Generative AI Leader Practice Test 45 — 15 Questions

Question 1

A development team is integrating a large language model into a healthcare application. They need to reduce the risk of generating harmful medical advice. Which THREE measures should they implement? (Choose three.)

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Use a safety filter to block outputs containing harmful medical terminology.. Implementing a safety filter that blocks outputs containing harmful medical terminology directly mitigates the risk of generating dangerous advice. This acts as a post-processing guardrail, intercepting model outputs that include terms associated with diagnoses, dosages, or procedures that could lead to patient harm. It is a standard practice in high-stakes domains to layer such filters on top of the generative model.

Answer

Include a disclaimer in the system instruction that the model is not a doctor.

Answer

Set the temperature to a very high value to ensure diverse outputs.

Question 2

A company is prompt engineering a model for customer support. They want to reduce hallucination (false information) in responses. Which TWO techniques are most effective? (Choose two.)

Accepted Answer

Implement RAG to retrieve relevant documents for context. Retrieval-Augmented Generation (RAG) grounds the model's output in external, verifiable documents retrieved from a knowledge base. By providing relevant context at inference time, RAG significantly reduces the likelihood of the model fabricating information, as it can reference and paraphrase from the retrieved sources rather than relying solely on its parametric memory.

Answer

Provide 3 few-shot examples of conversations

Answer

Reduce max output tokens to 150

Answer

Increase temperature to 1.2

Question 3

A company needs to fine-tune a foundation model on Vertex AI for a custom text classification task with only 500 labeled examples. They want to minimize cost while achieving high accuracy. What is the MOST cost-effective approach?

Accepted Answer

Design prompts with few-shot examples and test it with the available data.. The most cost-effective because it leverages prompt engineering with few-shot examples, which requires no training or infrastructure costs. With only 500 labeled examples, a well-designed prompt can often achieve high accuracy for custom text classification without the expense of fine-tuning or model distillation on Vertex AI.

Answer

Fine-tune the foundation model using full fine-tuning on the entire dataset.

Answer

Use model distillation to train a smaller student model.

Answer

Use Vertex AI LLM-based evaluation to compare multiple large models and select the best one.

Question 4

A healthcare organization needs a generative AI model to answer medical questions using proprietary clinical guidelines. They have a large dataset of doctor-patient interactions. Should they fine-tune a pre-trained model or use Retrieval-Augmented Generation (RAG)?

Accepted Answer

Use RAG to retrieve relevant guidelines during inference, avoiding frequent retraining.. RAG is preferred because it retrieves the most current clinical guidelines from an external knowledge base during inference, avoiding the need for constant retraining when guidelines change. This is especially important in healthcare where regulations update frequently. Fine-tuning a pre-trained model on proprietary interactions might lead to overfitting or outdated knowledge. Option A is incorrect because RAG does not necessarily reduce inference costs; it adds overhead for retrieval. Option C is incorrect because prompt engineering cannot encode all proprietary guidelines; it is limited by context window size. Option D is incorrect because fine-tuning requires retraining when guidelines change, which is less flexible than RAG.

Answer

Use RAG to reduce inference costs by skipping model updates.

Answer

Use prompt engineering to encode all guidelines into the system prompt.

Answer

Fine-tune the model on the clinical guidelines and interactions.

Question 5

A project manager wants to track the ROI of a generative AI feature that assists customer support agents. Which metric is MOST directly tied to productivity improvement?

Accepted Answer

Average handle time (AHT) per ticket. Average handle time (AHT) directly measures the time agents spend per interaction, so a reduction indicates productivity gain. CSAT measures satisfaction, not efficiency. Cost per API call is a cost metric. Adoption rate measures usage.

Answer

Adoption rate of the AI tool

Answer

Customer satisfaction (CSAT) score

Answer

Cost per API call

Question 6

A team set a budget alert for their GenAI API usage at $10,000. They received the alert with current spend of $12,500. Which business action is most appropriate as a first step?

Accepted Answer

Review usage patterns and optimize prompt lengths and frequencies. The first step in responding to a budget overrun should be to analyze usage patterns and optimize prompt lengths and frequencies. This approach identifies inefficiencies (e.g., unnecessarily verbose prompts, excessive retries) that directly reduce token consumption and cost without disrupting critical operations. It aligns with the principle of cost optimization before making architectural or policy changes.

Answer

Pause all non-critical use cases immediately

Answer

Switch to a cheaper model provider

Answer

Increase the budget by 50% to $15,000

Question 7

You are a data scientist at a financial institution. You are using Vertex AI to fine-tune a large language model (LLM) for generating financial reports. You have prepared a dataset of 10,000 examples. During fine-tuning, you notice that the training loss is decreasing steadily, but the validation loss is increasing after 5 epochs. The model's generated reports on the validation set contain many factual errors and nonsensical statements. You suspect overfitting. You have limited compute budget and need to improve generalization. What should you do?

Accepted Answer

Implement early stopping with a patience of 2 epochs. Early stopping with a patience of 2 epochs is the correct approach because it directly addresses overfitting by halting training when the validation loss fails to improve for a specified number of epochs. This preserves the model's generalization ability without requiring additional compute or data, which aligns with the limited budget constraint. In Vertex AI, early stopping is a built-in hyperparameter tuning strategy that monitors validation metrics and stops the job to prevent further degradation.

Answer

Increase the learning rate

Answer

Increase the number of training epochs to 20

Answer

Add more training examples from a public dataset

Question 8

A developer is using the Vertex AI PaLM API to generate code. They want to ensure the output is safe and adheres to company policies. Which THREE attributes can they configure in the safety_settings parameter?

Accepted Answer

Toxicity. The safety_settings parameter in the Vertex AI PaLM API accepts a list of safety categories, each with a configurable threshold, and the supported categories include Toxicity (C), Harassment (D), and Sexually explicit content (E), so these three are the attributes the developer can configure to filter unsafe output and enforce company policies. Toxicity (C) lets them block content that is rude, disrespectful, or otherwise harmful, Harassment (D) targets content that bullies or intimidates individuals or groups, and Sexually explicit content (E) filters sexually explicit material; each is a distinct safety category with its own threshold setting. Language detection (A) is not a safety category but a separate text-analysis capability, and sentiment analysis (B) is likewise an analytical feature rather than a configurable safety attribute, so neither belongs in safety_settings.

Answer

Language detection

Answer

Sentiment analysis

Question 9

A company is deploying a Gemini-based application and needs to ensure low latency for real-time user interactions. They also want to reduce cost. Which THREE strategies should they consider? (Select 3)

Accepted Answer

Use Gemini 1.5 Flash instead of Pro. Option A is correct because Gemini 1.5 Flash is a lighter, faster model than Gemini 1.5 Pro, delivering lower latency for real-time interactions and costing less per token, which directly addresses both the latency and cost goals. Option B is correct because caching responses to frequently repeated queries avoids redundant model invocations, cutting both latency (cache hits return instantly) and cost (fewer billed tokens). Option E is correct because transformer inference cost and latency scale with input length, so trimming the context window to only the necessary tokens reduces processing time and token charges. Option C is not appropriate because increasing max output tokens generates longer responses, raising latency and cost rather than reducing them. Option D is not appropriate because full fine-tuning is expensive and does not inherently make inference faster; latency gains typically come from smaller/distilled models or optimized serving, not full fine-tuning.

Answer

Increase the model's max output tokens to ensure comprehensive answers

Answer

Use full fine-tuning to make the model faster

Question 10

A company is deploying a generative AI model for medical diagnosis assistance. To comply with both Google's AI Principles and emerging regulations (e.g., EU AI Act), they must ensure appropriate human oversight. Which THREE measures should they implement?

Accepted Answer

Require a human clinician to review all AI-generated diagnoses before acting on them. Option C is correct because requiring a human clinician to review every AI-generated diagnosis before it is acted upon directly implements human-in-the-loop oversight, which Google's AI Principles and the EU AI Act demand for high-risk medical applications. Option D is correct because an override mechanism ensures the human reviewer retains meaningful authority to reject or correct the AI's recommendation, preventing automation bias and preserving human control. Option E is correct because documenting the human-in-the-loop process, including who reviews what and at which stage, provides the accountability and traceability that regulators and AI governance frameworks require. Option A is not appropriate because high confidence scores do not eliminate the need for human oversight in a high-risk medical context, and using them to bypass review undermines the required human control. Option B is also not appropriate because allowing autonomous action even for low-risk cases conflicts with the principle of maintaining human oversight for medical diagnosis assistance under the stated compliance requirements.

Answer

Use a model with high confidence scores to bypass human review

Answer

Allow the AI to act autonomously for low-risk cases to reduce workload

Question 11

A data scientist needs to fine-tune a foundation model for a sentiment analysis task without managing infrastructure. Which Google Cloud service should they use?

Accepted Answer

Vertex AI Model Garden. Vertex AI Model Garden is the correct service because it provides a curated hub of foundation models that can be fine-tuned with managed infrastructure, eliminating the need for the data scientist to provision or manage servers. It supports one-click deployment and fine-tuning workflows for sentiment analysis, directly addressing the requirement to avoid infrastructure management.

Answer

Compute Engine

Answer

BigQuery ML

Answer

Cloud Run

Question 12

A company wants to deploy a generative AI application that must comply with both Google's AI Principles and the EU AI Act. Which THREE practices should they adopt?

Accepted Answer

Use data minimization and purpose limitation to respect privacy. Option B is correct because data minimization and purpose limitation are core privacy principles aligned with both Google's AI Principles (which emphasize privacy by design) and the EU AI Act's data governance requirements for high-risk AI systems. Option C is correct because the EU AI Act mandates human oversight for high-risk AI systems, and Google's AI Principles require meaningful human control over consequential decisions. Option D is correct because deploying safety filters to block harmful content directly supports Google's AI Principle of avoiding creating or reinforcing unfair bias and harm, and aligns with the EU AI Act's requirements for accuracy, robustness, and cybersecurity. Option A is incorrect because using the largest possible model does not inherently ensure compliance with either framework and may increase energy use and risk without addressing governance obligations. Option E is incorrect because disabling logging undermines transparency, traceability, and accountability, which are explicitly required by the EU AI Act and expected under Google's responsible AI practices.

Answer

Use the largest possible model for best accuracy

Answer

Disable logging to minimize data retention

Question 13

A developer wants to integrate Gemini multimodal capabilities (text + image) into a mobile app using Python. Which Google Cloud client library should they use?

Accepted Answer

Vertex AI client library (google-cloud-aiplatform). The Vertex AI client library (google-cloud-aiplatform) provides the Generative AI SDK that supports multimodal capabilities, including the ability to send both text and image inputs to Gemini models. This library directly exposes the `GenerativeModel` class with methods like `generate_content()` that accept `Part` objects containing image data (e.g., `Part.from_image()` or `Part.from_uri()`), making it the correct choice for integrating Gemini multimodal features into a Python mobile app backend.

Answer

Dialogflow CX

Answer

Cloud Vision API

Answer

Natural Language API

Question 14

A developer uses the Vertex AI Python SDK to call a Gemini model for structured JSON output. However, the model often returns malformed JSON. Which parameter should the developer set in the generation configuration to enforce valid JSON output?

Accepted Answer

Set the 'response_mime_type' parameter to 'application/json'.. Setting `response_mime_type` to `'application/json'` in the generation configuration instructs the Gemini API to constrain the model's output to valid JSON format. This parameter leverages the model's native structured output capability, ensuring the response adheres to JSON syntax without relying on post-processing or prompt engineering.

Answer

Set the temperature to a lower value (0.1) to reduce variation.

Answer

Include few-shot examples of the desired JSON format in the system prompt.

Answer

Switch to a smaller model to reduce complexity.

Question 15

A company is developing a generative AI application that will be used by customers in the EU. To comply with the GDPR and the upcoming EU AI Act, which THREE measures should they implement? (Select 3)

Accepted Answer

Store and process all user data within the EU or in regions with adequacy decisions. Option A is correct because GDPR data-transfer rules (Chapter V) require personal data to remain in the EU/EEA or be transferred only to countries with an adequacy decision from the European Commission, so localizing storage and processing is a valid compliance measure. Option D is correct because GDPR Articles 13-15 grant data subjects the right to meaningful information about and explanation of automated decision-making, and the EU AI Act reinforces transparency and explainability obligations for AI systems. Option E is correct because the EU AI Act requires transparency so that users are informed when they are interacting with an AI system rather than a human, and GDPR fairness/transparency principles support this disclosure. Option B is not required: open-source models are neither mandated by GDPR nor by the AI Act, and transparency obligations apply regardless of model licensing. Option C is not required: GDPR does not impose a fixed 30-day log deletion period; retention must be limited to what is necessary, but the specific 30-day value is arbitrary and not a legal requirement.

Answer

Use only open-source models to ensure transparency

Answer

Automatically delete all user logs after 30 days