hardMultiple Select
Generative AI Leader Practice Question: A financial institution wants to deploy a…
A financial institution wants to deploy a generative AI system for automated report generation. They require that the model does NOT expose sensitive information from its training data and that outputs are factually accurate. Which THREE techniques should they combine?
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Use Retrieval-Augmented Generation (RAG) to ground outputs in verified documents
Option A is correct because Retrieval-Augmented Generation (RAG) grounds the model's outputs in an external, verified document store at inference time, so factual claims are anchored to authoritative sources rather than the model's parametric memory, directly improving factual accuracy. Option B is correct because training with differential privacy (e.g., DP-SGD) adds calibrated noise to gradients and bounds per-example influence, providing a formal guarantee that the model cannot memorize and expose sensitive training-data details. Option D is correct because Reinforcement Learning from Human Feedback (RLHF) fine-tunes the model against human preference and safety signals, aligning its behavior to refuse leakage and to prefer truthful, accurate responses. Option C does not belong because simply scaling parameters increases capacity and can worsen memorization of training data without any privacy or factuality guarantee. Option E does not belong because raising temperature to 2.0 increases randomness and hallucination, degrading factual accuracy rather than improving it.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✓
Use Retrieval-Augmented Generation (RAG) to ground outputs in verified documents
Why this is correct
RAG grounds generation in retrieved, verified documents rather than parametric memory, directly satisfying the factual-accuracy constraint. Because the model conditions on supplied source text instead of recalling training data, sensitive information from training is less likely to surface. It does not, however, prevent leakage from the underlying model itself.
- ✓
Train the model with differential privacy
Why this is correct
Differential privacy adds calibrated noise during training, bounding any single record's influence on learned parameters. This directly satisfies the stem's non-exposure constraint: sensitive training data cannot be reconstructed or memorised from outputs. It does not, however, address factual accuracy, so it must combine with retrieval grounding and output verification.
- ✗
Use a larger model with more parameters
Why it's wrong here
Parameter count does not prevent memorised training data from being reproduced, nor does it ground outputs in verifiable sources. Scaling up is tempting because larger models often score higher on reasoning benchmarks, and would be the right choice when the constraint is task accuracy on complex prompts rather than privacy or factual grounding.
- ✓
Apply Reinforcement Learning from Human Feedback (RLHF) to align model behavior
Why this is correct
RLHF aligns model outputs with human preferences for factual accuracy, directly satisfying the stem's accuracy constraint. However, it does not prevent training-data leakage; that requires complementary techniques such as differential privacy or data sanitisation. RLHF therefore contributes the alignment component only.
- ✗
Increase the temperature to 2.0 for creativity
Why it's wrong here
Temperature 2.0 increases sampling randomness, which raises hallucination and verbatim-recall risk instead of suppressing sensitive disclosure. High temperature is tempting for brainstorming or creative drafting, where varied phrasing is wanted; here the requirement is factual accuracy and non-disclosure, which need retrieval grounding and output filtering.
Go deeper
Related to this question
About these practice questions
One of 1,008 original Generative AI Leader practice questions on Courseiva, each with a full explanation and wrong-answer analysis — not exam dumps or protected exam content. Learn why practice questions differ from exam dumps →
JA
Written by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
This Generative AI Leader practice question is part of Courseiva's free Google Cloud certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the Generative AI Leader exam.