NCA-GENM - Trustworthy AI (5% of the exam) - Section 7.2
Build trustworthy generative models.
Apply responsible AI principles - including fairness, transparency, and safety guardrails - when designing and evaluating generative models. Recognise how bias in training data, unfiltered outputs, and lack of explainability can undermine trust in a deployed generative system.
Responsible AI
Practice question for this objective
lock_openFree sampleTrustworthy AImedium
A team building a generative model wants to reduce demographic bias that originates in the training corpus itself rather than correcting it after generation. In principle, which characteristic defines a data-level mitigation as opposed to an inference-time mitigation?
AIt applies a fixed post-processing filter that rewrites each generated output to remove sensitive terms before the response is returned to the user.
BIt adjusts the decoding temperature and top-p sampling at generation time so that lower-probability tokens describing under-represented groups appear more often.
CIt rebalances, reweights, or curates the training examples so that the distribution the model learns from is more representative before any training begins.check_circle Correct
DIt inserts a runtime safety classifier that blocks any response whose content the classifier scores as demographically skewed.
Recognise that data-level bias mitigation reshapes the training distribution before training, distinct from inference-time output controls. Bias frequently enters a generative model through a skewed corpus; rebalancing, reweighting, or curating those examples changes the distribution the model is optimised against, addressing the source rather than masking symptoms in finished outputs.
Why A is wrong: Tempting because it reduces visible bias, but rewriting outputs happens at inference time and leaves the learned representation untouched, so it is not a data-level mitigation.
Why B is wrong: Tempting because sampling changes outputs, but temperature and top-p operate during inference and do not change what the model learned, so they are not data-level fixes.
Why C is correct: Correct: data-level mitigations act on the corpus, through rebalancing, reweighting, or curation, so the model never internalises the skew in the first place.
Why D is wrong: Tempting because classifiers govern outputs, but a runtime gate is an inference-time guardrail acting on finished generations, not an intervention on the training data.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-GENM bank for this domain.
Apply RLHF with human annotators who rate outputs for demographic balance after the model has been trained on the original dataset.
errorWhy it is wrong: RLHF can nudge model behaviour after training but does not correct the skewed distribution already baked into the pre-training corpus; the underlying representations remain biased.
Increase the classifier-free guidance scale at inference so the model follows prompts more literally.
errorWhy it is wrong: Higher guidance sharpens prompt adherence and is a real generation control, but it acts at inference on an already biased model and cannot remove an association the weights have already absorbed from skewed data.
Bake the current policy into the training objective and retrain a fresh model whenever the rules are updated.
errorWhy it is wrong: Encoding rules directly into the weights can enforce policy, but it requires a costly retrain on every change, which is exactly the expense the team must avoid.