Glossary
AI watermark: visible marks and invisible signals
An AI watermark is a marker added to generated content to signal its origin, either a visible logo or label rendered into the picture, or an invisible signal embedded in the pixels, audio or text that a matching detector can recognise after metadata has been removed.
Illustration · PNG file
- Signature + IHDR
- iTXt XML:com.adobe.xmprecord
- tEXt parameters
- tEXt prompt / workflow
- tEXt invokeai_metadata
- eXIf
- caBX C2PA manifest
- IDAT pixels
- IEND
AI source type declared
The file labels itself as AI-made.
IPTC DigitalSourceType such as trainedAlgorithmicMedia is read from XMP and shown with its exact value.
Check an image →Types of watermark
Visible watermarks include Snapchat’s ghost-and-sparkle mark on exported AI images, Sora’s moving watermark and Meta AI’s “Imagined with AI” marker. Invisible watermarks include Google’s SynthID, which OpenAI also applies to its generated images, and Meta’s invisible watermark on Meta AI images. The C2PA specification treats invisible watermarks as a soft binding that can reconnect a stripped file to its manifest.
A visible mark can be cropped; an invisible one needs the vendor’s detector to confirm.
How isGenAI relates to watermarks
isGenAI does not detect visible or invisible watermarks. It reads file records and matches copies, and it says so in every report where a watermark would be the only remaining signal. A visible watermark is something you can observe yourself and note as a platform or vendor observation beside the file report.
Watermarks under Article 50
Article 50(2) names watermarks among possible machine-readable marking techniques, alongside metadata and cryptographic methods, requiring them to be effective, interoperable, robust and reliable as far as technically feasible. Invisible watermarks are robust but not interoperable across vendors; C2PA is interoperable but not robust to stripping, which is why vendors combine them.
For checkers, the practical rule is simple: note any visible watermark yourself, run the vendor’s detector for its invisible one, and use isGenAI for the file record and matching copies. Each answers a different question, and none of them replaces the original file.
How to use this term when checking a file
- Upload the original file to the AI image detector or the tool named on the page.
- Find the finding that uses this term under “How we checked” and read the source field it names.
- Weigh it by kind: a signed record, a declared field, a saved prompt and a matching copy are different strengths of evidence.
- Remember that absence of a record is not evidence of a human author.
Frequently asked questions
Can isGenAI detect AI watermarks?
No. It reads C2PA records, IPTC source types, AI statements, saved prompts and matching copies. Use the vendor’s detector for its watermark.
Which AI tools add watermarks?
Google (SynthID), OpenAI (SynthID and Sora’s visible mark), Meta AI (visible marker and invisible watermark) and Snapchat (visible mark) document them as of October 2026.
Does a watermark survive a screenshot?
A visible one does unless cropped; vendors state their invisible ones are designed to survive common edits. Verify with the vendor.
Is a watermark the same as Content Credentials?
No. A watermark is in the content; Content Credentials are signed metadata. They are complementary.
Sources
Definitions follow the published standards linked above; the isGenAI behaviour described is the deployed checker as of October 2026.