AI Images: IPTC DigitalSourceType and C2PA Metadata

To mark an image as AI-generated inside the file itself, write the right value into the IPTC photo metadata field called Digital Source Type. For an image produced entirely by a trained model, Google's image metadata page lists trainedAlgorithmicMedia. C2PA is a separate layer that can carry a signed record of how a file was made. A wedding videography studio that publishes both filmed work and generated pieces needs to know which layer does what, and where each can go wrong.

Field names and values are given in English because that is how the vocabulary is published. The audience here is the editor or web manager who must decide what goes into the file, what goes onto the page and who records it.

Why provenance labels exist

A viewer looking at a still from a wedding film cannot tell whether it came from a camera, a paintbrush or a model. Google's page on generative AI content says that sharing how content was created can give readers context, and it recommends adding image metadata, particularly for AI-generated images, to disclose their origin. A label is therefore a courtesy to the viewer and a signal to systems that read metadata.

Three audiences, three labels

The first audience is the visitor, who reads a caption. The second is whoever downloads the file and reuses it elsewhere, who meets the embedded fields. The third is software that reads image metadata or page markup. A caption speaks to the first, embedded fields to the second and third, and markup to the third. Skipping one leaves an audience without information.

What a label does not promise

Google's page states that it does not guarantee metadata will appear in search results. Google's documentation has no statement about ranking benefits from labelling. Treat the label as honest documentation and nothing more.

Embedded metadata versus page markup

Google's image metadata page (Opens in a new tab) describes two ways to supply image information and says one is enough. Structured data is added to each page where an image appears, so it must be repeated for every instance. IPTC photo metadata is embedded in the file, travels with the image and needs adding only once.

Practical suggestion

Use all three, and make sure they agree. When an image is re-edited, the file, the markup and the caption all change together, and one record feeds them.

The four DigitalSourceType values Google lists

Google's page names the Digital Source Type field as the one that identifies how an image was made, and lists four IPTC NewsCodes values for it.

Each value has an address under cv.iptc.org/newscodes/digitalsourcetype/, ending in the value's name. The full definitions sit in the IPTC Digital Source Type vocabulary (Opens in a new tab).

The wider IPTC vocabulary

The vocabulary covers far more than AI. It includes digitalCapture for a recording made with a digital camera, computationalCapture for several real frames merged automatically, negativeFilm, positiveFilm and print for scans of physical originals, humanEdits for corrections by people using non-generative tools, and digitalCreation for human-made work from non-generative tools. It also lists screenCapture, composite, compositeCapture and virtualRecording.

Retired terms

The vocabulary keeps minorHumanEdits, digitalArt and softwareImage for history but no longer for new use, pointing to humanEdits, digitalCreation and more specific terms respectively. A studio with an old archive may meet these in existing files.

Choosing a value for a mixed image

Real work is rarely one step. A videographer grabs a still from a venue shoot, then uses a generative tool to remove a stranger, or extends the sky. One value cannot tell the whole story, so the choice depends on what was done.

Three questions for choosing the value

1. How did the first version come into being: camera, hand, formula or model?
2. What was done afterwards: only correction, or content added or removed with a generative tool?
3. Was anything combined with other images, and was any piece synthetic?

Practical suggestion: write the three answers into the record next to the chosen value, so that a colleague can see why it was chosen. Where a generative tool changed part of a real photograph, add a plain-language caption naming the part.

Credit, creator and copyright fields to keep

Digital Source Type is one of several fields Google reads. The others are Copyright Notice, Creator, Credit Line, Licensor URL and Web Statement of Rights. Creator names the photographer or an organisation where appropriate. Credit Line names the person or organisation to be credited on publication. Copyright Notice names the current owner of the copyright.

Do not strip rights fields

Google's page advises against removing rights-related metadata, especially the creator, credit line and copyright notice fields. Check after every processing step that they are still present.

Who is the creator of an AI-assisted image?

Google's documentation has no statement on who counts as the creator of an AI-assisted image. The studio decides, and records the decision. Practical suggestion: name the person who directed the work, name the tool in the internal record and keep the wording the same in the caption, the file and the markup.

Licensing fields

Licensor URL points to a page explaining how to license the image, and Web Statement of Rights points to a page describing its licence. Google's page says the latter is required for the licensable badge. These fields are independent of the AI label, but a studio that licenses images can fill them in alongside the provenance field.

What C2PA adds

C2PA develops technical standards for verifying where media comes from and how it has changed. The C2PA specifications site (Opens in a new tab) collects its documents and groups the technical specifications under "Content Credentials" and "Attestations", with a separate group of guidance documents that includes guidance for AI and machine learning. How a manifest or signature is built is defined in the linked specifications, which are outside this discussion.

What Google says about C2PA

If an image contains C2PA metadata, Google can extract those details and may show them in "About this image". That can reveal how an image was created or whether AI tools edited it. The metadata must come from a signer that uses C2PA 2.1 or later, and the manifest must be signed by a certificate on the C2PA Trust List. As with IPTC, Google does not guarantee display.

C2PA versus IPTC

An IPTC value is a field somebody writes. C2PA is a signed record. Google's documentation has no statement on whether one replaces the other or whether having one makes the other unnecessary, so treat them as separate layers.

The conflict rule

Google's page says that if structured data and embedded IPTC information conflict, Google will use the structured data. That matters when a web developer maintains the markup and a photo editor maintains the file.

Preventing conflict

• One record per image feeds both the file fields and the markup.
• Edits to the image trigger an update to both.
• Whoever changes one confirms the other.

Keeping metadata through upload and conversion

Files change hands: from camera to editor, from editor to the content system, from the system to the page. Whether the embedded fields survive depends on the tool and its settings, and Google's documentation has no general statement about it. Measure rather than assume.

A verification routine

1. Take the final file as served on the live site.
2. Read the Digital Source Type, Creator, Credit Line and Copyright Notice fields with a metadata reader.
3. If a field is missing, repeat the processing steps one at a time to find where it vanished.
4. Rewrite the fields at the last step and note the result in the record.

Rewriting is not always enough

Practical suggestion: if you use C2PA, check the signed record after every processing step, and consult C2PA's guidance documents for how signatures behave.

Captions, alt text and on-page labels

Alt text describes what the image shows for people who cannot see it. A provenance note says how the image was made. They are different jobs. For alt text and licensing basics, see article image alt text and stock image licensing.

Sample wording

• Caption for a fully generated image: "Generated with an AI image tool, directed and edited by our team."
• Caption for a mixed image: "Still from our wedding film; the sky was extended with a generative tool."
• Alt text: "Bride and groom walking along a harbour wall at sunrise."

Images of unknown origin

A studio sometimes inherits images whose production steps nobody remembers: a former editor's archive, a donated set, a file received without notes. Do not fill the gap with a guess. Practical suggestion: add a line "origin not confirmed" to the record, leave the Digital Source Type field empty and state in the caption only what is known. If the facts arrive later, update the field, the caption and the markup together. A wrong value is worse than an empty field, because it reads as a firm statement.

The director's own words

A director who works with generative tools may want to describe the process in their own voice. The page can carry that as the director's statement, while the embedded field and the caption stay short, neutral and checkable. Keep the artistic statement and the provenance label apart.

A studio's workflow: who records what

Provenance belongs in the record at intake, not in a search after publication. For each image keep one row: title, method, tool, value, responsible person, date.

Roles

The contributor reports how the image was made. The editor chooses the value. The web manager writes the file fields and the markup. A second person reads the final file with a metadata reader. For how bylines and accountability appear on the site, see how to build an author page.

Bringing the archive into line

An archive of older photographs lacks provenance metadata, and fixing everything at once is not necessary. Set priorities: pages that are live first, new uploads next, the dormant archive last. Give each group a target date and follow progress in the record table.

• List live images with incomplete provenance records.
• Start with images made or altered with generative tools.
• After each fix, read the file back with a metadata reader.

Different tools, different field names

The same file may be shown with different labels by different metadata readers. A field that one tool calls Digital Source Type may appear under another name elsewhere. Do not tie the check to one tool. Looking with two different readers separates a labelling quirk from a true loss.

Shared vocabulary inside the studio

Everyone who works with provenance labels needs to use the same terms in the same sense. An editor who says "AI-assisted" may mean only a colour correction, while a web manager hears "output of a generative model". Practical suggestion: define three team terms in the record form and map each to its IPTC counterpart.

The same note on every surface

One image can appear on a project page, in a newsletter, in a social post and in a printed brochure. If each surface words the origin differently, viewers see contradictory accounts. Set one short master wording in the record and let every surface reuse it. When it changes, all surfaces change together.

A template for the on-page note

Use a short, consistent template for the provenance note on the page. It has three parts: how the image was made, who directed it and any later intervention. For example: "This image was made with a generative image model, directed by a member of the studio and colour-corrected afterwards." The template ties the member's statement, the caption and the file fields to one record.

Decision rule: which value, which layer?

Worked example: a wedding videography studio

Example scenario: a wedding videography studio publishes a project page about a harbour-side wedding. It includes a still from the film with one distracting power line removed by a generative fill, and a generated illustration of the venue's courtyard as it might have looked long ago.

What the editor does

1. The still is recorded as a camera original with a generative edit. The value compositeWithTrainedAlgorithmicMedia is written and the caption says that a power line was removed with a generative tool.
2. The illustration is recorded as fully generated. The value trainedAlgorithmicMedia is written and the caption says the scene is an imagined reconstruction.
3. Creator and Credit Line name the videographer and the studio, and Copyright Notice names the owner.
4. The files are uploaded and read back with a metadata reader. The fields are present.
5. The page markup is updated to repeat the same information.
6. The record is completed and the approving editor's name is entered.

Where the decision rule runs out

The decision table steers the first reaction but cannot settle every case. A picture may combine camera frames, generative additions and hand-drawn layers. One value may then fall short. Pick one of the composite values and describe in the caption which part was made how. Keep the value and the caption consistent. Anyone unsure should reread the vocabulary definition and write the reasoning into the record.

A label is not a verdict on quality

An image made with generative tools is not a lesser image. The label states how the picture was made, nothing more. A neutral tone respects both the contributor and the viewer.

Misreadings

"I wrote the field, so no caption is needed"

Visitors do not see embedded fields. A caption may still be needed.

"With C2PA the IPTC field is redundant"

Google's documentation has no statement that says so. Treat them as separate.

"A label lifts rankings"

Google's documentation has no such claim. The purpose is transparency about origin.

"Written once, kept forever"

Processing can change fields. Check the file that is actually served.

Recording the reader you used

Note which metadata reader you used for each check. The same file can show different field names in different readers, so the tool itself is part of the evidence. Add the reader's name, the date and the checker's name to the record line. If a field disappears later, you can tell which file version and which date the loss was noticed.

Monitoring after launch

Publishing is not the end of the job. If the site's image pipeline changes, for example when a new optimisation tool is added or file formats are switched, fields that were once kept can vanish. Practical suggestion: after every change to the pipeline, reread the metadata of a few live images and add the result to the record. A sample is enough, as long as it includes both generated and non-generated images.

Pre-publication checklist

• The production steps of the image are on record.
• The Digital Source Type value matches the vocabulary's definition.
• Creator, Credit Line and Copyright Notice are filled and retained.
• The served file was read back with a metadata reader.
• Markup and file say the same thing.
• A visible caption and alt text were written separately.
• If C2PA is used, the signer version and certificate were checked.
• The record is complete and a named person is responsible.

For editorial review of AI-assisted material, see AI-written articles and human editing and the editor review checklist. For product images, Google's generative AI content page (Opens in a new tab) says Merchant Center requires AI-generated images to carry IPTC metadata with the TrainedAlgorithmicMedia value.

A quarterly review

Every three months, review the record form, the embedded fields and the page markup together. Sample a few images at random and confirm that the three layers say the same thing. If they differ, find out why: a re-edit, a tool that dropped the metadata or a record that was never updated. Carry the finding back into the record form and the workflow so the same problem does not recur.

Questions from viewers

Viewers may ask how a picture was made. The record is the basis for the answer, and using it as the single source keeps answers consistent with the file and the caption.

Reviewing the record after a major tool update

When a camera, an editing program or a generative tool receives a major update, its metadata behaviour can change. Practical suggestion: after any such update, run one test file through the whole chain and read the fields back. Treat the first file after an update as a test case before real work depends on it. Note the tool version in the record so that a later loss of fields can be traced to a specific change.

Training new contributors

A short session for each new contributor pays for itself. Show one real image from the archive, its record row, the embedded fields read back with a metadata reader and the caption that appears on the page. Seeing the same information in four places makes the rule concrete. Ask the newcomer to fill in a record row for one of their own images and check it together.

Frequently Asked Questions

Which IPTC value marks an AI-generated image?

For an image produced entirely by a trained model, Google's page lists trainedAlgorithmicMedia. For mixed work there are compositeSynthetic and compositeWithTrainedAlgorithmicMedia.

What is C2PA?

It is a coalition's set of technical standards for verifying where media comes from and how it changed. Google can extract C2PA details and may show them in "About this image".

What happens if markup and the file disagree?

Google uses the structured data. Keep both aligned.

Will the metadata show in search results?

Google does not guarantee it. The metadata is a way to state origin clearly.

May I remove the rights fields?

Google's page advises against removing creator, credit line and copyright notice fields.

Does metadata survive processing?

Google's documentation has no general statement on this. Check the final served file with a metadata reader.

Conclusion

Describe an AI image's origin in three layers: the IPTC Digital Source Type in the file, a signed C2PA record where you use one, and a visible note on the page. Choose the value from the vocabulary's definitions, keep the rights fields, read back the processed file and keep markup and file in agreement.

If you want an editor-supervised workflow for your content, the content writing service page lists the options.