Trustworthy AI

A media organisation must distinguish between two content-authenticity technologies for its AI image pipeline: one embeds a cryptographically signed record of an asset's origin and edit history as metadata travelling alongside the file, while the other embeds an imperceptible signal directly into the pixels so it survives recompression and resharing. Which statement most accurately characterises the conceptual difference between these two approaches?

• ABoth invisible watermarking and provenance metadata cryptographically sign the asset, so either one alone proves the editing history and neither can be removed without invalidating the file.
• BProvenance metadata is embedded in the pixels and survives recompression, whereas invisible watermarking is external metadata that is removed the moment the file is re-exported.
• CProvenance metadata standards such as C2PA carry rich signed assertions but can be stripped when a file is re-exported, whereas invisible watermarking embeds a signal into the content itself that tends to survive transformations like compression.check_circle Correct
• DBoth technologies are visible overlays applied to the rendered image, differing only in whether the overlay is a logo or a QR code that viewers can scan.

Why A is wrong: Tempting because both relate to authenticity, but watermarking encodes a bit-pattern rather than a signed edit history, and metadata can be stripped without corrupting the visible content, so this overstates both.

Why B is wrong: Tempting as a plausible labelling, but it reverses the two: it is the watermark that lives in the signal and the manifest metadata that is strippable, so the roles are swapped.

Why C is correct: Correct: a C2PA manifest is signed metadata bound to the file and is lost if the metadata is removed, while an invisible watermark lives in the signal and persists through many transformations, so the two are complementary rather than equivalent.

Why D is wrong: Tempting because visible credential badges exist, but the question describes imperceptible and metadata-borne mechanisms, neither of which is a visible overlay, so this mischaracterises both.

A trust-and-safety team relies solely on a binary AI-image detector trained in 2023 to flag synthetic photographs. Which limitation should the team be most aware of when assessing its ongoing reliability?

• AThe detector cannot run on GPU hardware without recompilation, limiting throughput
• BBinary detectors always misclassify JPEG-compressed images as synthetic
• CDetection accuracy degrades on images produced by generators released after the detector's training cutoffcheck_circle Correct
• DPerceptual hashing prevents the detector from reading image pixels correctly

Why A is wrong: Hardware compatibility is an infrastructure concern unrelated to the fundamental accuracy challenge; it does not explain why a well-deployed detector would fail to catch new synthetic images.

Why B is wrong: This overstates a known sensitivity issue; compression artefacts can affect detector confidence but modern classifiers are calibrated against them, and the claim that compression always causes misclassification is inaccurate.

Why C is correct: Synthetic-media detectors learn statistical patterns from the specific generators included in their training data. When newer generators - especially those with improved architectures or fine-tuning pipelines - produce images with different artefact profiles, the detector's generalisation fails, leading to increased false-negative rates.

Why D is wrong: Perceptual hashing is a separate content-matching technique; it does not interact with or obstruct a pixel-based neural classifier and has no bearing on detection accuracy for synthetic media.

An engineering team deploying a customer-facing image generation service wants to prevent the model from producing content that violates usage policies, while preserving the ability to update policy rules without retraining the model. Which safety architecture best satisfies both constraints?

• AEncode all policy rules directly into the model's system prompt and rely on the model's instruction-following capability to self-enforce them at inference time.
• BDeploy a modular, rule-based safety layer such as NeMo Guardrails between the model and end users, with policies defined in a separately maintained configuration that can be updated without model retraining.check_circle Correct
• CRetrain the model periodically on datasets that omit policy-violating content so that the generation distribution no longer includes prohibited outputs.
• DUse CLIP-based post-generation filtering to score every generated image against a set of concept embeddings and discard outputs above a threshold similarity to prohibited concepts.

Why A is wrong: Prompt-only enforcement is brittle; adversarial inputs or prompt injection attacks can bypass system-prompt instructions, and updating a rule requires prompt engineering rather than a governed policy change.

Why B is correct: A modular safety layer such as NeMo Guardrails externalises policy logic from model weights into a maintainable configuration, allowing teams to add, remove, or modify rules at any time without a new training run, which directly meets the update-without-retraining requirement.

Why C is wrong: Periodic retraining can shift the generation distribution away from prohibited content but is expensive, slow to update, and cannot react to policy changes that arise between training cycles.

Why D is wrong: CLIP-based post-generation filtering can catch many prohibited images but operates solely on the output rather than controlling the generation process; updating policies requires rebuilding concept embeddings and the scoring pipeline, and latency increases due to the additional inference pass.

A multimodal generative model is trained on a large image-caption dataset scraped from the web. Audits reveal that captions describing people in professional roles skew heavily toward one demographic group. Which mitigation strategy most directly targets this bias at the data level before model training begins?

• AApply RLHF with human annotators who rate outputs for demographic balance after the model has been trained on the original dataset.
• BAdd a NeMo Guardrails safety layer at inference time that blocks outputs containing demographic references in professional contexts.
• CIncrease the model parameter count so the larger capacity can learn to represent minority demographics more accurately from the same corpus.
• DResample and rebalance the training corpus so that professional-role captions are distributed equitably across demographic groups before any training run.check_circle Correct

Why A 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.

Why B is wrong: Inference-time blocking suppresses outputs but leaves the biased representation in the model weights intact; it also risks over-blocking legitimate content.

Why C is wrong: Scaling model size amplifies patterns in the training data; a larger model trained on a skewed corpus typically reproduces and reinforces that skew rather than correcting it.

Why D is correct: Resampling the corpus to achieve demographic balance removes the skewed prior from the training distribution itself, addressing the root cause rather than downstream symptoms.

A text-to-audio generation system is observed to produce confident-sounding narration that includes factual details not present in the grounding document provided at inference time. Which architectural or pipeline mechanism most directly reduces this hallucination behaviour?

• AImplement retrieval-augmented generation so that every factual claim in the narration is traceable to a retrieved passage from the grounding document.check_circle Correct
• BSwitch the underlying language model backbone to a diffusion-based architecture so that outputs are sampled from a learned distribution rather than generated autoregressively.
• CIncrease the temperature sampling parameter during decoding to encourage more diverse output, which reduces the model's overconfidence in specific factual claims.
• DFine-tune the model on a curated dataset of factually accurate audio samples to shift the prior distribution toward truthful outputs.

Why A is correct: Retrieval-augmented generation (RAG) anchors generation to explicit source passages, providing a verifiable grounding chain that constrains the model from producing facts outside the provided context.

Why B is wrong: Diffusion-based models for audio generation address signal quality and style, but they do not inherently constrain factual content to a grounding document; hallucination is a knowledge-grounding problem, not an architecture-family problem.

Why C is wrong: Raising temperature increases output diversity and randomness but makes hallucination worse, not better; the model is more likely to fabricate varied content rather than staying faithful to the source.

Why D is wrong: Fine-tuning on accurate samples improves baseline factuality but does not bind the model to a specific grounding document at inference time; the model can still hallucinate beyond the provided context.