Skip to content
CreatorPrivacyKit
Field notes / AI provenance
C2PA · SynthID · metadata

What is an AI watermark?

People use “AI watermark” for a signed credential, a pixel mark such as SynthID, or a label a platform drew. Check those layers on the file you still have.

Published September 15, 2026 · Updated October 12, 2026

What is an AI watermark?

People use “AI watermark” for a signed credential, a pixel mark such as SynthID, or a label a platform drew. Check those layers on the file you still have. C2PA can associate a file with a participating generator or edit history when the credential is still on the file.

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

Signed claim

C2PA can associate a file with a participating generator or edit history when the credential is still on the file.

Vendor mark

SynthID is useful inside the ecosystems that embed and detect it. Absence is not a global “not AI.”

Missing ≠ human

No supported signal is inconclusive. Screenshots and re-saves throw evidence away.

Three different things get called a watermark

LayerExampleWhat it actually is
Signed provenanceC2PA / Content CredentialsA signed record attached to the file and its history
Imperceptible markSynthIDA signal in the content, detected by the systems that know how to look for it
Statistical guessVisual “AI detector”A model score about pixel patterns. It can be wrong.

C2PA is not a certificate that the scene happened in the real world. SynthID is not a universal stamp on every generator. A visual score is not cryptography.

Signed provenance is a claim about a file

When a C2PA manifest is present, read the signer and the actions — c2pa.created is not the same as inpainting on a camera photo. When it is absent, the file may never have had credentials, or they may have been stripped by a screenshot, a messenger, or an export.

Optional official OpenAI verification can look for supported OpenAI C2PA and SynthID. That step sends the selected image to OpenAI only after you click. Local inspection of containers and metadata does not.

SynthID is vendor-scoped

Google’s SynthID detector speaks to supported Google AI content. A miss there does not rule out ChatGPT images, Midjourney, Firefly, or an open model. Do not treat one vendor’s watermark as a census of the whole internet.

A screenshot is a new file

C2PA lives with the original container. Capture the pixels and you usually lose that packet. Pixel-domain marks may survive some transforms and fail others. If provenance matters, keep the original download, not the phone screenshot of the post.

More on that split: what survives a screenshot and how to check Content Credentials.

How to read a result without inventing a percentage

  1. Positive C2PA: record the issuer and verbs, not “AI: yes.”
  2. Positive watermark: name the system that matched.
  3. Classifier score: keep it labeled as a model output.
  4. Nothing found: say exactly that. It is not “proven human.”

If you need to inspect a file, use the Content Credentials checker. The C2PA viewer online opens the same inspector. This article explains the phrase. It is not a second detector.

Platform labels are a third layer

Instagram and Facebook can show “AI info” (the older public wording was “Made with AI”). TikTok can label from Content Credentials, a creator disclosure, or a watermark it says only TikTok can read. YouTube can carry a C2PA credential that the whole video was made with AI into “How this content was made,” and it can add a label the creator cannot remove. Those labels are not EXIF. The walkthrough is Instagram AI info. The Learn split is what people mean by an AI watermark.

Questions

What is an AI watermark?

It is not one object. In files you can hold, it is usually a C2PA Content Credential, an imperceptible mark such as SynthID, or a badge a website drew. Metadata tools only touch the credential layer. A pixel watermark such as SynthID remains after that strip.

Does a missing watermark mean the image is human?

No. A miss is a miss. Screenshots drop manifests. Not every generator writes SynthID. Read what SynthID is before you treat a blank result as a camera original.