AI training data and image licensing

How to write a dataset card for a photographic archive

A Chanel No. 19 spray bottle with a gold top and bubbles in the liquid.
Chanel No. 19 voile parfumé.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

A dataset card is two or three pages that tell a buyer what an image collection is, where it came from and what it is not. Here is how to write one for boxes of film.

In March 2018 a group of researchers led by Timnit Gebru posted a paper to arXiv with a deliberately plain title, "Datasheets for Datasets". Its argument was borrowed from electronics: every component ships with a datasheet describing its operating characteristics, so every dataset should ship with a document describing its purpose, composition, collection process and recommended uses. Eight years on, that idea is the first thing a data buyer asks an archive for. It is usually called a dataset card.

This guide is for archive owners who need to write one, and for buyers who want to know what a good one looks like when the "dataset" is a room of transparencies, negatives and prints.

This article is general information, not legal advice. Checked on 17 November 2026.

What is a dataset card, and where did the format come from?

Three sources shaped the format. The datasheet paper grouped its questions under motivation, composition, collection process, preprocessing, uses, distribution and maintenance. In April 2022 three Google researchers proposed "Data Cards", described as structured summaries of essential facts about a dataset, written for the range of people who touch it across its life. And dataset hosting platforms such as Hugging Face made a card the default front page of every dataset repository, which turned a research proposal into a habit.

Regulation has since given the habit a commercial edge. Under Article 53 of the EU AI Act, providers of general-purpose models must publish a summary of training content on a Commission template, published on 24 July 2025. A developer filling in that template needs facts from its suppliers. A supplier that hands over a clear card has done part of the buyer's paperwork.

Why a photographic archive needs its own version

The research templates assume the data already exists as files. An archive is different in four ways, and each needs its own section.

State. Some of the collection is digital and some is physical. A card that gives one headline number without splitting it misleads.

Counting. Archives count frames, sheets, rolls, boxes and "images", and the same picture can exist as a transparency, a duplicate, a print and a magazine tear sheet. The card must say which unit it is using.

Description. Cataloguing in a working photographer's archive is usually at job or box level. Item-level captions may not exist yet.

Rights. A photograph's copyright is separate from rights in what it shows. The card should describe the review process, not simply assert an outcome.

What sections should a photography dataset card contain?

The table below is a working template. The right-hand column shows how the Robert Harper Photograph Library fills each line, using only figures from the archive's own inventory.

Section What to state Example: Robert Harper Photograph Library
Identity Name, owner, contact Robert Harper Photograph Library; 825 Ventures Ltd, London, owner and rights holder; info@archive.london
Provenance Creator, period, how the work was made One photographer; about thirty-five years of professional work, from the 1960s/1970s to the digital era; editorial and commercial commissions
Size Unique images, and the counting basis 95,487 unique photographs; 177,143 frames including duplicates and variants
State Digital against analogue 42,803 digital; 52,684 analogue, largely awaiting digitisation
Formats Physical and file formats Mostly medium-format colour transparencies; also black-and-white negatives, contact sheets, larger-format film, 35mm slides, more than 4,000 Polaroid prints, magazine pages
Subjects Categories, with overlaps noted 330 described collections across 24 principal categories; collections may overlap
Description Level of captioning and keywording Collection-level descriptions online; item-level cataloguing is part of the digitisation programme
Rights position Who licenses; what review is done Single rights holder; licences by negotiation, subject to rights review; no clearance claims
Exclusions Categories removed or flagged Portraits and identifiable people, branded products and magazine pages reviewed case by case and excludable
Limits What is unknown or unverified Counts not yet checked frame by frame against the physical material; analogue condition and scan specification to be confirmed per batch
Access How to evaluate Sample set under NDA; first-look arrangements as scanning proceeds

Two lines deserve a closer look.

Size and state. Give the unique count and the gross count, and say which is which. A buyer who discovers after signing that a third of the "images" are variants of the same frame will not come back. Then split digital from analogue. For film, say what scanning will produce: the film scan resolution calculator converts a format and a sampling rate into pixel dimensions, which is the figure an ML team actually wants.

Limits. This is the section owners are tempted to cut and buyers read first. The datasheet authors asked dataset creators to record errors, noise and redundancies; the archive equivalent is to say that the inventory has not been checked frame by frame against the physical material, that collections overlap, that dates are known at job level and not always at frame level. None of this weakens the offer. A reviewer who finds the caveats stated in the card stops looking for the ones that were hidden.

How to describe rights without overclaiming

The rights section is where cards most often go wrong, in one of two directions. Some say nothing, which forces the buyer's counsel to assume the worst. Others use words such as "cleared" or "safe", which are promises about third parties that an archive owner is rarely in a position to make.

The better course is to describe three things factually. First, who owns the copyright in the photographs and will sign the licence. Second, what the review process examines: identifiable people, trade marks and product designs, third-party artwork, the terms of original commissions where known. Third, what happens to material that raises a question: flagged, held back, or excluded from the licensed set. Our article on what a rights review covers before images are licensed for AI sets out the categories.

It also helps to say what machine-readable signals the files carry. The IPTC Photo Metadata Standard has included a "Data Mining" property since version 2023.1, approved on 4 October 2023, with values that allow or prohibit AI and machine-learning use. A card can state which value is embedded in published files and that licensed deliveries are governed by the contract instead.

Common mistakes

  • One big number. A total with no split by state, format or uniqueness.
  • Adjectives in place of facts. "High quality" tells a buyer nothing; pixel dimensions, bit depth and colour space do.
  • Silent exclusions. Removing sensitive categories without saying so makes the composition look different from what will be delivered.
  • No date. A card describes a moving target, especially while scanning is under way. Date it and version it.
  • Performance claims. A card describes data. It should not predict what a model will do with it.

Keeping the card honest as scanning proceeds

For an archive that is still being digitised, the card is a living document. Each scanned batch changes the digital count, adds real technical specifications and usually corrects the inventory. Publishing the running figures on a public archive facts page and matching the card to them keeps both consistent, and lets anyone quoting the archive use the same numbers.

Buyers assess a card alongside the questions in what AI developers look for in a licensed image dataset, so it is worth writing the two against each other.

The Robert Harper Photograph Library's card

The archive's own dataset card is now published, alongside the complete 503-folder inventory it is drawn from.

The example column above is not hypothetical. Those are the facts on Archive.London's own fact sheet, with the same caveats, for a collection that runs from medium-format transparencies and other analogue originals to digital files. The card is a description and not a warranty; licences are negotiated directly with the single rights holder, subject to rights review.

If you would like the current version, with a sample set under NDA, you can request the dataset card through the AI training data licensing page.

Frequently asked questions

What is a dataset card?

A dataset card is a short structured document describing a dataset for the people who might use it: what it contains, who made it and why, how it was collected and processed, what it is suitable for, what its known limits are, and on what terms it is available. The format grew out of the 2018 research paper 'Datasheets for Datasets' and is now standard on dataset hosting platforms.

What should a dataset card for images include?

At minimum: owner and contact; creator and period; counts by format and by digital or analogue state; subject categories; technical specification of files; level of captions and metadata; rights position and what review has been done; categories excluded; known gaps and biases; intended and unsuitable uses; and how to obtain a sample. Every number should say where it comes from and how reliable it is.

No. It is a description, not a licence or a warranty, and it should say so. The licence agreement governs what a buyer may do. The card's job is to let technical and legal reviewers decide quickly whether a collection is worth a closer look, and to give them accurate facts to quote internally. Anything it states should be something the owner can stand behind.

How long should a dataset card be?

Two to four pages is usual for a first-contact card. The original datasheet proposal lists several dozen questions, and a full answer set can run much longer, but a procurement reader needs the essentials first: counts, formats, provenance, rights position and limits. Longer technical appendices, such as a file manifest or metadata schema, can follow under NDA.

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