Plan and manage an Azure AI solution 1.4 Implement responsible AI across generative AI and agentic systems

Your app generates marketing images with a supported Foundry image model.

Study guide skill: Implement auditing through trace logging, provenance metadata, and approval workflows

Question B3-008

Your app generates marketing images with a supported Foundry image model. Partners often re-save the images, which strips file metadata. A trust team wants to check programmatically whether an image came from Microsoft AI and needs to understand what a positive result means. Which statement is correct?

Select one answer.

  • A Use the Content Provenance Detection API, which looks for both C2PA Content Credentials and invisible watermarks; a watermark might still be found after metadata is removed, and a detection shows origin, not that the content is accurate or trustworthy ✓ Correct answer
  • B Treat a detected provenance marker as proof of which user prompted the image and who owns it
  • C Rely only on C2PA Content Credentials, because the signed manifest survives any re-save, crop, or format conversion
  • D Use the Content Provenance Detection API as the moderation step, because a positive provenance result also confirms the image isn't harmful

Answer: A

Foundry marks supported AI-generated content with C2PA Content Credentials (cryptographically signed metadata) and invisible watermarks, and the Content Provenance Detection API checks for both. Unlike metadata, watermark signals might remain when metadata is removed, although provenance might not survive every transformation. The API doesn't assess safety or determine trustworthiness, and provenance isn't proof of authorship or ownership.

Deep dive

How it works

Foundry marks supported AI-generated content in two ways:

  • Content Credentials (C2PA): a cryptographically signed metadata manifest embedded in the file. It can carry the creator, generation tools, edit history and timestamps.
  • Invisible watermarks: machine-readable signals embedded in the content itself, imperceptible during normal use.

The Content Provenance Detection API checks media for both. That matters in this scenario, because re-saving in an app that doesn't preserve metadata can strip the C2PA manifest. Unlike metadata, watermark signals might remain available even when metadata is removed. The detection page even notes that provenance can still be detected through the watermark when only a cropped portion of the content is analyzed.

Two limits shape what a positive result means:

  • Provenance isn't trustworthiness. It tells you about origin and generation history, not whether the content is accurate, truthful, harmful or trustworthy.
  • Provenance isn't authorship. It might not identify who prompted the model, who owns the content, or who publishes it.

Also remember that provenance might not survive every transformation: editing, re-saving, cropping, resizing, filters, format conversion and compression can alter or remove it. A negative result isn't proof that content is human-made.

The other options, and when you'd use them

  • Treat a marker as proof of the prompting user and owner: Microsoft says not to interpret provenance signals as proof of ownership, authorship, intent or endorsement.
  • Rely only on C2PA: C2PA is metadata, and the re-saves in this scenario are exactly the kind of workflow that can strip it. That's why watermarks exist alongside it.
  • Use provenance as moderation: Provenance doesn't determine whether content is harmful. Use content safety filtering for that.

Exam tip

C2PA = signed metadata (can be stripped). Watermark = in the content (might survive). A positive detection = origin signal only, never "safe", "true" or "who made it".

References

Last reviewed:

Ready to test yourself? The quiz asks this and every other AI-103 question, with instant feedback. Practice this in a quiz →

All Plan and manage questions · AI-103 study guide

Thanks for your feedback!