NCA-GENM - Trustworthy AI (5% of the exam) - Section 7.1
Understand content authenticity in generated media.
Describe the challenge of verifying whether media - text, image, audio, or video - was generated by an AI model versus captured from reality. Recognise mechanisms such as digital watermarking, provenance metadata, and content credentials that support authenticity verification.
Content authenticity
Practice question for this objective
lock_openFree sampleTrustworthy AImedium
A team must distinguish invisible watermarking from C2PA-style content credentials when planning how to mark AI-generated media. Which statement most accurately captures the difference between the two approaches?
AWatermarking hides a signal inside the content signal itself, while content credentials attach a signed metadata manifest that can be stripped from the asset.check_circle Correct
BBoth embed an external signed sidecar file, so neither survives if the media is re-encoded or stripped of its metadata wrapper.
CContent credentials hide a signal inside the pixels, while watermarking attaches a signed manifest of assertions that travels alongside the file.
DBoth techniques require cryptographic key infrastructure to verify, so neither can confirm provenance without contacting a central signing authority at read time.
Watermarking embeds a signal within the content itself, whereas content credentials attach a separable signed metadata manifest describing provenance. An invisible watermark modulates the media signal so the marker rides inside the pixels or audio and can persist through some re-encoding, while a C2PA content credential is metadata bound to the file that confirms origin only as long as it is not stripped, so the two address durability and provenance in complementary ways.
Why A is correct: This is the core distinction: a watermark lives in the pixels or audio samples and can persist through some transformations, whereas a C2PA manifest is attached metadata that survives only while it remains bound to the file.
Why B is wrong: It correctly notes a shared fragility risk, which is tempting, but watermarking embeds a signal in the pixels or samples themselves rather than a sidecar, so the premise that both rely on external files is wrong.
Why C is wrong: This swaps the two mechanisms, which is an easy confusion to make under exam pressure, but it inverts the actual definitions of both techniques.
Why D is wrong: Content credentials do rely on signing keys, which makes the claim feel reasonable, but many watermarking schemes are detected with a shared decoder and no per-asset signature or live authority lookup.
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.
Store the model name and edit log in a sidecar database and join it to each image by filename when verification is requested.
errorWhy it is wrong: A filename-keyed database is tempting because it keeps a history, but it is not bound to the file, does not travel with it, and is trivially broken by renaming or re-export, so it is not tamper-evident.
Both survive equally, because the watermark and the manifest are stored in the same metadata block and the transcoder copies that block verbatim into the new container.
errorWhy it is wrong: They are not stored together: the watermark lives in the pixel data while the manifest lives in a metadata block. Assuming a single shared block ignores how the two mechanisms are embedded and where the pipeline strips data.
Both 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.
errorWhy it 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.