IPTC Digital Source Type Explained: The AI Tag Inside Your Image

When a platform says it reads “industry-standard AI metadata,” one of the things it usually means is a single field: IPTC Digital Source Type. It is a short URI written into the image’s XMP metadata, and its value states how the image came to exist. One value means a camera capture, another means created with generative AI, and a third means a real photo edited with generative AI. It is small, easy to read, and the reason a lightly retouched photo can end up with an AI label.

#Where the field lives

The IPTC Photo Metadata Standard defines the property as Iptc4xmpExt:DigitalSourceType in XMP, which ExifTool reads as XMP-iptcExt:DigitalSourceType. It holds at most one value, and that value is a URI from the IPTC Digital Source Type vocabulary, written in full as http://cv.iptc.org/newscodes/digitalsourcetype/ followed by a term such as trainedAlgorithmicMedia. The property has existed since 2009; the AI terms were added in 2022 and 2023 and relabeled in 2024.

C2PA reuses the same vocabulary. Its specification lets each recorded action carry a digitalSourceType drawn from the IPTC terms, and for media created from scratch by a generative model it requires the first action, c2pa.created, to record one. So the same URI can appear twice in one file: once as loose XMP, and once inside the signed manifest.

#The values that mean AI

The vocabulary also covers media that is not AI-generated, which matters just as much when reading a file: digitalCapture for a normal camera capture, computationalCapture for multi-frame merges such as smartphone HDR, algorithmicallyEnhanced for sharpening or noise reduction, screenCapture, humanEdits, digitalCreation, and a general composite. In the 2024 revision IPTC retired minorHumanEdits and digitalArt in favor of humanEdits and digitalCreation.

#Who writes it and who reads it

• Generators. In 2023 Midjourney and Shutterstock agreed to embed the property in generated images, and IPTC recommended trainedAlgorithmicMedia for them. Google uses both trainedAlgorithmicMedia and compositeWithTrainedAlgorithmicMedia, the latter for edits such as Magic Eraser on Pixel phones. OpenAI writes it through C2PA. Meta AI images carry trainedAlgorithmicMedia.
• Social platforms. Meta named IPTC alongside C2PA as the signals behind its “AI info” label. Pinterest says it analyzes a Pin’s metadata following the IPTC standard, together with classifiers, for its “AI modified” label.
• Search. Google Search Central lists Digital Source Type as a recommended image metadata property and supports the AI values, while noting it does not guarantee the metadata will be shown.

#How to read it yourself

1. Run exiftool -XMP-iptcExt:DigitalSourceType image.jpg. An empty result means the XMP field is absent, not that the image is a photograph.
2. Open the file at contentcredentials.org/verify to see whether a signed C2PA manifest also records a source type.
3. Compare the two. A signed value is evidence of what the signing tool declared; a loose XMP value is only what the last program to save the file wrote.

#Removing it, and what that does not change

Because the field is ordinary metadata, anything that rewrites a file without its XMP drops it: ExifTool, many exports, a canvas re-encode, or the C2PA removal flow that also clears the manifest. That is a reasonable step on images you own, for privacy or before a re-edit. It removes a declaration, not the facts. An invisible watermark such as SynthID stays in the pixels, platforms may apply their own classifiers, and writing digitalCapture into an AI image to pass it off as a photograph is misrepresentation, not cleanup. The provenance layers primer covers how the pieces fit together.