isGenAI.com

Glossary

Machine-readable marking: how AI output is made detectable

Machine-readable marking is the EU AI Act Article 50(2) requirement that providers mark synthetic audio, image, video and text outputs in a format software can read, so the content is detectable as artificially generated or manipulated, using techniques such as C2PA metadata, IPTC fields and watermarks.

Definitions and isGenAI behaviour as of October 2026. Updated .

See live marking observationsFrequently asked questions
Code-drawn illustration for Machine-readable marking

Illustration · PNG file

  1. Signature + IHDR
  2. iTXt XML:com.adobe.xmprecord
  3. tEXt parameters
  4. tEXt prompt / workflow
  5. tEXt invokeai_metadata
  6. eXIf
  7. caBX C2PA manifest
  8. IDAT pixels
  9. 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 →
AI source type declared: where this record sits inside a PNG file and what the isGenAI report says about it. The picture is a code-drawn illustration.

The techniques

The Article and the Commission’s guidance treat several mechanisms as candidates: signed provenance metadata (C2PA Content Credentials), declared metadata (IPTC Digital Source Type in XMP), invisible watermarks (SynthID and vendor equivalents) and cryptographic methods. Each trades interoperability against robustness: metadata is readable by anyone but stripped by platforms; watermarks survive stripping but need the vendor’s detector. The law asks for solutions that are effective, interoperable, robust and reliable as far as technically feasible.

How isGenAI checks for a mark

Upload an output to the AI image detector or C2PA checker: the report shows whether a C2PA manifest, an IPTC source type or a saved record is present, each named by field. The Article 50 machine-readable marking page gathers live observations from public scorecards. isGenAI reads metadata marks; it does not detect watermarks, and says so.

Robustness in practice

A mark that disappears on the provider’s own export, or on every social upload, exists but does not work. The platform pages document what survives each service, which is the evidence a provider needs when choosing a technique.

Testing a mark means testing real outputs through real pipelines: export from the product, upload to the platforms users choose, download again and check what remains. isGenAI’s compare tool makes the before-and-after difference explicit for every record type.

How to use this term when checking a file

  1. Upload the original file to the AI image detector or the tool named on the page.
  2. Find the finding that uses this term under “How we checked” and read the source field it names.
  3. Weigh it by kind: a signed record, a declared field, a saved prompt and a matching copy are different strengths of evidence.
  4. Remember that absence of a record is not evidence of a human author.

See live marking observations

Frequently asked questions

What counts as machine-readable marking?

Any provider-side mark software can read: C2PA manifests, IPTC metadata, invisible watermarks or other cryptographic methods, meeting the effectiveness and robustness test as far as feasible.

Is a visible label machine-readable?

Not by itself. Visible disclosure is the deployer’s duty under 50(4); machine-readable marking is the provider’s under 50(2).

How do I check whether an image carries a mark?

Upload the original to isGenAI; the report lists any C2PA manifest, IPTC source type or saved record it finds.

Which technique is best?

Vendors increasingly combine signed metadata with a watermark, because each covers the other’s weakness.

Sources

Definitions follow the published standards linked above; the isGenAI behaviour described is the deployed checker as of October 2026.