12 Google Generative AI Leader (GAIL) Practice Questions with Explanations

Explanation

Unsupervised learning finds structure in unlabeled data, and clustering is the textbook way to discover groupings nobody has defined yet. Supervised learning (A) needs labeled examples of the categories, which the retailer does not have and is trying to discover. Reinforcement learning (C) learns actions from a reward signal over time; there is no environment or reward here. Prompting a text model (D) ignores the photos entirely and does not discover visual similarity.

Explanation

Gemma is Google's family of lightweight, open models that organisations can download, adapt and run on their own infrastructure, which is exactly what the bank describes. Imagen (A) generates images and Veo (B) generates video; neither fits a general-purpose model the bank hosts itself. Gemini (C) is Google's most capable multimodal model family, but it is consumed as a managed service rather than downloaded and run privately, so it misses the "run inside our own infrastructure" requirement.

Explanation

Unstructured data has no predefined schema of rows and columns; audio recordings, images, video, emails and free-text documents are typical examples, and they are where gen AI often adds the most value. Options B, C and D are structured data: each has a fixed schema with defined fields, which is why they fit naturally in tables, spreadsheets and relational databases.

Explanation

Gemini Enterprise is Google's prebuilt offering for employees that brings search across company data sources together with agents that can act on that information. Gemini for Google Workspace (A) assists people inside individual Workspace apps; it is not designed as a central search-and-agent hub across third-party systems like SharePoint. Google AI Studio (B) is a developer prototyping tool, not an employee-facing product. The Customer Engagement Suite (D) serves external customers through contact-center tools, not internal employee knowledge work.

Explanation

Agent Assist supports human agents during live interactions with real-time suggestions, transcription and knowledge surfacing. Conversational Insights (B) is in the same suite but analyses conversations to find trends and sentiment; it is an analytics tool rather than live in-call help. Imagen (C) generates images and does not help agents mid-call. Model Garden (D) is where builders choose models on Agent Platform; it is a component you might build with, not a finished capability for contact-center staff.

Explanation

Model Garden is the catalog on Agent Platform (Google's platform formerly branded Vertex AI) for discovering, testing and deploying Google, third-party and open models, which matches the "single catalog" requirement. Gems (A) are customised versions of Gemini for repeated personal tasks, not a model catalog. Agent Assist (B) is a contact-center tool. Security Command Center (C) is a security and risk management service; it does not provide a model selection catalog.

Explanation

Google AI Studio is the quick, browser-based environment for developers to try Gemini models, iterate on prompts and generate an API key for prototyping. Conversational Insights (A) analyses contact-center conversations, not developer prompts. Gemini for Google Workspace (C) helps users inside Workspace apps but does not give a developer API access for an app. AutoML (D) trains custom models from labeled data, which is slower and unnecessary when the goal is to try an existing foundation model today.

Explanation

Including a handful of input-output examples in the prompt so the model copies their pattern is few-shot prompting, and it is one of the cheapest ways to enforce a consistent format. Zero-shot (A) means no examples at all. Chain-of-thought (B) asks the model to reason step by step, which helps with multi-step logic rather than output formatting. Fine-tuning (D) changes the model's weights using a training dataset; pasting examples into a prompt does not modify the model.

Explanation

Temperature controls randomness in token selection; lowering it makes the model favour its most likely tokens, so repeated runs become more consistent. Raising top-p (B) widens the pool of candidate tokens and tends to increase variety, the opposite of the goal. Raising temperature (C) also increases variation. Output length (D) caps how long a response can be but does nothing to make two responses more alike.

Explanation

The problem is the knowledge cutoff, and Grounding with Google Search lets the model base its answer on current, public web results and return sources, which addresses both freshness and citations. Monthly fine-tuning (A) is costly and would still lag behind a weekly news cycle. Few-shot examples (B) shape format but add no new knowledge. Editor approval (C) is human-in-the-loop review: valuable for risk, but it does not give the model access to recent facts and creates a bottleneck at scale.

Explanation

Pseudonymization replaces direct identifiers with tokens, and the mapping key is held separately, so the model team cannot identify patients while the compliance office can re-link records when authorised. Anonymization (A) is irreversible, which breaks the re-linking requirement. Encryption in transit (C) protects data while it moves, but the team would still see identifiable records once received. An NDA (D) is a legal control, not a technical privacy safeguard, and exposes identities to the team.

Explanation

Gen AI initiatives should be measured against business outcomes defined up front, so comparing KPIs like resolution time and customer satisfaction with a pre-pilot baseline shows whether the assistant created value. Prompt volume (A) and license counts (D) measure activity and adoption, not impact; heavy use of an unhelpful tool would still look good. Model parameter count (B) is a technical attribute that says nothing about whether this pilot improved the business.