AI training data and image licensing

Image provenance for AI: datasets, model cards and audit trails

A sparkling necklace of linked flower motifs drapes across textured tree bark.
Floral necklace on bark.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

Ask where a model's training images came from and the answer is often a shrug in several layers. Provenance is the paperwork that replaces the shrug, and it starts with whoever supplies the pictures.

In 2023 a group of researchers calling themselves the Data Provenance Initiative audited more than 1,800 text datasets used in AI. On the popular sites where such datasets are hosted, they found licence information missing for over 70% and wrongly categorised in over half. Their audit concerned text, not pictures, but it describes the general condition well: a great deal of training data arrives with no dependable account of where it came from or what may be done with it.

Provenance is the remedy, and for image data it begins with the supplier.

This article is general information, not legal advice. Checked on 18 November 2026.

What does provenance mean for training data?

In the art trade, provenance is the chain of ownership of an object. For training data the idea is wider. It covers four things about every item in a dataset:

  • Origin. Who made it, when and how.
  • Rights. Who owns it, and on what terms the developer holds it.
  • Custody. The route by which it reached the training set: from the owner directly, through an aggregator, or from a crawl.
  • Transformation. What was done to it on the way: scanning, resizing, cropping, captioning, filtering.

For a photograph the first two overlap with what picture buyers have always asked, as set out in provenance for photographs: what buyers ask for. The last two are specific to data, and they are where records most often run out.

What are dataset cards and model cards?

Two papers from 2018 set the pattern. "Datasheets for Datasets", by Timnit Gebru and colleagues, first posted on 23 March 2018, proposed that every dataset be accompanied by a document recording its purpose, composition, collection process and recommended uses, on the model of the datasheets that accompany electronic components. "Model Cards for Model Reporting", by Margaret Mitchell and colleagues, submitted on 5 October 2018, proposed a short document for each trained model stating its intended uses, its measured performance and how it was evaluated.

The two are links in one chain:

Document Written by Describes Feeds into
Dataset card or datasheet The data supplier One collection: contents, origin, rights, limits The developer's records
Model card The developer One model: uses, performance, evaluation, data used Customers and auditors
Public training content summary The developer All sources used for a general-purpose model, in categories Regulators and the public

A developer cannot write the second and third honestly if the first does not exist. That is the supplier's practical interest in documentation, and the reason for the guide to writing a dataset card for a photographic archive.

What do the disclosure rules require?

Decided. In the EU, Article 53 of the AI Act requires providers of general-purpose AI models to publish a sufficiently detailed summary of the content used for training, on a template the Commission published on 24 July 2025. The obligations have applied since 2 August 2025; models already on the market before that date have until 2 August 2027. The Commission's enforcement powers began on 2 August 2026. The template's data sources section covers publicly available datasets, private datasets, data scraped from online sources, user data and synthetic data. For scraped data it asks for the crawlers used, the collection period and a list of the top 10% of domains scraped, with a lower threshold for small and medium-sized providers. It is a summary: no itemised list of works is required.

In California, AB 2013 has required developers of generative AI systems to post documentation about their training datasets since 1 January 2026. A federal court refused xAI's request for a preliminary injunction against the law on 4 March 2026, as reported by PPC Land; we have not confirmed what has happened in that case since.

In the UK there is no statutory duty. The government's report of 18 March 2026 said it would work with industry to explore good practice on transparency and did not propose legislation, as summarised by Reed Smith.

Pending. How strictly the EU summaries will be policed is not yet known; no enforcement action could be confirmed on the date above. In the United States, federal proposals that would give rights holders a route to discover whether their works were used, such as the TRAIN Act, have been introduced and not enacted. The detail of the European rules is in the EU AI Act's training data summary rules.

What goes wrong without provenance?

Three well-documented episodes show the pattern.

Unknown contents. LAION-5B, an index of several billion image links and captions, was withdrawn after a Stanford Internet Observatory report published in December 2023 identified links to suspected child sexual abuse material. A revised version was released on 30 August 2024 with 2,236 links removed. In June 2024 Human Rights Watch reported finding photographs of identifiable children among the links. LAION points out that its datasets contain links and metadata, not the images. The difficulty is the same either way: nobody had reviewed, item by item, what the index pointed to.

Contamination. Bloomberg reported in April 2024 that about 5% of the images used to train Adobe's Firefly were AI-generated, having entered through Adobe's stock library. A dataset described at the level of "our licensed stock collection" did not reveal what individual contributors had uploaded.

Contracts read later. Old agreements are being interpreted for a use nobody had in mind when they were signed, which is why a chain of title has to include the documents and not only the assertion. That is covered in the checklist for preparing a photo archive for AI licensing.

How do you build an audit trail for images?

An audit trail is provenance at the level of the individual file. It needs surprisingly few fields, kept consistently.

Field Purpose
Stable identifier Names the image, independent of filename
SHA-256 hash of the delivered file Proves exactly which file was supplied
Perceptual hash Finds near-duplicates and resized copies
Source reference Points to the original: negative, transparency, raw file, with its inventory number
Creator and rights holder Who made it, who licenses it
Rights category and review date What was checked about people, brands and third-party content, and when
Caption source Whether description came from records, a cataloguer or a machine
Licence reference The agreement under which this file was delivered
Version and change log What was added, corrected or withdrawn, and when

Three practices make the table work.

Hash at delivery. Both sides keep the manifest. Any file can be hashed again later and checked against it. The image hash and fingerprint tool shows both kinds of hash for a single file; in production the same calculation runs in a script.

Record removals. If an image is withdrawn after delivery, the log should show when and why, and the licence should say what the developer must do about it.

Keep the physical link for analogue material. A scan's provenance runs back to a piece of film in a numbered sleeve. Recording that number in the file ties the digital object to something that can be inspected.

Content Credentials can sit alongside this. A C2PA manifest is a signed record of how a file was made and edited, and the Content Credentials checker shows whether a file carries one. It is evidence about the asset. It does not state rights and is not a substitute for the manifest, as explained in Content Credentials for photo archives.

What should a buyer ask a supplier?

  1. Who created the images, and how do you know?
  2. Who owns the rights, and what documents show it?
  3. Has any of this material been licensed to stock libraries or aggregators before, and on what terms?
  4. How did the files come to exist: camera originals, scans from film, copies of copies?
  5. What rights review has been done, by category, and when?
  6. Where did the captions come from?
  7. Can you supply a manifest with identifiers and hashes?
  8. Does the collection contain any AI-generated or AI-altered images?
  9. What happens if an image has to be withdrawn?
  10. Is there a dataset card, and what does it say the collection is not suitable for?

A supplier who says plainly which of these it cannot yet answer is more useful than one who answers all ten with adjectives.

Provenance at the Robert Harper archive

The Robert Harper Photograph Library starts from a simple chain: one photographer, whose career is documented, and one owner, 825 Ventures Ltd, the owner and rights holder. The public dataset card describes the collection and its limits, and the archive inventory lists all 503 folders, from jewellery and fragrance to corporate and technology subjects. By inventory count there are 95,487 unique photographs, a figure to which the 503 inventory lines add up exactly, though it has not been checked frame by frame against the physical material.

What lies ahead is the file-level trail described above. More than half of the archive is still on film and has yet to be scanned, so identifiers, hashes, caption sources and rights categories remain to be recorded as digitisation proceeds; the dataset card states that item-level captions and keywords are not yet available. The archive's position on third-party rights is set out under rights and stewardship. Licences are negotiated directly with the single rights holder, subject to rights review, and data teams who want to see the documentation can begin at the AI training data licensing page.

Frequently asked questions

What is training data provenance?

Training data provenance is the documented history of the data used to build a model: where each item came from, who created and owns it, under what terms it was obtained, and what was done to it before training. For images it links a file in a training set back to a photographer, a rights holder and a licence.

What is the difference between a dataset card and a model card?

A dataset card describes a collection of data: its purpose, contents, how it was gathered, its rights position and its limits. A model card describes a trained model: its intended uses, how it performs and how it was evaluated. A model card should refer to the datasets used, so good dataset documentation is what allows a model card to be complete.

Does the EU AI Act require AI companies to list every image they trained on?

No. Article 53 requires providers of general-purpose models to publish a sufficiently detailed summary of training content using the Commission's template. The template asks for categories of source, lists of large public datasets and the most-scraped domains. It does not require an itemised list of individual works. This is general information, not legal advice.

How do you prove which images were licensed for AI training?

By agreeing a manifest at delivery: a list giving each file's identifier and cryptographic hash, with the licence reference and date. Both parties keep a copy. If a question arises later, any file can be hashed again and checked against the manifest, which shows whether it was part of the licensed set.

Do Content Credentials prove an image is licensed for training?

No. Content Credentials, the C2PA standard, record how a file was created and edited, in a signed manifest. They say nothing about who may use the image or for what. They can support a provenance record, for instance by showing that a file came from a particular camera or was generated by AI, but rights are documented separately.

Sources

This article is general information, not legal advice. It reflects the position as we understood it on the date shown and may be out of date. Published by Archive.London. Corrections are welcome under our editorial policy.

Licensing the archive for AI and data projects

95,487 photographs from one photographer and one rights holder. Read the dataset card, search the 503-folder inventory, discuss a sample set under NDA, or arrange a first look as scanning proceeds.

See AI and data licensingWrite to the archive