Skip to content
CreatorPrivacyKit
Learn Blog Privacy @mollycoddle330

“AI watermark” is three different mechanisms

The phrase shows up in tool roundups as if one detector could find “the watermark.” In files you actually hold, you are usually looking at a signed credential, an invisible pixel pattern, or a label a website drew on top. The field note that answers the phrase directly is the AI watermark article.

Reviewed 12 October 2026

What is an AI watermark?

The phrase shows up in tool roundups as if one detector could find “the watermark.” In files you actually hold, you are usually looking at a signed credential, an invisible pixel pattern, or a label a website drew on top. The field note that answers the phrase directly is the AI watermark article.

C2PA: Coalition for Content Provenance and Authenticity; C2PA specification 2.4; Instagram Help: label AI content before sharing; YouTube Help: disclosing use of GenAI content

Last updated

The three uses

  • Provenance credentials — C2PA / Content Credentials in the file. Inspectable, strippable. C2PA.
  • Invisible pixel marks — for example SynthID. Not removed by header stripping. Detection is probabilistic and often unavailable.
  • Visible or UI badges — “Made with AI” on a platform. May come from a credential, a classifier, or the upload path. Not in EXIF.

Why detectors disappoint

A tool that only reads C2PA will say “not detected” on a screenshot of a generated image. That is correct and incomplete. A classifier that outputs “87% AI” is inventing precision this site will not copy.

Use AI Image Origin Detector to read what is in the file. Treat a miss as a miss.

Screenshots

A screenshot usually destroys C2PA. It may or may not preserve a pixel watermark, and you should not claim either way without a test on that generator. It always preserves whatever was on screen. Screenshots and metadata.

Common questions

Can I remove an AI watermark with Metadata Cleaner?
You can drop supported C2PA, EXIF, and XMP from a JPEG, PNG, or WebP sharing copy. A pixel watermark such as SynthID remains. An Instagram AI info label, a Facebook label, a TikTok label, or a YouTube altered-content label can remain too, because those are product decisions.
Which page explains the phrase “AI watermark”?
The field note is What is an AI watermark?: credentials, SynthID, and labels as three mechanisms. This Learn page keeps the same split and links C2PA and SynthID. The file check is the Content Credentials checker.

Want the layers in one place? AI & authenticity