IPTC Digital Source Type values explained: what trainedAlgorithmicMedia and the others mean
The IPTC Digital Source Type field holds a short web address that says how a digital image was made, according to whoever wrote it. IPTC's list has 17 current values, from digitalCapture for camera photos to trainedAlgorithmicMedia for generative AI output. The field is a label, not proof, and anyone can add or remove it.
What is the IPTC Digital Source Type field?
IPTC maintains a photo metadata standard. Its Photo Metadata User Guide describes the field as one that "indicates the media source from which the digital image was created". The guide says the field was added in 2008, extended in 2022 for generative AI, and updated again in 2024.
The value written into the field is the identifier of a term, not its display name. Each identifier is a web address built the same way, for example http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia. The full list of terms and definitions is IPTC's Digital Source Type vocabulary.
The same vocabulary also appears inside Content Credentials. Section 18.14.4.5 of the C2PA specification, version 2.2 says an action in a signed record may carry a digital source type, and that it must be an IPTC term or one of two C2PA values. One C2PA value is for effectively empty media. The other is for data, such as a CSV file, that a trained model produced instead of media.
What do trainedAlgorithmicMedia and compositeWithTrainedAlgorithmicMedia mean?
These are the two values most people search for.
- trainedAlgorithmicMedia. IPTC's label is "Created using Generative AI". Its definition covers digital media created algorithmically with an AI model trained on captured content. The user guide gives text-to-image output from a prompt as an example.
- compositeWithTrainedAlgorithmicMedia. IPTC's label is "Edited using Generative AI". It covers augmentation, correction or enhancement with a generative model, such as inpainting or outpainting. The user guide's example is a photo changed with a generative fill tool.
A third value, compositeSynthetic, covers a mix of several elements where at least one is generative. The three share one idea: the file says generative AI had a part in it. They differ in how it took part: creating the whole image, editing it, or being one element in a mix.
What do digitalCapture and the other values mean?
The table lists the 17 current terms as IPTC's page showed them on October 30, 2026. The last column says how the New Watermark checker groups each one.
| Value | IPTC's meaning in short | Checker groups it as |
|---|---|---|
| digitalCapture | Captured from real life with a digital camera or recording device. | Another source |
| computationalCapture | Several real-life frames merged automatically, as in smartphone HDR. | Another source |
| negativeFilm | Digitised from a transparent negative. | Another source |
| positiveFilm | Digitised from a transparent positive. | Another source |
| Digitised from a non-transparent medium such as a photographic print. | Another source | |
| humanEdits | Correction or enhancement by people using non-generative tools. | Another source |
| compositeCapture | A mix of several elements that are all captures of real life. | Another source |
| digitalCreation | Made by a person with non-generative digital tools. | Another source |
| dataDrivenMedia | Digital media that represents data, made through human programming or creativity. | Another source |
| screenCapture | A capture of what is on a computer or phone screen. | Another source |
| algorithmicallyEnhanced | An algorithm fixes or refines the media without changing its main content, as sharpening does. | Another source |
| algorithmicMedia | Made purely by a formula-driven algorithm, with no sampled training data. | Another source |
| composite | A mix of several elements, any of which may or may not be generative. | Another source |
| virtualRecording | A live recording of a virtual event based on generative AI and/or captured elements. | Another source |
| trainedAlgorithmicMedia | Created by an AI model trained on captured content. | AI declaration |
| compositeWithTrainedAlgorithmicMedia | Edited with a generative AI model, as in inpainting. | AI declaration |
| compositeSynthetic | A mix of elements, at least one of them generative. | AI declaration |
The vocabulary shows 20 terms in all. Three are retired: softwareImage (June 14, 2022), minorHumanEdits and digitalArt (both September 17, 2024). IPTC says to use humanEdits and digitalCreation in place of the last two. Older files can still carry a retired value, and our checker marks it as retired.
What did we find in real files?
We ran 31 files that public collections and vendor samples present as AI-made through our checker on October 27, 2026. Wikimedia Commons volunteers uploaded most of them, so we cannot know how each was handled before upload. Nine of the 31 declare an IPTC digital source type. Seven declare trainedAlgorithmicMedia and one declares compositeWithTrainedAlgorithmicMedia, all inside signed Content Credentials records. One declares trainedAlgorithmicMedia in unsigned XMP. The other 22 declare none.
Two details show how untidy real files are. First, the address form varies: one Adobe file signs https://cv.iptc.org/... and the others use http://. The checker accepts both. Second, in two older ChatGPT files the value sits in the record of an earlier version, called an ingredient, and not in the active record. The tables for every file that carries a record are on our accuracy page.
A value is a claim by whoever signed or wrote it. In our files made to test forged records, a Midjourney-filed image re-signed with a self-made certificate carries a signed digitalCapture claim and a certificate that names a camera maker. The checker reads it as "Content Credentials found; signer trust not established".
Why does Google Merchant Center care about the value?
Google's Merchant Center help page on AI-generated content says images created with generative AI must contain metadata indicating that, using the IPTC DigitalSourceType trainedAlgorithmicMedia tag. It also says not to remove embedded tags such as the IPTC DigitalSourceType property. An earlier help page dated February 2024 names the trainedAlgorithmicMedia tag. This is Google's rule for sellers, not a rule for everyone. Google Merchant Center and AI-generated images covers it in full.
What this means when you check an image
Open the checker and choose the file. The headline tells you how the value arrived.
- "Metadata declares AI use". Unsigned metadata carries a value that the checker treats as generative: trainedAlgorithmicMedia, compositeWithTrainedAlgorithmicMedia or compositeSynthetic. The detail says anyone can add or remove it, so it is a label, not proof.
- "Metadata declares another source". Unsigned metadata carries a value such as digitalCapture. It does not say generative AI was used, and it may be out of date.
- "Trusted Content Credentials declare AI use". A signer on the current C2PA Trust List signed a record whose source type is one of those three. That shows who signed and that the bytes match. It does not show the claim is true.
No value found tells you little. In our set, 22 of 31 files carried no IPTC value, and sharing an image often removes metadata. Nothing found does not mean an image is real or human-made. For where else a declaration can sit, see where AI declarations live in an image file.
Where to read more
This guide belongs to How to tell if an image is AI. Related guides: Content Credentials versus EXIF, IPTC and XMP and can EXIF data be faked.
Frequently asked questions
- What does trainedAlgorithmicMedia mean?
- It is IPTC's term for digital media created algorithmically with an AI model trained on captured content. IPTC labels it "Created using Generative AI".
- Does a trainedAlgorithmicMedia value prove an image was made by AI?
- No. It is what the file says about itself. Unsigned metadata can be added or removed by anyone, and a signed record shows only who signed it and that the bytes match.
- Is Digital Source Type the same as Content Credentials?
- No. It is one field that can sit in unsigned XMP or inside a signed Content Credentials record. The signed record adds a signature and a signer.
- Why do some files use https and others http in the address?
- We saw both in real files. IPTC lists its addresses with http. Our checker accepts both forms.