Generative AI Leader Practice Test 36 — 15 Questions

Question 1

A company is deploying a GenAI application that must meet SOC 2 compliance. Which three Google Cloud offerings can be used in a compliant manner? (Choose three.)

Accepted Answer

Vertex AI. Vertex AI (C) is a fully managed, enterprise-grade Google Cloud service covered by SOC 2 attestation, with data residency, CMEK, VPC Service Controls, and IAM controls that let a GenAI workload run compliantly. Document AI (D) is likewise a Google Cloud enterprise service included in Google's SOC 2 report scope, so processing documents through its processors satisfies the compliance requirement. BigQuery ML (E) runs inside BigQuery, which is a SOC 2-attested Google Cloud service, allowing model training and inference on governed data with Cloud IAM, audit logging, and CMEK. Colab consumer (A) and the free tier of Google AI Studio (B) are consumer-oriented offerings whose terms and controls do not provide the SOC 2 enterprise assurances required for this deployment.

Answer

Colab (consumer version)

Answer

Google AI Studio (free tier)

Question 2

A company notices that their GenAI-powered code review assistant sometimes suggests code with security vulnerabilities. They want to reduce this risk without sacrificing review speed. Which approach is MOST effective?

Accepted Answer

Include the company's secure coding guidelines in the system prompt and provide few-shot examples of secure code. It directly addresses the root cause of insecure code generation by providing the model with explicit secure coding guidelines and few-shot examples. This technique, known as in-context learning, steers the model's output toward desired behaviors without altering its architecture or speed. It effectively reduces security vulnerabilities while maintaining the assistant's rapid review capability.

Answer

Disable the code review assistant for security-sensitive modules

Answer

Manually review every suggestion from the assistant before accepting it

Answer

Switch to a smaller, less capable model that is less likely to generate complex code

Question 3

A healthcare company needs to process medical records at scale, extracting structured information from forms and enabling natural language queries. They require HIPAA compliance and want to avoid training custom models. Which TWO services should they use? (Choose 2)

Accepted Answer

Natural Language AI. Document AI (option C) is correct because it provides pre-trained, HIPAA-eligible processors for form parsing and OCR, extracting structured information from medical forms at scale without custom model training. Natural Language AI (option A) is also correct because its pre-trained models perform entity extraction, sentiment analysis, and syntax analysis on unstructured clinical text, supporting natural language understanding of medical records without training custom models. Both can be used under HIPAA compliance when configured appropriately.

Answer

BigQuery ML

Answer

Vertex AI Workbench

Answer

Vision AI

Question 4

Which TWO safety features are available in Vertex AI Gemini API? (Select TWO.)

Accepted Answer

Safety filters for categories like hate speech and harassment. Option A is correct because the Vertex AI Gemini API applies configurable safety filters that screen both prompts and responses against harm categories such as hate speech, harassment, sexually explicit content, and dangerous content. Option B is correct because these safety filters operate using configurable thresholds (for example BLOCK_LOW_AND_ABOVE, BLOCK_MEDIUM_AND_ABOVE, BLOCK_ONLY_HIGH, BLOCK_NONE) that let developers tune how aggressively content is blocked per category. Together, A and B describe the built-in safety attributes exposed through the API's safetySettings. Option C is not a Gemini API safety feature but a general Google Cloud storage/platform control (encryption at rest is handled by the underlying infrastructure, not the API's safety settings). Option D is incorrect because the Gemini API does not automatically redact PII as a built-in safety feature. Option E is incorrect because Cloud DLP is a separate Google Cloud service that must be integrated manually and is not a native safety feature of the Gemini API.

Answer

Model-level encryption at rest

Answer

Automatic redaction of personally identifiable information (PII)

Answer

Integration with Cloud Data Loss Prevention (DLP)

Question 5

A global news agency is using a generative AI model to summarize breaking news articles in real-time. The model is deployed on Vertex AI across multiple regions (us-central1, europe-west4, asia-southeast1) for low latency worldwide. The agency has a Service Level Objective (SLO) of 99.9% availability and p99 latency under 2 seconds. Recently, during a major event, traffic spiked 10x, and the europe-west4 region experienced latency spikes over 5 seconds and some 503 errors. The team suspects the regional endpoint is under-provisioned. Which combination of actions should they take to meet the SLO consistently?

Accepted Answer

Enable the global endpoint feature in Vertex AI with automatic traffic splitting, and increase the minimum replicas for each regional endpoint. It enables the global endpoint feature with automatic traffic splitting, allowing traffic to be routed to healthy regions and providing failover. Additionally, increasing minimum replicas per region ensures each regional endpoint has baseline capacity to handle spikes, preventing under-provisioning. Option B only increases max replicas in europe-west4, which does not address traffic shifts, and reducing min replicas elsewhere risks capacity issues. Option C suggests Cloud CDN, which is for static content, not model inference. Option D configures a global load balancer with a single endpoint, which does not optimally use Vertex AI's regional endpoints and may not meet latency SLO.

Answer

Increase the maximum replicas for the europe-west4 endpoint and reduce the min replicas in other regions

Answer

Implement Cloud CDN caching for common summaries and reduce the number of regions to two

Answer

Configure a global load balancer with a single Vertex AI endpoint and increase max replicas globally

Question 6

A company wants to deploy a generative AI chatbot for customer service but is concerned about cost unpredictability due to variable usage. Which pricing model should they choose to best manage costs?

Accepted Answer

Provisioned throughput. Provisioned throughput provides fixed capacity with predictable monthly cost, ideal for managing cost uncertainty. Option A (committed use discounts) requires a commitment but costs can still vary if usage exceeds the committed amount. Option B (free tier) is too limited for a full-scale chatbot. Option C (pay-as-you-go) is variable and leads to cost unpredictability.

Answer

Committed use discounts

Answer

Free tier

Answer

Pay-as-you-go

Question 7

A company wants to generate high-quality product images from text descriptions for an e-commerce catalog. They need photorealistic results. Which model and approach should they choose?

Accepted Answer

Use Imagen on Vertex AI with appropriate prompts. Imagen on Vertex AI is specifically designed for high-quality, photorealistic text-to-image generation, making it the ideal choice for creating product images from text descriptions. It leverages advanced diffusion models to produce detailed and visually accurate outputs that meet the requirements of an e-commerce catalog.

Answer

Use Veo for video generation and extract frames

Answer

Fine-tune Gemini 1.5 Pro on product images

Answer

Use Codey to generate code that renders images

Question 8

A media company is using Vertex AI's Imagen model to generate images for marketing campaigns. They have a set of prompts that describe desired scenes, but the generated images often contain artifacts such as distorted faces or unnatural lighting. The team has tried varying the prompt wording but the issues persist. They are using the default parameters (no modifications). They have a budget for additional compute resources and want to improve image quality without switching to a more expensive model. The team has access to a small set of high-quality images in the same style as their target outputs. What should the team do?

Accepted Answer

Fine-tune the Imagen model on the small set of high-quality images to improve output quality.. Fine-tuning the Imagen model on the small set of high-quality images allows the model to learn the desired style and reduce artifacts like distorted faces and unnatural lighting, improving output quality without switching to a more expensive model. Option A is incorrect because increasing the guidance scale may cause the model to overfit to the prompt and potentially introduce more artifacts rather than fix them, and the team already has issues with prompt adherence. Option B is incorrect because using more detailed prompts with negative prompts might help but the team already tried varying wording without success; the root cause is the model's lack of specific training on the desired quality, which fine-tuning directly addresses. Option D is incorrect because generating more images per prompt does not improve the per-image quality; it only increases the chance of finding a good one, and the team wants to improve overall image quality.

Answer

Increase the guidance scale parameter to make the model follow prompts more closely.

Answer

Use a more detailed prompt style with negative prompts to avoid artifacts.

Answer

Increase the number of images generated per prompt and manually select the best ones.

Question 9

An e-commerce company fine-tunes a model on customer reviews to generate product feedback summaries. They want to ensure the model does not reproduce toxic language from the training data. Besides filtering the training data, which additional technique is most effective at inference time?

Accepted Answer

Pass the model output through a toxicity detection model and conditionally regenerate or block. It directly addresses the safety requirement at inference time by introducing a secondary guardrail. A toxicity detection model (e.g., a classifier trained on the Jigsaw Toxic Comment dataset) can score the generated output in real time; if the score exceeds a threshold, the system can either block the response or trigger a regeneration with adjusted parameters. This is the only technique that actively filters for toxic language after generation, rather than merely reducing output variance or exploring alternative sequences.

Answer

Set temperature to 0.0 to reduce variance

Answer

Set top-k to 10 to limit token choices

Answer

Use beam search with a high beam width

Question 10

A financial technology company has deployed a custom-tuned PaLM 2 model on Vertex AI to generate personalized investment recommendations for retail clients. The model was fine-tuned on a corpus of historical market data and advisory transcripts. Recently, the compliance team flagged that several recommendations contradicted SEC guidelines, and the model sometimes repeated prohibited statements from outdated training materials. The team has already implemented safety filters (e.g., blocking toxic content) and adjusted the model's system instructions to be more conservative. However, the issues persist. The model's deployment parameters are: temperature=0.4, top_p=0.9, max_output_tokens=500, and no grounding. The company must maintain compliance without significantly increasing latency. What should they do next?

Accepted Answer

Configure Vertex AI grounding using a curated data store of real-time SEC regulations and market data. Configuring Vertex AI grounding with a curated data store of real-time SEC regulations directly addresses the root cause: the model is generating outputs that contradict current compliance rules. Grounding forces the model to base its responses on authoritative, up-to-date sources, which is more effective than safety filters or system instructions alone, and it avoids the latency increase of a second model or the risk of catastrophic forgetting from additional fine-tuning.

Answer

Increase temperature to 0.7 to allow more diverse responses, and add a second model to verify outputs

Answer

Perform an additional fine-tuning round exclusively on the most recent SEC regulatory filings and compliance-approved content

Answer

Implement a chain-of-thought prompting technique that requires the model to explain its reasoning step by step

Question 11

A company is using generative AI for code generation and wants to evaluate the quality of generated code for security vulnerabilities. Which metric is most appropriate?

Accepted Answer

Automatic static analysis. (Automatic static analysis) is correct because it directly scans code for security vulnerabilities, making it the most appropriate metric for this purpose. Option A (BLEU score) measures text similarity, not security. Option C (Human evaluation) is subjective and less scalable. Option D (Perplexity) measures language model confidence, not code security.

Answer

BLEU score

Answer

Human evaluation

Answer

Perplexity

Question 12

A retail company wants to use gen AI for customer service chatbots. They have a large volume of customer interactions. What is the primary business consideration for deploying a gen AI solution?

Accepted Answer

Ensuring data privacy and compliance. Ensuring data privacy and compliance is the primary business consideration when handling customer data, especially with regulatory requirements like GDPR or CCPA. Option A is wrong because while latency matters, minimizing it at any cost is not the primary consideration and can lead to excessive expenses. Option B is wrong because using only open-source models may limit flexibility and can raise data privacy concerns if not properly vetted. Option C is wrong because choosing the most complex model often increases cost and latency without necessarily improving business outcomes.

Answer

Minimizing latency at any cost

Answer

Using open-source models only

Answer

Choosing the most complex model

Question 13

A media company wants to add generative AI features to its mobile app but must control costs and prevent unexpected spend as usage grows. Leadership wants visibility into consumption by product team. Which governance approach should the GenAI Leader recommend on Google Cloud?

Accepted Answer

Issue each product team its own Google Cloud project and apply quotas and budget alerts on the Vertex AI usage in those projects.. Project-level separation combined with quotas and budget alerts gives each product team an attributable cost boundary and a hard ceiling on consumption. The other approaches either report spend only after it happens, blur attribution by sharing identity, or block the capability entirely instead of governing it.

Answer

Rely on the monthly consolidated billing report and review Vertex AI charges with finance after each invoice arrives.

Answer

Apply a single organization-wide quota for Vertex AI and share one service account across all product teams.

Answer

Turn off Vertex AI API access for the organization and require teams to request manual approval for each generative feature.

Question 14

A company is deploying a generative AI system for medical diagnosis. Which TWO measures are essential for responsible AI in this high-stakes domain?

Accepted Answer

Ensure a human medical professional reviews all AI-generated diagnoses. Option B is correct because in a high-stakes medical diagnosis scenario, keeping a qualified human medical professional in the loop for every AI-generated diagnosis ensures accountability and mitigates the risk of harmful errors, which is a core principle of responsible AI (human oversight). Option D is correct because Model Cards provide structured documentation of a model's intended use, performance metrics, and known limitations, enabling clinicians to understand when and how much to trust the system's outputs. Together, these measures support transparency and human accountability, which are essential in clinical decision-making. Option A is not appropriate because autonomous AI decisions in emergencies remove the necessary human oversight and could cause serious harm. Option C is wrong because publishing all patient data would violate privacy and confidentiality regulations such as HIPAA, and transparency does not require exposing raw patient data. Option E is incorrect because restricting AI to administrative tasks does not address responsible AI for diagnosis and is not an essential measure for the described use case.

Answer

Allow the AI to make autonomous decisions in time-sensitive emergencies

Answer

Publish all patient data used for training to ensure transparency

Answer

Use the AI only for administrative tasks, not diagnosis

Question 15

A model generates biased output. Which technique is least effective?

Accepted Answer

Set frequency penalty to 1.0. Setting the frequency penalty to 1.0 is least effective for reducing biased output because frequency penalties reduce repetition of tokens based on their frequency in the generated text, not their association with protected attributes or fairness. This parameter controls lexical diversity, not demographic parity or representational harm, so it does not address the root cause of bias in the model's training data or inference logic.

Answer

Use adversarial debiasing

Answer

Apply safety filters

Answer

Fine-tune on diverse data