NCA-GENL Trustworthy AI Practice Question
A media company is deploying an LLM that writes first drafts of news briefs. The editorial board wants safeguards that reduce the risk of the model emitting defamatory or unverified claims about named individuals before a human editor reviews the draft. Which two measures best address this risk? (Choose two.)
⚠ Common exam trap
The trap here is treating determinism or a larger context window as a factual safeguard, when only grounding and output moderation actually constrain unsupported claims about people.
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
✓
Add an output moderation rail that flags or blocks sentences making unverified assertions about named people before the draft is shown to the editor.
The risk is that unverified or defamatory assertions reach the editor before review. An output moderation rail provides a runtime filter that flags or blocks such sentences, while retrieval-augmented grounding requires person-specific claims to be supported by source documents. Together they reduce fabrication at generation and catch problematic statements before display. Context size, sampling settings, and logging changes do not constrain claim veracity.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✓
Add an output moderation rail that flags or blocks sentences making unverified assertions about named people before the draft is shown to the editor.
Why this is correct
An output moderation rail inspects generated text before it reaches the editor and can flag or suppress sentences that assert unverified claims about named individuals. This creates a pre-review safety layer that reduces the chance a defamatory statement is surfaced even casually. It directly targets the risk described and complements, rather than replaces, human editorial judgment. The rail should be tuned to avoid excessive false positives that would erode editor trust.
- ✗
Lower the model's top-p sampling value to make the generated text more deterministic.
Why it's wrong here
Reducing top-p narrows the sampling distribution, which can make phrasing more consistent, but it does not prevent the model from asserting false claims. A deterministic model can still confidently produce an unverified statement if that pattern is likely given the prompt. Determinism improves reproducibility, not factual grounding. This measure does not reduce the defamation risk the editorial board is trying to mitigate.
- ✗
Increase the model's context window so it can ingest the entire archive of past articles for every draft.
Why it's wrong here
A larger context window allows more input tokens but does not ensure the model uses them accurately or avoids fabrication. Stuffing an archive into context can introduce irrelevant or contradictory material and increase latency and cost. It also provides no mechanism to verify that a specific claim about a person is supported. Context size is a capacity parameter, not a grounding or moderation control, so it does not address the risk.
- ✗
Disable logging of prompts and outputs to protect the privacy of the individuals mentioned in drafts.
Why it's wrong here
Removing logs may reduce one privacy exposure, but it eliminates the audit trail needed to investigate how a problematic claim was generated and to improve safeguards. It also does nothing to prevent the model from emitting defamatory content in the first place. The scenario is about pre-review safety, not data retention. Disabling logging weakens governance and is unrelated to the risk being addressed.
- ✓
Ground the drafting step in a retrieval-augmented pipeline that requires each claim about a person to be supported by a retrieved source document.
Why this is correct
Retrieval-augmented generation constrains the model to content drawn from an approved corpus, and requiring source support for person-specific claims reduces fabrication. When the pipeline can attach a retrieved document to each assertion, the editor can verify provenance quickly and unsupported claims can be withheld. This addresses the root cause of unverified assertions rather than only filtering symptoms. It pairs well with an output rail as a defense-in-depth measure.
About these practice questions
Courseiva writes every NCA-GENL question from scratch — 367 in total, each with an explanation and a wrong-answer breakdown. None are copied from real exams or dumps. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official NVIDIA exam blueprint
This NCA-GENL practice question is part of Courseiva's free NVIDIA 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 NCA-GENL exam.