Google Cloud GenAI-Leader Practice Test

GenAI-Leader practice test: what's covered

• Gemini
• Vertex AI
• Prompting
• RAG
• Agents
• Responsible AI

Free GenAI-Leader sample questions

1. A product leader is briefing executives and must explain what makes a large language model 'generative' rather than a traditional rules-based system. Which characteristic best captures this distinction?

   • A. It produces novel content by predicting likely sequences learned from large training corpora
   • B. It only retrieves exact answers from a fixed relational database of curated facts
   • C. It executes hand-written if-then branches that a developer defined for every input
   • D. It compresses files so they occupy less storage on disk without any modeling
   Show answer

   Correct: A. Generative models create new content (text, images, code) by statistically predicting the next most probable token based on patterns learned during training. Retrieving exact answers from a database is information retrieval, not generation. Hand-written if-then branches describe rules-based systems, the opposite of a learned model. File compression is unrelated to generative modeling, so those options are wrong.

2. An engineering team is debugging why their prompt is being truncated before the model finishes a long document summary. They learn the model splits text into smaller units before processing. What are these units called?

   • A. Tokens, the chunks of text a model reads and generates within its context window
   • B. Pixels, the smallest addressable color elements that compose a raster image file
   • C. Gradients, the numeric slopes used only during backpropagation in training
   • D. Checkpoints, the periodic saved snapshots of model weights during fine-tuning
   Show answer

   Correct: A. Models process text as tokens (sub-word units), and context windows are measured in tokens, so exceeding the limit causes truncation. Pixels are image elements, not text units, so that is wrong. Gradients are training math, not text chunks. Checkpoints are saved weights, not the units the model reads, making those options incorrect.

3. A data scientist explains embeddings to a business stakeholder who wants to build semantic search. Which statements accurately describe what embeddings are and why they matter? (Choose TWO)

   • A. Embeddings are numeric vectors that capture the semantic meaning of text or images
   • B. Items with similar meaning end up close together in the embedding vector space
   • C. Embeddings are plain-text passwords that authenticate users to the search index
   • D. Embeddings increase a model's parameter count every time a query is submitted
   • E. Embeddings permanently delete training data to comply with retention policies
   Show answer

   Correct: A and B. Embeddings convert content into dense numeric vectors where semantic similarity maps to spatial proximity, which is exactly what powers semantic search and recommendations. They are not passwords or an authentication mechanism, so that option is wrong. Submitting a query does not change parameter count, and embeddings do not delete training data, making those choices incorrect.

4. During a pilot, a chatbot confidently states a refund policy that the company has never had, presenting the fabricated detail as fact. What is the industry term for this failure mode of generative models?

   • A. Hallucination, where the model generates plausible-sounding but factually false output
   • B. Overfitting, where a model memorizes its training set and fails on new data
   • C. Latency, the measurable delay between a request and the first returned token
   • D. Quantization, reducing the numeric precision of weights to shrink a model
   Show answer

   Correct: A. A hallucination is when a model produces confident, fluent output that is not grounded in reality or its sources, exactly as described. Overfitting is a training problem about poor generalization, not fabricated facts at inference. Latency is a speed metric, and quantization is a compression technique, so neither describes inventing false information.

5. A leader wants the simplest correct distinction between a foundation model and a narrow, task-specific machine learning model when planning the company's AI investment. Which statement is most accurate?

   • A. A foundation model is pre-trained broadly and can be adapted to many downstream tasks
   • B. A foundation model can only ever classify spam and nothing else by design
   • C. A foundation model requires no data and is programmed entirely by hand each time
   • D. A foundation model runs exclusively on a single laptop with no accelerators at all
   Show answer

   Correct: A. Foundation models are trained on broad data at scale and then adapted through prompting, tuning, or grounding to many tasks, which is their defining value. Being limited to spam classification describes a narrow model, the opposite. Foundation models are trained on massive data, not hand-programmed, and large models typically need accelerators, so the remaining options are wrong.

Last updated: 2026-10-19

GenAI-Leader practice test FAQ

Related practice tests

• Google Cloud ACE practice test
• Google Cloud Professional Cloud Architect practice test
• Google Professional Cloud DevOps Engineer practice test
• Google Professional Cloud Network Engineer practice test
• Google Cloud PCSE practice test
• Google Cloud Professional Data Engineer practice test