other options are wrong
- A. Applying software patches is an ongoing operational task for deployed systems, not the critical concern while collecting training data.
- B. Monitoring outputs for unexpected results belongs to the deployment and monitoring stage, after the model is trained.
- C. Ethical guidelines for responses address responsible AI behavior, not securing the data being collected.
What people say about the full Generative AI Leader pack 5.0 from 1
This exam wasn’t easy, but the practice questions I used helped me a lot. I practiced with exam questions, and it made the real exam feel much more manageable.
Question 11
The office of the CISO wants to use generative AI (gen AI) to help automate tasks like summarizing case information, researching threats, and taking actions like creating detection rules.
What agent should they use?
Show answer and explanation
Correct answer: B. Security agent
Security agents apply generative AI to cybersecurity work. They help security teams summarize case information, research and investigate threats, and take actions such as writing detection rules, which reduces manual effort and speeds up response. Because the office of the CISO wants to automate exactly these security operations tasks, a security agent is the appropriate choice. Google Cloud brings these capabilities into its security products through Gemini.
Why the other options are wrong
- A. A data agent helps users analyze and query data, not investigate threats or create detection rules.
- C. A customer service agent answers customer questions and resolves support issues, which is unrelated to security operations.
- D. A code agent helps developers write, review, and debug application code rather than handling security cases and threat research.
Question 12
An organization needs an AI tool to analyze and summarize lengthy customer feedback text transcripts. You need to choose a Google foundation model with a large context window.
What foundation model should the organization choose?
Show answer and explanation
Correct answer: A. Gemini
Gemini models are known for their long context windows, which allow them to take in very large amounts of text in a single request. This lets the organization provide entire lengthy customer feedback transcripts and have the model analyze themes, sentiment and key points and produce accurate summaries without splitting the material into small pieces. Because the task is text analysis and summarization of long documents, Gemini is the foundation model that fits the requirement.
Why the other options are wrong
- B. CodeGemma is specialized for code generation and completion, not analysis and summarization of lengthy customer feedback.
- C. Imagen generates images from text and is not designed to analyze or summarize text transcripts.
- D. Chirp is a family of speech models for converting between audio and text, not for summarizing text that is already transcribed.
Question 13
An order fulfillment team has an agent that automatically processes orders, updates inventory, sends shipping notifications, and handles returns.
What type of agent is this?
Show answer and explanation
Correct answer: D. A workflow agent
A workflow agent automates a multi-step business process from start to finish. In this scenario the agent processes orders, updates inventory, sends shipping notifications, and handles returns, taking actions across several systems in a defined sequence. Its focus is completing operational tasks within a process rather than holding conversations with users or helping individual employees with their own work, which is what makes it a workflow agent.
Why the other options are wrong
- A. A conversational agent centers on natural-language dialogue with users, not on running back-end order processes.
- B. A customer service agent focuses on answering customer questions and resolving issues, not on executing fulfillment operations.
- C. An employee productivity agent helps individual workers with tasks such as drafting, searching, or summarizing, not end-to-end order processing.
Question 14
What are core hardware components of the infrastructure layer in the generative AI landscape?
Show answer and explanation
Correct answer: B. TPUs and GPUs
The infrastructure layer is the foundation of the generative AI landscape, supplying the computing power and hardware needed to train and serve models. Its core hardware components are accelerators such as Tensor Processing Units (TPUs), which Google designed specifically for machine learning workloads, and Graphics Processing Units (GPUs). These processors handle the large-scale parallel computation that foundation models require, and the model, platform, agent, and application layers all run on top of them.
Why the other options are wrong
- A. User interfaces belong to the application layer, where people interact with gen AI, not the hardware infrastructure.
- C. Tools and services for building AI models make up the platform layer rather than the hardware infrastructure.
- D. Pre-trained models form the model layer, which runs on the infrastructure rather than being part of it.
Question 15
A software engineering team is experimenting with generative AI within their coding process.
What is a benefit of using gen AI to create unit tests for code?
Show answer and explanation
Correct answer: A. Reducing manual validation effort.
Writing unit tests by hand is repetitive and time-consuming, and gen AI can draft test cases quickly by reading the code and proposing checks for normal paths and edge cases. This reduces the manual effort developers spend validating their code, freeing them to focus on design and more complex problems. Developers still review the generated tests, because AI output can contain mistakes and tests can only reveal problems in the scenarios they cover. The realistic benefit is faster, less laborious validation, not perfect code or the removal of human oversight.
Why the other options are wrong
- B. Generating tests checks existing code behavior and does not by itself make the code simpler.
- C. Tests, whether AI generated or not, cannot guarantee code is free of errors.
- D. Generated tests still need human review, so human involvement remains necessary.
Question 16
A large multinational corporation with geographically dispersed teams struggles with knowledge silos and inconsistent access to crucial internal information.
What is a key business benefit of using Google Agentspace in this scenario?
Show answer and explanation
Correct answer: B. Seamless knowledge sharing and collaboration across internal systems.
Google Agentspace brings together enterprise search, AI assistance and agents on top of an organization's internal systems. By connecting to the applications and repositories where company knowledge lives, it lets employees in any location find and use the information they need through a single interface. For a multinational with dispersed teams and knowledge silos, the key business benefit is seamless knowledge sharing and collaboration across internal systems, giving everyone consistent access to crucial information.
Why the other options are wrong
- A. Automating employee performance reviews is not the purpose of Agentspace and does not address knowledge silos.
- C. IT infrastructure management across offices is a different concern from employees' access to internal knowledge.
- D. Encryption and compliance for internal communications are security functions, not the knowledge access problem Agentspace solves.
Question 17
A company is evaluating different generative AI (gen AI) platforms and wants to understand the role of the infrastructure layer in supporting the development and deployment of gen AI models.
What is the function of the infrastructure layer in the gen AI landscape?
Show answer and explanation
Correct answer: C. To provide the compute resources needed to run and train AI models and store training data.
The infrastructure layer is the foundation of the gen AI landscape. It supplies the compute resources, such as TPUs and GPUs, along with the storage and networking needed to train and run AI models and to hold the large datasets used for training. Every layer above depends on it: the model layer offers pre-trained models, the platform layer provides tools for building and deploying solutions, and the application layer delivers user-facing experiences. When evaluating platforms, the infrastructure layer is where a company assesses raw processing power, scalability and data storage capacity.
Why the other options are wrong
- A. Providing access to pre-trained models is the role of the model layer.
- B. A user-friendly interface for interacting with models belongs to the application layer.
- D. Tools and resources for building and deploying gen AI applications describe the platform layer.
Question 18
A company wants a generative AI platform that provides the infrastructure, tools, and pre-trained models needed to build, deploy, and manage its generative AI solutions.
Which Google Cloud offering should the company use?
Show answer and explanation
Correct answer: D. Vertex AI
Vertex AI is Google Cloud's unified AI platform for building, deploying and managing machine learning and generative AI solutions. It provides access to pre-trained foundation models through Model Garden, tools for prompting, tuning and evaluating models, capabilities for building agents and managed infrastructure for deploying and monitoring models in production. Because the company wants one platform that combines infrastructure, tools and pre-trained models across the full generative AI lifecycle, Vertex AI is the offering to use.
Why the other options are wrong
- A. BigQuery is a data warehouse for storing and analyzing data, not an end-to-end platform for building and managing generative AI solutions.
- B. GKE runs containerized applications but does not provide pre-trained models or dedicated generative AI tooling.
- C. Cloud Storage stores objects and files and does not offer models or tools for building and deploying AI solutions.
Question 19
A pharmaceutical company’s research and development department spends significant time manually reviewing new scientific papers to identify potential drug targets. They need a solution that can answer questions about these documents and provide summarized insights to researchers without requiring extensive coding expertise.
What should the organization do?
Show answer and explanation
Correct answer: A. Use Vertex AI Agent Builder to create a custom AI agent.
Vertex AI Agent Builder lets organizations create AI agents grounded in their own data with little or no code. The R&D team can connect the scientific papers as a data store and build an agent that answers researchers' natural language questions and returns summarized insights drawn from those documents. This directly reduces the time spent on manual review while keeping the solution accessible to staff without extensive coding expertise, which is exactly what the company needs.
Why the other options are wrong
- B. Keyword-based search only helps locate documents; it does not answer questions or provide summarized insights from the papers.
- C. Gemini for Google Workspace assists individual users inside Workspace apps, rather than providing a dedicated question-answering solution over a body of research papers.
- D. Classifying papers into predefined research areas with AutoML sorts documents but does not answer questions or summarize their findings.
Question 20
What is a characteristic of Google Cloud as a generative AI company?
Show answer and explanation
Correct answer: B. Google Cloud has an AI-first focus that enables innovation, with continuous updates and broad integration across its platform.
Google Cloud positions itself as an AI-first company, building on decades of Google AI research and bringing new models and capabilities to customers on an ongoing basis. Generative AI is integrated broadly across its offerings, from infrastructure and Vertex AI to Google Workspace and its security and data products. This AI-first focus, with continuous updates and broad integration across the platform, is the characteristic that defines Google Cloud as a generative AI company.
Why the other options are wrong
- A. Google Cloud takes an open approach that includes open models and third-party models, rather than relying only on proprietary, closed-source technology.
- C. Google Cloud does not isolate all generative AI models and data from external networks; fully disconnected operation is a specialized deployment option such as Google Distributed Cloud air-gapped, so this absolute claim is inaccurate.
- D. AI agents still require configuration, governance and oversight, so zero management overhead is not a real characteristic of Google Cloud.
Question 21
A company is developing a generative AI application to analyze customer feedback collected through online surveys. Stakeholders are concerned about potential privacy risks associated with this data, as the feedback contains personally identifiable information. They need to mitigate these risks before using the data to train the AI model.
What action should the company prioritize?
Show answer and explanation
Correct answer: D. Applying data anonymization techniques to remove or obscure sensitive data.
Data anonymization removes or obscures personally identifiable information, using techniques such as redaction, masking, or tokenization, before the data is used for training. This directly addresses the stakeholders' concern because the model does not learn from the removed or obscured details, which greatly reduces the risk of exposing them in its outputs. Anonymizing the survey feedback lets the company keep the valuable content of the responses while reducing privacy risk, which makes it the action to prioritize before training.
Why the other options are wrong
- A. Collecting only quantitative data in future surveys does nothing about the PII already in the current data and discards useful feedback.
- B. Access controls limit who can view the raw data, but the PII would still be present in the data used to train the model.
- C. A large and diverse dataset helps with model quality and fairness, but it does not remove personally identifiable information.
Question 22
What is the continuous cycle of observing, interpreting, planning, and acting that makes up an AI agent's “thinking process”?
Show answer and explanation
Correct answer: B. Reasoning loop
An AI agent works toward a goal through a reasoning loop: it observes its environment and inputs, interprets what it has observed, plans the next steps, and acts, often by calling tools. It then observes the results of that action and repeats the cycle until the goal is reached. This iterative process is what lets an agent handle multistep tasks and adapt as new information arrives, and it is often described as the agent's thinking process. Training, integration and interfaces are separate aspects of building or using an AI system, not the cycle the agent runs while working.
Why the other options are wrong
- A. Model training happens before deployment and is not the cycle an agent runs while completing a task.
- C. Platform integration concerns connecting systems, not the agent's cycle of observing, planning and acting.
- D. A user interface is how people interact with an application, not the agent's internal decision cycle.
Question 23
What is a key advantage of using Google’s custom-designed TPUs?
Show answer and explanation
Correct answer: C. TPUs are specialized AI processors that excel at parallel processing for machine learning workloads.
Tensor Processing Units are custom-designed AI accelerators that Google built specifically for machine learning. Their architecture is optimized for the large matrix operations at the heart of neural networks, performing massive amounts of computation in parallel. This lets TPUs train and serve large models, including generative AI foundation models, quickly and cost-effectively at scale. The key advantage is therefore specialized, highly parallel processing for machine learning workloads, rather than general-purpose computing, storage or edge deployment.
Why the other options are wrong
- A. TPUs are compute accelerators and do not add storage capacity or speed up data retrieval.
- B. Cloud TPUs are data center accelerators for large-scale training and inference, not lightweight edge processors.
- D. TPUs are specialized for machine learning math rather than for speeding up general virtual machine workloads.
Question 24
A company collects customer feedback through open-ended survey questions where customers can write detailed responses in their own words, such as “The product was easy to use, and the customer support was excellent, but the delivery took longer than expected.”
What type of data is this?
Show answer and explanation
Correct answer: D. Unstructured data
Unstructured data has no predefined format or schema that fits neatly into rows and columns. Open-ended survey answers written in customers' own words are free-form text that can vary in length, wording, and topic, so they are unstructured. This kind of data is well suited to generative AI and natural language processing, which can extract themes, sentiment, and key points from text that traditional structured analysis cannot easily handle.
Why the other options are wrong
- A. Quantitative data consists of numerical measurements, not free-form written responses.
- B. Structured data follows a fixed schema such as database tables, which open-ended text does not.
- C. Labeled data carries tags such as categories or sentiment, and these responses have not been tagged.
Question 25
A company wants to adopt generative AI and is concerned about vendor lock-in. They want to maintain flexibility in their technology stack.
What Google Cloud strength would ease their concerns?
Show answer and explanation
Correct answer: C. Google Cloud’s AI solutions have an open approach that supports customer choice across offerings.
Google Cloud takes an open approach to AI that emphasizes customer choice. Customers can use Google's own models, open models and third-party models, for example through Model Garden, and build with open-source frameworks and standards that work across environments. This flexibility lets an organization select and combine the technologies that best fit its needs and change course later, which directly addresses concerns about vendor lock-in and the desire for a flexible technology stack.
Why the other options are wrong
- A. Strict reliance on proprietary technologies would increase vendor lock-in rather than ease that concern.
- B. Removing customization and integration options reduces flexibility, which is the opposite of what the company wants.
- D. Replacing human jobs with automation has nothing to do with vendor lock-in and does not describe a Google Cloud strength.
That was 25 of 99.
The full Google Generative AI Leader pack has all 99 questions, each with the answer, the explanation and why the other options are wrong, plus a questions-only copy for timed runs. US$39, paid once, with free monthly updates and a pass-or-your-money-back guarantee.
Not ready yet? Get the first 49 free as one PDF by email. Get the Half Pack →