Free Google Generative AI Leader Sample Questions

Correct answer: B — Context window

• A. Modality refers to the types of data a model can handle, such as text, images, or audio. While the task requires a model with text modality, it does not determine how much text the model can process in a single request.
• B. The context window represents the maximum amount of input text (measured in tokens) that a model can process at one time. For 200-page documents, a large context window is the critical factor that allows the entire document to be ingested and analyzed without truncation or the need for multi-step processing.
• C. Image generation performance relates to the model's ability to create visual content. This is irrelevant to the requirement of processing and summarizing long-form text legal documents.
• D. Reinforcement learning capabilities, such as Reinforcement Learning from Human Feedback (RLHF), are used to align a model's outputs with human preferences and improve safety or style. However, these techniques do not expand the model's physical input capacity or context window size.

Correct answer: C — Prompt tuning / Fine-tuning

• A. Data ingestion is the process of collecting and importing raw data into a system for storage or immediate use. It does not involve modifying the internal parameters or the behavior of a pre-trained model through task-specific examples.
• B. Unsupervised learning involves training models on unlabeled data to discover underlying patterns or structures without explicit human guidance. In contrast, the scenario describes using specific examples to adapt a model's behavior for a specialized task.
• C. This describes the process of adapting a pre-trained model to a specialized domain or task. Fine-tuning involves updating model weights on a small, labeled dataset, while prompt tuning (including few-shot learning) provides context to guide the model's output. Both are efficient ways to leverage a model's existing knowledge for new requirements without retraining from scratch.
• D. Model deployment is the final stage of the machine learning lifecycle where a finished model is integrated into a production environment to serve live requests. It is the act of making a model available for use, not the act of adapting its behavior using specific examples.

Correct answer: B — Generative AI can create net-new content like text and images.

• A. Incorrect. Generative AI is not exclusively used for numerical predictions; numerical forecasting and regression are typical tasks for traditional supervised machine learning models. Generative AI is designed to produce diverse forms of content such as text, code, and imagery.
• B. Correct. Generative AI models are specifically designed to learn patterns from training data to create net-new content. This makes them ideal for creative tasks like writing marketing copy, as they can synthesize original and contextually appropriate language.
• C. Incorrect. Grouping data into clusters is a task associated with unsupervised learning (clustering) within traditional AI. While generative models understand data distributions, their primary purpose is to synthesize new examples rather than merely organizing existing ones into groups.
• D. Incorrect. Generative AI requires massive amounts of training data to learn the structures and patterns of language or imagery. Without extensive prior data, these models cannot produce coherent or meaningful outputs.

Correct answer: C — Platforms

• A. The Infrastructure layer refers to the underlying compute, storage, and networking resources (such as GPUs, TPUs, and cloud clusters) used to host and run models. It is a lower-level layer that does not typically provide developer-facing features like prompt management or safety filters.
• B. The Models layer consists of the actual LLMs (the algorithms and weights). While models generate responses, they do not inherently provide the management tools, orchestration APIs, or integrated safety guardrails required for application development.
• C. The Platforms layer provides developer-facing services and tools to integrate AI models into business logic. This includes easy-to-use APIs, prompt orchestration, safety and content filters, and monitoring tools, making it the ideal layer for teams building complex integrations like NPC dialogue systems.
• D. The Applications layer represents the end-user products (such as the game itself or a specific chatbot interface). While the company is building an application, the question asks for the layer that provides the services and tools necessary to connect that application to an LLM.

Correct answer: A — Infrastructure

• A. Correct. The Infrastructure layer encompasses the physical and virtual resources required to run AI applications, including compute hardware, networking, and storage. Moving hardware to the factory floor (edge computing) to reduce latency is a direct modification of the infrastructure layer.
• B. Incorrect. The Models layer refers to the AI algorithms themselves, including their architecture, training, and weights. Relocating the physical hardware does not change the mathematical structure or the weights of the AI model.
• C. Incorrect. The Platforms layer consists of the software frameworks, orchestration tools (like Kubernetes), and development environments used to deploy and manage AI. While a platform might run on the new hardware, the act of moving physical servers is an infrastructure change.
• D. Incorrect. The Agents layer refers to autonomous or semi-autonomous software entities that perform specific tasks or workflows. Changing the location of compute hardware may improve agent performance via reduced latency, but it does not modify the design or logic of the agents themselves.

Correct answer: A — Completeness

• A. Correct. Completeness refers to the degree to which all required data values are present and sufficient for the intended task. With 40% of the income data missing, the dataset lacks the necessary information to represent the population accurately. This level of missingness can introduce bias and significantly degrade the model's predictive performance.
• B. Incorrect. Consistency refers to the uniformity of data across different systems or within the same dataset (e.g., ensuring a customer's address is the same in two different tables). The issue here is the total absence of data, not a conflict in existing data values.
• C. Incorrect. Format refers to the structure, encoding, or organization of the data (e.g., using a numeric field instead of a string). If the incomes were present but entered in different currencies or date formats, it would be a format issue. In this case, the data is simply missing.
• D. Incorrect. Relevance refers to whether the data collected is pertinent and useful for the specific problem being solved. Income is highly relevant for predicting loan defaults; the problem is that the institution failed to capture this relevant data for a large portion of the records.

Correct answer: B — Structured data

• A. Incorrect. Unstructured data refers to information that does not have a predefined data model or organized format, such as free-form text, images, or audio. The bank's data is organized into specific columns, which contradicts the definition of unstructured data.
• B. Correct. Structured data is organized into a fixed schema, typically in rows and columns, making it highly searchable and analyzable. Because this bank information is organized into defined fields, it is straightforward to query with SQL, aggregate in spreadsheets, and feed into models that expect tabular inputs.
• C. Incorrect. Unlabeled data typically refers to datasets lacking target tags for supervised machine learning tasks. While the data may or may not be labeled for a specific ML task, it is primarily categorized as 'structured' due to its tabular format and defined column headers.
• D. Incorrect. Multimedia data describes content such as audio, video, or images. The bank's dataset consists of textual and numeric fields in a table, which does not constitute multimedia assets.

Correct answer: C — Veo

• A. Gemini is a family of multimodal models capable of processing and generating text, code, and images. While it has advanced reasoning capabilities and can help craft scripts or prompts, it is not Google's primary model specifically dedicated to high-fidelity video generation from text.
• B. Imagen is Google’s text-to-image diffusion model designed to generate high-quality still images from text prompts. It is not the dedicated model for generating cinematic video content.
• C. Veo is Google’s most capable model specifically designed for high-quality, text-to-video generation. It can generate 1080p resolution videos that follow a provided script or description, making it the correct choice for creating a promotional video clip.
• D. Gemma is a family of lightweight, state-of-the-art open models built from the same research and technology used to create Gemini. It is primarily used for text-based tasks and does not specialize in producing video content.

Correct answer: D — Gemini

• A. Gemma is a family of lightweight, open-weight text-to-text models optimized for language understanding and generation. It lacks native multimodal capabilities and cannot process live audio feeds directly.
• B. Imagen is a text-to-image generative model. It is designed to generate or edit visual content based on text prompts and is not intended for audio processing, transcription, or translation tasks.
• C. Veo is Google's most advanced generative video model. While it focuses on the video modality, its primary function is video generation (text-to-video) rather than the transcription, translation, and summarization of live audio streams.
• D. Gemini is a natively multimodal foundation model designed from the ground up to reason across text, images, video, code, and audio. Its ability to process audio inputs directly makes it the ideal choice for real-time transcription, multilingual translation, and summarization of audio feeds.

Correct answer: C — Imagen

• A. Veo is Google's generative video model designed for creating high-quality, cinematic video content. It is not optimized for editing static images or performing text-guided photo manipulation.
• B. Gemma is a family of lightweight, open-source large language models (LLMs). While highly capable for text generation and reasoning, it is not an image generation or editing model.
• C. Imagen is Google's foundational text-to-image model. Through Vertex AI, Imagen provides advanced editing capabilities, such as inpainting and outpainting, which allow users to modify specific areas of an existing image (like clothing color or patterns) using natural language descriptions.
• D. Gemini is a multimodal model capable of understanding and reasoning across text, images, video, and code. While it can analyze and describe images, the specialized task of high-fidelity, pixel-level image editing is the primary domain of the Imagen model family.

Correct answers: A, B, D — Extensions; Data stores; Functions

• A. Correct. Extensions (such as Vertex AI Extensions) allow agents to connect to external APIs and services. They provide the mechanism for agents to perform real-world actions like booking a flight, searching a specific database, or interacting with third-party software.
• B. Correct. Data stores (including knowledge bases and vector databases) enable agents to perform Retrieval-Augmented Generation (RAG). This allows the agent to retrieve relevant, grounded information from private or specific datasets to answer queries accurately.
• C. Incorrect. Hyperparameter tuning is an optimization process used during the development and training of a machine learning model to improve its performance. It is not a mechanism for an agent to interact with external environments at runtime.
• D. Correct. Functions (often referred to as function calling) allow agents to invoke predefined code or logic. This enables the agent to translate a user's natural language request into a structured programmatic call to perform tasks like calculations or database updates.
• E. Incorrect. Gradient descent is an optimization algorithm used during the training phase to minimize the model's loss function. It is not a tool or interface used by agents to achieve tasks or interact with external systems.

Correct answers: A, B — Document AI API; Cloud Vision API

• A. Correct. Document AI is specifically designed to process scanned documents and PDFs, offering OCR and structured data extraction (e.g., form parsers and specialized processors) to pull fields and metadata from forms. It is the most appropriate choice for handling scanned PDF forms and extracting their structured content.
• B. Correct. Cloud Vision API is built to analyze images of physical objects, offering capabilities like label detection, object detection, and optical character recognition for images. It is the primary API used to handle images of physical products for recognition and metadata extraction.
• C. Incorrect. The Speech-to-Text API is used for converting spoken language into written text and does not provide capabilities for processing scanned PDFs or static images.
• D. Incorrect. The Text-to-Speech API is used for converting written text into spoken language (audio generation). It does not handle the input processing of scanned PDF forms or product images.
• E. Incorrect. The Cloud Video Intelligence API is intended for analyzing video content (e.g., shot detection, object tracking over time, and explicit content detection) rather than static scanned PDFs or standalone product images. For still images and documents, Vision API and Document AI are more appropriate.

Correct answer: B — Conversational Agents

• A. Conversational Insights is typically used to analyze and derive insights from conversation data, such as sentiment analysis and trend identification. It helps measure and improve conversations after they have occurred but does not provide the dialog management and flow-building capabilities needed for a voice bot.
• B. Conversational Agents (often associated with Dialogflow and Google Cloud’s Vertex AI/CCAI suite) are specifically designed to build multi-turn conversational flows. They handle state management, slot-filling, and branching logic, making them the ideal tool for constructing question sequences and voice prompts needed for eligibility checks.
• C. Agent Assist is designed to support human customer service representatives in real-time by surfacing suggested responses and knowledge base articles. It augments human performance rather than serving as the primary engine for building autonomous, multi-turn conversational bots.
• D. Google Search is a web-based information retrieval tool. It lacks the intent recognition, dialog management, and voice integration features required to orchestrate a structured multi-turn conversation between a user and a bot.

Correct answer: C — Vertex AI Search

• A. Incorrect. Conversational Insights is focused on extracting analytics, sentiment, and trends from conversation data (such as contact center logs) rather than providing a retrieval-augmented conversational search over a custom media archive.
• B. Incorrect. Agent Assist is specifically built to support contact center agents during live interactions by surfacing real-time suggestions and context. It is not designed to provide a subscriber-facing conversational search and summarization service over a media repository.
• C. Correct. Vertex AI Search is designed for enterprise search applications, allowing users to perform semantic and conversational searches over large organizational datasets. It enables RAG (Retrieval-Augmented Generation) to provide summarized answers that are grounded in the specific articles and content provided by the company.
• D. Incorrect. Google Cloud Contact Center as a Service (CCaaS) is a comprehensive solution for managing contact center operations and customer support workflows. It is not a specialized tool for conversational search and summarization of a media archive.

Correct answer: A — Pre-built and customizable AI solutions.

• A. Google Cloud offers purpose-built solutions like Contact Center AI, Dialogflow, and Vertex AI Search and Conversation. These are specifically designed for customer service scenarios, allowing departments to deploy structured, conversational agents quickly and customize them without building the underlying architecture or integration workflows from scratch.
• B. Tensor Processing Units (TPUs) are specialized hardware designed to accelerate machine learning workloads. While they provide high performance for model training and inference, they do not offer the pre-built application architecture or customer-service-specific workflows required in this scenario.
• C. While Google Cloud supports open-source models, using them typically requires significant engineering effort to fine-tune, integrate, and build a customer-service workflow from the ground up. This contradicts the customer's preference for a solution that is already structured.
• D. Bare-metal infrastructure provides low-level control over physical hardware resources. It lacks the application-layer intelligence and pre-configured agents needed for customer service, requiring the organization to build the entire solution from the ground up, which increases operational complexity.

Correct answer: A — Its enterprise-ready, reliable, and globally scalable infrastructure.

• A. Correct. Google Cloud's AI platform is built on enterprise-ready, reliable, and globally scalable infrastructure. This include multi-region/multi-zone deployments, global load balancing, and robust SLAs that provide the high availability and fault tolerance required for mission-critical, 24/7 global operations.
• B. Incorrect. Focusing on experimental or beta-only features does not guarantee production-grade reliability. Such features often lack stability and are not covered by the standard production SLAs necessary for critical banking applications.
• C. Incorrect. Relying on single-region data centers increases the risk of downtime during regional outages and limits redundancy. High availability for global services requires a multi-region architecture rather than a single-region approach.
• D. Incorrect. While Google Cloud supports open source, it is not an 'open-source only' stack. Reliability in an enterprise context is derived from the underlying managed infrastructure and operational support, not exclusively from the software licensing model.

Correct answers: A, B, C — Responsible AI principles and safety filters; Secure and private data handling; Highly scalable infrastructure

• A. Correct. Google Cloud integrates Responsible AI principles and safety filters (e.g., in Vertex AI) to help detect and mitigate risks such as bias and toxicity. These guardrails ensure fairness, explainability, and safety, which are essential for enterprise compliance and risk management.
• B. Correct. Secure and private data handling is a cornerstone of the platform. Features like encryption at rest and in transit, VPC Service Controls, and Customer-Managed Encryption Keys (CMEK) ensure that enterprise data is protected and meets regulatory privacy requirements.
• C. Correct. Google Cloud provides a highly scalable infrastructure, including purpose-built AI accelerators like TPUs and GPUs, along with managed services like Vertex AI that support large-scale training and inference workloads required for production.
• D. Incorrect. Mandatory public sharing of training data is not a characteristic of Google Cloud; enterprises typically keep training data private and under their exclusive control. Google Cloud provides secure, private data storage and does not require public disclosure of customer data.
• E. Incorrect. Google Cloud provides robust, enterprise-grade Service Level Agreements (SLAs) for production environments. Reliability and performance guarantees are standard for Google Cloud services, not experimental-only terms.
• F. Incorrect. Google Cloud offers comprehensive Identity and Access Management (IAM), fine-grained role-based access controls, and audit logging to manage and secure resources, which is fundamental for an enterprise-ready environment.

Correct answer: A — Vertex AI Model Garden

• A. Vertex AI Model Garden is the single point of entry to discover, test, and deploy a wide variety of foundation models, including Google's own models, open-source models, and third-party models (like Anthropic's Claude). It provides the necessary security, governance, and hosting infrastructure to prototype applications within the Google Cloud environment.
• B. Vertex AI Search is a tool for building search and retrieval-augmented generation (RAG) applications over private enterprise data. It is not the primary catalog for discovering and deploying a specific foundation model.
• C. Vertex AI Agent Builder is a high-level platform for orchestrating AI agents and workflows. While it uses models, it is not the first stop for finding and deploying a specific third-party foundation model for prototyping.
• D. Cloud Functions is a serverless compute service for small snippets of code triggered by events. It is not designed to host large-scale foundation models or serve as a discovery catalog for third-party AI models.

Correct answers: A, C, E — Building conversational agents that can take action; Orchestrating multiple tools and APIs to fulfill user requests; Creating virtual assistants that interact with enterprise systems

• A. Correct. Vertex AI Agent Builder is specifically designed to create conversational agents that can engage in natural language dialogue and perform actions. It focuses on high-level dialog management and safe execution of actions based on user intent.
• B. Incorrect. Training custom image recognition models is a computer vision task best suited for Vertex AI Training or AutoML. Agent Builder is centered on conversational workflows and tool orchestration rather than low-level model training.
• C. Correct. A core strength of Agent Builder is its ability to orchestrate multiple tools, APIs, and services. This allows agents to fetch data from external systems or trigger business workflows to fulfill complex user requests.
• D. Incorrect. Infrastructure migration for virtual machines is handled by Google Cloud's migration services (like Migrate to Virtual Machines). This is outside the scope of AI agent development.
• E. Correct. Developing virtual assistants that integrate with enterprise systems (such as CRMs, ERPs, and internal databases) is a primary use case. Agent Builder provides the connectors and secure orchestration needed for enterprise-grade automation.
• F. Incorrect. Designing network topologies is a infrastructure and networking engineering task. Vertex AI Agent Builder is a software development platform for generative AI applications and does not have network design capabilities.

Correct answer: B — Gems

• A. Multimodal search is designed to search across and retrieve content from various input types like text, images, and videos. It does not provide a mechanism for creating, saving, or reusing specific assistant personas.
• B. Gems in Gemini Advanced allow users to create and save custom versions of Gemini with specific instructions, personas, and specialized knowledge. This feature enables teams to encapsulate a brand's voice—such as a lead copywriter—into a reusable tool that maintains consistency without requiring the user to rewrite complex prompts for every new session.
• C. Gemini for Google Workspace refers to the integration of AI capabilities directly into applications like Docs, Gmail, Sheets, and Slides. While it facilitates AI-powered workflows within these apps, 'Gems' is the specific feature within the Gemini ecosystem used to build and manage persistent, reusable persona templates.
• D. NotebookLM (and any associated APIs) focuses on knowledge-centric exploration, allowing users to interact with and synthesize information from specific uploaded documents. It is not the standard feature used for creating branded, reusable assistant personas for general team interaction.

Correct answers: A, B, C — Draft a new document based on a text prompt.; Rewrite existing text to change the tone or make it more concise.; Summarize the content of the current document.

• A. Correct. Gemini in Google Docs can generate new content from a user prompt, producing a draft document or sections of text based on instructions and templates. This core productivity feature leverages generative AI to help users overcome the 'blank page' problem.
• B. Correct. Gemini offers built-in editing capabilities to rewrite existing text. Users can request the AI to change the tone (e.g., make it more formal or professional), make the text more concise, or expand upon existing points.
• C. Correct. Gemini can analyze the active document to extract key points and generate a summary. This is a standard feature designed to help users quickly digest long documents or create executive overviews.
• D. Incorrect. Gemini for Google Workspace is a productivity and content assistant. It does not have the capability to perform e-commerce transactions or purchase domain names, which would require external payment authorization and specific domain registrar workflows.
• E. Incorrect. While Gemini can write or explain code snippets, Google Docs does not include an integrated execution environment to run Python code directly in the document margins. Such functionality is typically found in tools like Google Colab.
• F. Incorrect. Provisioning cloud infrastructure is a complex administrative task handled through the Google Cloud Console or Infrastructure as Code (IaC) tools. It is outside the scope of Gemini's document-based assistance features.

Correct answer: A — Ground the model's responses strictly in the provided medical articles.

• A. Correct. Grounding involves anchoring the model's responses in specific, verifiable source material (such as the medical articles). This practice, often implemented via Retrieval-Augmented Generation (RAG), reduces hallucinations by constraining the output to provided facts and allowing for citations, which is critical in domain-sensitive fields like healthcare.
• B. Incorrect. While model upgrades can improve general performance, they do not inherently prevent hallucinations. Proactive grounding is required to ensure a model remains faithful to a specific set of source documents.
• C. Incorrect. Vertex AI Feature Store is designed for managing and serving structured ML features for predictive models. It does not provide the grounding or retrieval mechanisms necessary to prevent factual inventions in generative text summarization.
• D. Incorrect. Security patches are vital for system integrity and are entirely unrelated to the factual accuracy or the generation logic of the foundation model. Lowering security measures would only introduce vulnerabilities.

Correct answer: A — Ground the model in the company's internal IT policy documents.

• A. Correct. Grounding the model in internal IT policy documents (often using Retrieval-Augmented Generation or RAG) ensures that the chatbot's responses are based on the provided context rather than general pre-trained knowledge. This technique effectively overcomes limitations related to data freshness and hallucinations by providing a specific source of truth for the model to reference at inference time.
• B. Incorrect. Fine-tuning on public IT forums would introduce external, non-official information into the model's weights. This would likely lead to the chatbot providing generic or potentially conflicting advice that deviates from the enterprise's specific internal policies.
• C. Incorrect. Automatic model upgrades update the underlying foundation model version (e.g., from Gemini 1.0 to 1.5). While this may improve reasoning capabilities, it does not restrict the model's knowledge base to internal company documents or prevent it from using general internet knowledge.
• D. Incorrect. Vertex AI Feature Store is designed for storing, serving, and sharing machine learning features for structured data (like user profiles or transaction history) for predictive modeling. It is not intended for grounding LLMs in unstructured documents, nor is it used for storing chatbot UI components.

Correct answer: B — First-party enterprise data

• A. World data refers to general knowledge about the world, such as public websites and encyclopedias. This data lacks the specific, internal information regarding a company's unique return policies and shipping procedures, leading to generic responses that do not reflect the organization's specific practices.
• B. First-party enterprise data consists of information generated and owned by the company itself, such as internal policy documents, knowledge bases, and procedure manuals. This is the most suitable grounding source because it provides the authoritative, up-to-date, and relevant information needed to accurately represent the company's unique internal rules.
• C. Third-party data is obtained from external sources or industry reports. While it may provide general guidance on retail norms, it will not contain the specific, proprietary details of the retailer's unique policies, which could lead to non-authoritative or incorrect answers.
• D. Synthetic data is artificially generated and is often used for model training or privacy-preserving tests. It is not an appropriate source for factual grounding of operational policies, as it does not replace the actual source documents and could introduce hallucinations or inaccuracies regarding legal and procedural rules.

Correct answer: B — Increase the temperature closer to 1.0

• A. Decreasing the temperature to 0.0 makes the model's output more deterministic and predictable by consistently choosing the most likely next token. This reduces variance and diversity, which is counterproductive when the goal is to generate unique and creative character names.
• B. Increasing the temperature closer to 1.0 (or higher, depending on the model) introduces more randomness into the token selection process. This allows the model to sample from a wider distribution of possible next tokens, including lower-probability ones, resulting in more varied, creative, and surprising outputs.
• C. Decreasing the maximum output tokens only limits the length of the generated text. It does not inherently increase the diversity or unpredictability of the content; that is controlled via sampling parameters like temperature or top-p/top-k.
• D. Increasing safety settings to the highest level filters out content that violates safety guidelines. While important for compliance, highly restrictive safety settings can sometimes suppress creative or unconventional outputs without actually increasing the diversity of the results.

Correct answer: C — Zero-shot prompting

• A. Incorrect. Few-shot prompting involves providing the model with a few examples (input-output pairs) to demonstrate the desired format and style. Since the manager provided no examples of names, this is not few-shot prompting.
• B. Incorrect. Chain-of-thought prompting encourages the model to produce intermediate reasoning steps or a line of thought before giving a final answer. This prompt simply requests names without instructions to show reasoning.
• C. Correct. Zero-shot prompting involves giving the model an instruction or task without providing any examples or demonstrations. The model must rely entirely on its pre-training and the direct instruction to produce the list of names.
• D. Incorrect. Role prompting involves assigning a specific persona or role (e.g., 'You are a professional branding expert') to the model to influence the tone and perspective of the response. No role was assigned in this prompt.

Correct answer: B — Prompt chaining

• A. Incorrect. Chain-of-thought prompting asks the model to produce intermediate reasoning steps (often using 'let's think step-by-step') to arrive at an answer within a single output. It focuses on the model's internal logic rather than a sequential workflow where the output of one prompt is manually or programmatically passed to another.
• B. Correct. Prompt chaining involves breaking a complex task into smaller, sequential sub-tasks where the output of one prompt (the outline) serves as the input or context for the next prompt (the introduction). This technique helps manage complexity and ensures higher quality by focusing the model on one specific task at a time.
• C. Incorrect. Few-shot prompting involves providing the model with a few examples (input-output pairs) within the prompt to demonstrate the desired format or behavior. This scenario describes a multi-step workflow, not the use of training examples within a prompt.
• D. Incorrect. Zero-shot prompting involves asking the model to perform a task without any prior examples or demonstrations in the prompt. While the initial request for an outline might be zero-shot, the overall process of using subsequent outputs defines prompt chaining.

Correct answers: A, B, D — It allows the team to review and refine the output at each intermediate step.; It reduces the complexity the LLM has to handle in a single generation.; It helps maintain focus and consistency across the different components of the campaign.

• A. Correct. Prompt chaining enables a 'human-in-the-loop' workflow where teams can validate and adjust the output of a specific sub-task (e.g., the audience profile) before it is used as the foundational context for the next step (e.g., ad copy).
• B. Correct. Breaking a complex multi-step task into smaller, discrete prompts reduces the reasoning load on the LLM. This minimization of complexity often leads to higher accuracy and better adherence to specific instructions for each component.
• C. Incorrect. Prompt chaining is a sequence of model interactions and does not inherently grant the LLM the ability to browse the internet; such capabilities require specific tool-calling or agentic integrations.
• D. Correct. By passing the output of one step as the context for the next, prompt chaining ensures that the landing page and ad copy remain grounded in the previously established audience profile, leading to higher thematic and strategic consistency.
• E. Incorrect. Prompt chaining still involves sending and receiving text, which consumes tokens. In many cases, chaining can actually increase total token usage because the context is passed repeatedly through multiple API calls.
• F. Incorrect. Chaining is a prompting technique used at inference time. It does not update the model's weights or perform fine-tuning, which is a separate training process requiring a dedicated dataset.

Correct answer: C — Text generation

• A. Image generation focuses on creating or altering visual content and graphics from textual descriptions; it is not designed for natural language processing tasks like summarization or email drafting.
• B. Code generation is specialized for producing source code and assisting with programming tasks based on specifications; it is not the primary tool for natural language summarization or language translation.
• C. Text generation models, such as Large Language Models (LLMs), are specifically designed to produce and transform natural language. They are the ideal choice for tasks that involve understanding and creating human-like text, including summarization, email composition, and translation.
• D. Speech-to-text generation is primarily used for transcription (converting spoken audio into written text). While it produces text as an output, it does not inherently perform the complex linguistic processing required for summarizing or translating that text.

Correct answer: C — Text generation and translation models

• A. Incorrect. While image generation can help create visuals and improve engagement, it focuses on visual content and does not address the core requirement of translating article text into multiple languages.
• B. Incorrect. Code generation for the website backend can accelerate development tasks and improve functionality, but it does not provide automated language translation or solve the bottleneck of slow manual content translation.
• C. Correct. Text generation and translation models are specifically designed to automate language translation and enable multilingual content creation at scale. These models can be integrated into publishing workflows, often with a human-in-the-loop for quality assurance, to significantly speed up the time-to-publish for international audiences.
• D. Incorrect. Personalized user avatars are a tool for user engagement and personalization. They do not contribute to translating article content or addressing the organization's need for faster multilingual publication.

Correct answer: B — Identity and Access Management (IAM)

• A. Incorrect. Security Command Center (SCC) is Google Cloud's security posture management and threat detection platform. While it helps identify vulnerabilities and security risks, it is not used to define or enforce granular user-level permissions for modifying specific model parameters.
• B. Correct. Identity and Access Management (IAM) is the standard service for granting fine-grained, resource-level permissions in Google Cloud. It allows administrators to follow the principle of least privilege by assigning specific roles to data scientists (e.g., Vertex AI Administrator or Editor) and different roles to customer service representatives (e.g., Vertex AI User or Discovery Engine Viewer).
• C. Incorrect. Cloud Monitoring is part of the Google Cloud Observability suite. It provides insights into performance, metrics, and uptime. It is used to monitor system behavior and alert on issues, not to manage or restrict user access to resource configurations.
• D. Incorrect. Virtual Private Cloud (VPC) manages network-level isolation, subnets, and firewalls. While VPC can restrict network access to a service, it cannot distinguish between different users' identities or define what specific actions (like editing model parameters) a user is permitted to perform once they are on the network.

Correct answer: B — Vulnerabilities can be introduced at any stage, such as poisoned training data or compromised model weights.

• A. Security tools and methodologies are not limited to data preparation; they are applicable and necessary across data collection, training, evaluation, and monitoring. Restricting security to a single phase leaves the system vulnerable to attacks at other stages.
• B. Vulnerabilities can be introduced at any point in the ML lifecycle, such as the injection of malicious data (data poisoning), tampering with model weights during storage, or exploiting insecure CI/CD pipelines. A defense-in-depth approach ensures the integrity and confidentiality of the model from development through to production.
• C. The primary objective of end-to-end ML security is to protect the confidentiality, integrity, and availability of models and data. While it is a best practice, its main driver is risk mitigation rather than the reduction of cloud storage costs.
• D. AI models require robust security even after deployment, as they are susceptible to adversarial inputs, model extraction, and data leakage. Security is not a one-time event at deployment but an ongoing requirement including monitoring and runtime protection.

Correct answer: B — Transparency

• A. Data anonymization is the process of protecting private or sensitive information by erasing or encrypting identifiers that connect an individual to stored data. While a critical privacy practice, it does not address the issue of user awareness regarding the identity of an AI agent.
• B. Transparency involves being open and clear about the use of AI, ensuring that users are clearly informed when they are interacting with an AI system rather than a human. Failing to disclose this identity leads to a loss of trust and a sense of deception, directly violating transparency principles.
• C. Pseudonymization is a data management and de-identification procedure that replaces private identifiers with pseudonyms to reduce re-identification risk. It is a technical privacy measure and is not related to the requirement of identifying an agent's artificial nature to users.
• D. Explainability refers to the ability to make the decisions and outputs of AI systems understandable to humans. While it helps with trust in model outcomes, the primary issue of failing to disclose that the agent is a chatbot is a lapse in transparency, not explainability.

Correct answer: B — Data pseudonymization

• A. Data anonymization involves irreversibly removing or modifying personally identifiable information (PII) so that individuals cannot be re-identified. Because the hospital retains a separate file that links the IDs back to the names, re-identification remains possible, meaning the data is not fully anonymized.
• B. Data pseudonymization is the technique of replacing direct identifiers (like names) with pseudonyms or IDs while maintaining a separate, secure key that allows for re-identification when necessary. This practice reduces the risk of identity exposure during processing while preserving the link to the original data, which is common in privacy frameworks like GDPR.
• C. Data deletion involves the permanent removal of information from a dataset so it can no longer be retrieved or used. In this scenario, the hospital is not removing the data but rather restructuring how identifiers are stored and linked.
• D. Data augmentation is a machine learning technique used to increase the size or diversity of a training dataset by creating synthetic data or modified versions of existing data. It is not a privacy-preserving identification management technique.

Correct answer: A — Explainability

• A. Correct. Explainability (or interpretability) is the ability to understand the reasoning or logic behind an AI's output. In high-stakes fields like healthcare, clinicians require localized evidence (such as heatmaps or feature highlights) to validate the AI's findings. Without this transparency, the system is seen as a 'black box,' leading to a lack of trust and clinical utility.
• B. Incorrect. Anonymization is the process of removing personally identifiable information (PII) from data to protect patient privacy. While critical for regulatory compliance, it does not address the model's transparency or its failure to explain its diagnostic conclusions.
• C. Incorrect. Pseudonymization is a data privacy technique where personal identifiers are replaced with artificial identifiers (pseudonyms). This is a data management control and is unrelated to the AI model's decision-making process or visual output for clinicians.
• D. Incorrect. Data quality refers to the accuracy, completeness, and representativeness of the training or input data. While a model requires high-quality data to be accurate, the specific issue here is not the accuracy of the result, but the inability to see the underlying rationale for the result.