One metadata field now states whether a file came from a camera, a strip of film or a generative model. It has existed since 2008, and most archives have never filled it in.
The field that marks a picture as machine-made also marks it as a transparency; an archive that fills in neither has declined to say which.
There is a single metadata field whose job is to say what a digital image was made from: a camera, a negative, a transparency, a print, a piece of software or a generative model. It is called Digital Source Type, it belongs to the IPTC Photo Metadata Standard, and its film values have been available since 2008. The angle of this article is that the recent attention on labelling AI images has, almost by accident, handed film archives a tool of their own. The vocabulary that carries "created using generative AI" also carries "digitised from a transparent positive". Each value means more because the other exists.
The vocabulary and Google's documentation were checked on 18 November 2026.
What is IPTC Digital Source Type?
It is one property in the IPTC standard, stored in a file's XMP metadata as Iptc4xmpExt:DigitalSourceType. It does not take free text. Its value is a term from a controlled vocabulary that the IPTC maintains as one of its NewsCodes, written in full as a web address such as http://cv.iptc.org/newscodes/digitalsourcetype/positiveFilm. Using a controlled term is the point: software can act on it without having to interpret a caption.
The field answers one question, where the pixels came from. It does not say who owns the image, what it shows or whether it may be used. Those belong to other fields, which we cover in IPTC fields that matter for a photo archive.
What are the Digital Source Type values?
The published vocabulary contains 20 terms. Three are retired, which leaves 17 in use. The dates below are those shown in the vocabulary itself.
| Value | IPTC name | Use it for | Added |
|---|---|---|---|
| digitalCapture | Digital capture sampled from real life | A photograph from a digital camera | 2008 |
| computationalCapture | Multi-frame computational capture sampled from real life | Frames merged automatically, such as smartphone HDR | 17 Sep 2024 |
| negativeFilm | Digitised from a transparent negative | A scan of a negative | 2008 |
| positiveFilm | Digitised from a transparent positive | A scan of a transparency or slide | 2008 |
| Digitised from a non-transparent medium | A scan of a print or other opaque original | 2008 | |
| humanEdits | Human-edited media | Changes made by a person with non-generative tools | 17 Sep 2024 |
| algorithmicallyEnhanced | Algorithmically-altered media | Sharpening, noise reduction and the like, leaving the main content unchanged | 14 Jun 2022 |
| compositeWithTrainedAlgorithmicMedia | Edited using Generative AI | Inpainting, outpainting and other generative edits to an existing image | 26 Jul 2023 |
| trainedAlgorithmicMedia | Created using Generative AI | An image generated by a model trained on captured content | 14 Jun 2022 |
| algorithmicMedia | Pure algorithmic media | An image made by a formula, with no training data | 14 Jun 2022 |
| digitalCreation | Digital creation | Work made by a person with non-generative digital tools | 17 Sep 2024 |
| dataDrivenMedia | Data-driven media | Charts and other representations of data | 14 Jun 2022 |
| screenCapture | Screen capture | A screenshot | 17 Sep 2024 |
| virtualRecording | Virtual event recording | A recording of a virtual event | 14 Jun 2022 |
| composite | Composite of elements | A mix of elements, any of which may or may not be generative | 17 Sep 2024 |
| compositeCapture | Composite of captured elements | A composite in which every element is a capture of real life | 14 Jun 2022 |
| compositeSynthetic | Composite including generative AI elements | A composite with at least one generative element | 14 Jun 2022 |
The three retired terms still turn up in older files: softwareImage (2008, retired on 14 June 2022 in favour of more specific terms), minorHumanEdits (replaced by humanEdits on 17 September 2024) and digitalArt (replaced by digitalCreation on the same day).
The history is legible in the dates. Five terms in 2008, for a world of cameras and scanners. A burst in June 2022 as image generators arrived. Refinements in 2023 and 2024 as it became clear that the hard cases are mixtures. The addition of computationalCapture is a quiet admission that even a telephone photograph is now several exposures merged by software; World Press Photo's contest rules go further and exclude smartphone HDR and portrait modes altogether.
Which value should a film scan carry?
The definitions settle most cases, because they describe the original medium and not the equipment.
| Original | Value | Note |
|---|---|---|
| Colour transparency or 35mm slide | positiveFilm | Whether scanned on a film scanner, a drum scanner or with a digital camera on a copy stand |
| Black-and-white or colour negative | negativeFilm | The inversion to a positive is part of digitising, not a separate source |
| Photographic print | ||
| Polaroid print | An opaque original | |
| Contact sheet | The sheet is a print; the negatives it was made from are a separate matter | |
| Magazine page or tear sheet | The right value says nothing about who owns what is on the page | |
| Camera raw file and its exports | digitalCapture |
The note in the first row matters for anyone camera scanning film. The capture device is a digital camera, but the image is "digitised from a transparent positive", so positiveFilm is correct and digitalCapture is not. A file labelled digitalCapture claims to be a photograph of the world. A camera scan is a photograph of a photograph.
What about edits, composites and restorations?
Here the vocabulary gives the words but not always the rule, and what follows is our house reading, not IPTC guidance. The property holds one value, so a choice has to be made when a file is both a scan and something else.
- Master scan, with dust removal and colour balance: keep negativeFilm, positiveFilm or print. Routine digitisation work does not change what the file is.
- Restored derivative, using only non-generative tools: still the film value in our catalogue, with the restoration described in the caption and edit history. Some organisations will prefer algorithmicallyEnhanced or humanEdits; the important thing is to decide once and apply it throughout.
- Any derivative touched by a generative tool (AI upscaling, colourising, infill): compositeWithTrainedAlgorithmicMedia, without exception, and a separate file from the master. The reasoning is in AI colourising and upscaling of archive photographs.
- A montage of several scans: compositeCapture.
- A wholly generated image: trainedAlgorithmicMedia. It should never share a folder, a naming pattern or an identifier series with scans.
Who reads the field?
Google's documentation on image metadata lists four Digital Source Type values that it supports: trainedAlgorithmicMedia, compositeSynthetic, algorithmicMedia and compositeWithTrainedAlgorithmicMedia. All four concern synthetic content. The same page is careful about what follows from that. It says Google does not guarantee that IPTC metadata will appear in search results, and it attributes the AI information in "About this image" to C2PA metadata signed by a certificate on the C2PA Trust List.
Marketplaces run parallel systems of their own. Adobe Stock, for instance, requires contributors to tick "Created using generative AI tools" at upload, a platform flag that does the same job as the metadata value. The policies are compared in do picture agencies accept AI-generated images?.
For film values there is, so far as we can find, no public platform that displays them. Their readers are picture desks, asset management systems and, increasingly, the people who assemble and audit datasets. A dataset card that says "scanned from film" is a statement about the collection. The same statement inside each file is a property a pipeline can filter on.
How do you write it, and what are the limits?
Most cataloguing and asset management software exposes the field, and command-line metadata tools can write it in bulk. Set it at the point of scanning, as part of the vendor's delivery specification, and you never have to think about it again. Our own copyright metadata tool writes the rights fields, creator, copyright notice, credit line, usage terms and licensor URL, into a JPEG without recompressing it. It does not write Digital Source Type, and the two jobs are best seen as a pair: one says whose the picture is, the other says what it is. The metadata viewer will show what a file currently carries.
Three limits should temper enthusiasm.
- It is a declaration. Anyone can type positiveFilm into a generated image. The field is as reliable as its author.
- It is fragile. Many platforms strip embedded metadata on upload, so absence means nothing.
- It is single-valued. A file with a complicated history gets one word. The detail belongs in the catalogue.
The remedy for the first two is to hold the value in more than one place: in the file, in the catalogue record, and where the tooling exists, inside a signed manifest, since Content Credentials can carry the same vocabulary. We explain that layer in Content Credentials for photo archives.
How this applies to the Robert Harper archive
The Robert Harper Photograph Library's inventory gives its makeup by medium: 52,684 unique analogue photographs, principally 6 x 7 colour transparencies with black-and-white negatives, contact sheets, 35mm slides and 4,005 Polaroid prints, and 42,803 digital photographs held as Nikon NEF raw files, TIFFs and JPEGs. Mapped to the vocabulary, that is mostly positiveFilm, with negativeFilm, print and digitalCapture making up the remainder, and no synthetic values at all. The counts are the inventory's own and are not yet independently reconciled.
More than half of the archive is still on film and has yet to be scanned, which means the field can be set correctly for every file at the moment it is made. That requirement, along with identifiers and hashes, goes into the brief for any scanning work. Developers and dataset builders who need source type as a filterable, per-file property can raise it when discussing AI and data licensing, which is available by negotiation and subject to rights review.
Frequently asked questions
What is IPTC Digital Source Type?
It is a property in the IPTC Photo Metadata Standard that records what kind of source a digital image was created from. Its value is a term from an IPTC controlled vocabulary, such as digitalCapture for a photograph from a digital camera, positiveFilm for a scan of a transparency, or trainedAlgorithmicMedia for an image created with generative AI.
Which Digital Source Type should I use for an AI-generated image?
Use trainedAlgorithmicMedia for an image created wholly by a generative model. Use compositeWithTrainedAlgorithmicMedia where a real image has been edited with generative AI, for example inpainting or outpainting, and compositeSynthetic for a composite in which at least one element is generative. All three are among the values Google documents as supported.
Which Digital Source Type should I use for scanned film?
Use negativeFilm for a file digitised from a negative and positiveFilm for one digitised from a transparency or slide. Use print for anything digitised from an opaque original, including photographic prints, Polaroids, contact sheets and printed pages. The value describes the original medium, not the scanning device.
Does Google read the Digital Source Type field?
Google's documentation on image metadata lists four Digital Source Type values that it supports, all concerning synthetic or partly synthetic media. The same page states that Google does not guarantee that IPTC metadata will be shown in search results, and describes labels in its About this image feature as coming from C2PA metadata.
Can the Digital Source Type field be trusted?
Only as far as you trust whoever filled it in. It is ordinary embedded metadata: unsigned, easy to edit and often removed by platforms. Its value lies in honest parties declaring consistently. For stronger assurance the same vocabulary can be carried inside a signed Content Credential, and an archive should keep the value in its catalogue as well.
Sources
- IPTC NewsCodes: Digital Source Type vocabulary (terms, definitions, created and retired dates)
- Google Search Central: Image metadata in Google Images
- Adobe Stock: Generative AI content submission guidelines (updated 11 June 2026)
- World Press Photo: What counts as manipulation (2027 contest)
- IPTC Photo Metadata Standard specification
Published by Archive.London. Corrections are welcome under our editorial policy.

