NCA-GENL Trustworthy AI Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Fine-tune the base LLM on the full internal loan policy corpus and rely on the fine-tuned weights to reproduce policy text accurately.

  Why it's wrong here

  Fine-tuning bakes policy content into model weights, but it provides no per-response citation to the exact source paragraph, and fine-tuned models can still hallucinate or blend policies. When policies change, retraining is required, and there is no runtime mechanism to reject unsupported claims. This fails the traceability requirement even if accuracy on static test questions looks acceptable.

• ✗

  Deploy the model behind an API gateway that logs every prompt and response, then have compliance staff review the logs weekly.

  Why it's wrong here

  Logging creates an audit trail of what was said, but it does not prevent unsupported claims from reaching users in the first place, nor does it bind answers to specific source paragraphs. Weekly manual review is after-the-fact and cannot scale. The requirement is enforcement at inference time, which logging alone does not provide, so this approach addresses detection rather than prevention.

• ✗

  Increase the model's temperature so that responses draw on a wider range of internal knowledge, improving coverage of loan policy topics.

  Why it's wrong here

  Raising temperature increases sampling randomness, which makes outputs less deterministic and more likely to drift from source material. It does not create traceability to a specific paragraph, and it actively increases hallucination risk. For a compliance-bound loan policy assistant, higher creativity is the opposite of what is needed; the model should be tightly constrained and grounded, not encouraged to explore alternative phrasings.

• ✓

  Enable retrieval-augmented generation with citations and reject any answer whose claims are not grounded in the retrieved passages.

  Why this is correct

  RAG with citation grounding ties every generated claim to retrieved source passages, so the compliance team can audit the exact paragraph. Adding a grounding check that rejects ungrounded answers prevents plausible but unsupported statements from reaching users. This directly satisfies both traceability and the no-unsupported-claims requirement, using standard NeMo Guardrails and retrieval patterns rather than relying on the model's parametric memory.