Trustworthy AI

A financial services company is deploying a customer-facing chatbot powered by a large language model. The team wants to prevent the model from producing responses that contain account numbers, credit card details, or regulatory advice that could expose the company to liability. Which architectural approach most directly addresses both data-leakage and liability risks at the output stage?

• AApply a retrieval-augmented generation pipeline so the model draws only from pre-approved documents, removing any need for output scanning.
• BFine-tune the base model on a corpus of compliant responses so it learns to avoid sensitive outputs, replacing the need for runtime guardrails.
• CEnforce strict input validation on the user's query to block prompts that mention account numbers or financial terms before they reach the model.
• DAdd a dedicated output-validation layer that applies regex and classifier-based checks to detect and redact PII patterns and flag regulated-content categories before the response reaches the user.check_circle Correct

Why A is wrong: RAG limits the knowledge base used during generation, which reduces hallucination risk, but it does not guarantee that the model will never produce sensitive content or regulatory advice. A model can still compose harmful output from retrieved context, so RAG alone is insufficient for output-stage enforcement.

Why B is wrong: Fine-tuning on compliant examples can shift the model's distribution toward safer outputs, but it offers probabilistic rather than deterministic protection. Novel inputs can still elicit non-compliant responses, and fine-tuning alone cannot satisfy a hard policy requirement like blocking PII.

Why C is wrong: Input validation filters what enters the model, which is a useful safety layer, but it does not address output risk. The model can produce sensitive content from benign-looking prompts through indirect reasoning or retrieved context, so input-only controls leave the output stage unprotected.

Why D is correct: An output-validation layer is the correct mechanism for catching harmful content after generation and before delivery. Regex detects structured PII such as card numbers; classifiers handle semantic categories like financial advice. Together they enforce policies regardless of what the model produces.

An engineering team must serve a 13-billion-parameter chat model to many concurrent users on a fixed GPU budget. They are considering post-training INT8 weight quantisation to fit more concurrent sessions per GPU. Before approving this for production, which safeguard most directly protects against the quantised model degrading on the safety-critical behaviours the full-precision model passed?

• AConfirm the GPU supports INT8 tensor operations so the quantised weights execute on dedicated low-precision hardware paths.
• BIncrease the maximum serving batch size so the throughput gained from quantisation is fully realised under peak concurrent load.
• CRe-run the safety and refusal evaluation suite on the quantised model and require it to meet the same thresholds as the full-precision baseline before release.check_circle Correct
• DReduce the maximum output token length so any degraded responses from the quantised model are truncated before reaching the end user.

Why A is wrong: Hardware INT8 support governs throughput and is tempting because it is a real prerequisite for the speedup, but it says nothing about whether the quantised model still behaves safely, so it does not protect the safety-critical behaviours.

Why B is wrong: Raising batch size captures the efficiency benefit and is appealing for scaling, but it is an optimisation of serving capacity and provides no check on whether safety behaviours survived quantisation.

Why C is correct: Quantisation can shift model behaviour in subtle ways, so re-running the same safety and refusal evaluations and gating on the baseline thresholds is the direct safeguard that catches any regression introduced by the precision change.

Why D is wrong: Shorter outputs limit exposure and seem like a containment measure, but truncation does not verify behaviour and can cut off a response mid-refusal, so it fails to protect the safety-critical behaviours.

A retrieval-augmented enterprise assistant must give users a way to verify every factual claim it makes against the source documents, so reviewers can confirm answers are grounded rather than fabricated. The team wants a design change that builds verifiability into each response at scale. Which approach most directly achieves this?

• ARequire the model to return inline citations that link each factual claim to the specific retrieved passage that supports it.check_circle Correct
• BRaise the retrieval similarity threshold so that only the single most similar document chunk is ever passed to the generator.
• CLower the generation temperature so the model produces more conservative and cautious wording for each factual claim.
• DCache the most frequent answers so repeated questions return the same vetted response without re-querying the model each time.

Why A is correct: Inline citations attach each claim to the exact supporting passage, giving reviewers a direct path to verify grounding at scale and to spot any claim that lacks a source, which is precisely the verifiability the team needs.

Why B is wrong: A stricter threshold can improve retrieval precision and is tempting because it sounds like tighter grounding, but it does not let a reviewer trace which claim came from which source, and discarding context can omit needed evidence.

Why C is wrong: Conservative wording may reduce overt speculation and is appealing for trust, but it provides no link between claims and sources, so a reviewer still cannot confirm whether any given statement is grounded.

Why D is wrong: Caching vetted answers improves consistency and is tempting for reliability, but it only covers previously seen questions and adds no per-claim traceability for new responses, so it does not deliver general verifiability.

A team is training a large language model using reinforcement learning from human feedback (RLHF). After the supervised fine-tuning stage, they notice the reward model assigns high scores to verbose, flattering responses even when those responses contain factual errors. Which action most directly addresses this reward modelling failure?

• AIncrease the KL-divergence penalty coefficient in the PPO update to prevent the policy from straying too far from the supervised fine-tuned baseline model.
• BSwitch from PPO to direct preference optimisation (DPO) to remove the explicit reward model from the training loop entirely.
• CRetrain the reward model using annotator guidelines that explicitly penalise factual inaccuracies, and add calibration checks against a held-out factuality dataset.check_circle Correct
• DApply constitutional AI prompting rules during inference to filter out verbose responses before they reach the end user.

Why A is wrong: A larger KL penalty limits policy drift, which is a useful regularisation tool, but it does not fix a miscalibrated reward model. The reward model would still score verbose flattery highly; the policy would simply learn that behaviour more slowly rather than learning a different, more accurate behaviour.

Why B is wrong: DPO sidesteps a standalone reward model, which is a legitimate alignment technique. However, DPO still relies on preference data collected from human annotators; if those annotators reward verbosity over accuracy, the same bias will persist in the preference pairs used to train the DPO policy.

Why C is correct: The root cause is that annotators rewarded style over accuracy, so the reward model learned a proxy objective. Revising annotator guidelines to penalise inaccuracies and validating against a factuality dataset corrects the reward signal at its source, which is the direct fix for reward hacking through verbosity.

Why D is wrong: Constitutional AI rules applied at inference time are a safety layer, not a training-time fix. They may reduce surface harm but do not correct the underlying reward model, so the policy continues to learn the wrong objective during subsequent training rounds.

A product team must decide where to apply alignment effort for a customer-facing assistant. They have a strong base model and limited budget. One engineer argues that simply scaling the model larger will make it safe, another argues that capability and alignment are separate problems that scale independently. Which statement best reflects the responsible-AI consensus a team should plan around?

• AIncreasing model scale reliably improves alignment, so a sufficiently large model needs no dedicated safety training before deployment.
• BAlignment is achieved purely by filtering the pre-training corpus, so no post-training intervention is required once the data is clean.
• CAlignment is fully guaranteed by adding a single refusal phrase to the system prompt, which constrains every possible model output.
• DCapability and alignment are distinct objectives, so a capable model still requires explicit alignment work such as preference tuning and evaluation to behave safely.check_circle Correct

Why A is wrong: It is tempting because larger models are more capable and often more steerable, but scale improves capability without guaranteeing aligned behaviour, and unsafe behaviours persist or sharpen at scale, so relying on size alone is unsafe.

Why B is wrong: Data curation genuinely reduces some harms and is part of responsible practice, but it cannot by itself instil instruction-following, refusal, or honesty behaviours, which require dedicated post-training alignment steps.

Why C is wrong: A system-prompt instruction does shape behaviour and is cheap, but it is easily bypassed by rephrasing or role-play framings and provides no robust guarantee, so it cannot stand in for genuine alignment work.

Why D is correct: Responsible-AI practice treats raising capability and ensuring aligned behaviour as separate efforts, so even a strong base model needs explicit alignment and evaluation rather than capability gains alone to be deployable safely.