A vision model will tell you a photograph shows a ring on black velvet. It will not tell you who commissioned it, when, or for which magazine, and it may invent an answer if asked.
Show a current vision model a studio photograph of a diamond ring and it will say, correctly, that this is a ring, that the stone is cut and clear, that the background is dark and the lighting controlled. Ask it who commissioned the picture, in which year, and where it was published, and it has nothing to go on. Those facts are not in the pixels. A well-behaved system will say so. A less careful one will produce a confident sentence anyway.
That gap is the whole subject. Automated tagging is good at what can be seen and poor at what has to be known, and an archive's value lies mostly in the second category.
What kinds of AI tagging are there?
Three different technologies are sold under one label, and they fail in different ways.
Classifiers and object detectors assign terms from a fixed list, usually with a confidence score. They are predictable and limited to what is on the list.
Embedding models turn each image into a numerical vector so that pictures can be searched by a text phrase or by similarity to another picture. Nothing is written into the record. This is often called semantic search.
Vision-language models write free-text captions and keyword lists on request, and can follow instructions such as "use only terms from this vocabulary". They are the most flexible and the most prone to stating things that are not so.
A given product may combine all three. It is worth asking a vendor which is doing what.
What does auto-tagging get right?
- Generic visible content. Object types, animals, vehicles, food, clothing categories, interiors against exteriors, landscape types.
- Visual properties. Dominant colours, black-and-white against colour, close-up against wide, plain against busy backgrounds.
- Genre. Portrait, still life, street scene, aerial.
- Legible text. Printed lettering on packaging, signage or a slide mount can often be read. Handwriting on sleeves and contact sheets is far less dependable.
- Grouping. Clustering frames from the same set-up, and flagging near-duplicates and variants. An inventory that counts 95,487 unique photographs against 177,143 frames including duplicates and variants, as the Robert Harper Photograph Library's does, shows how much of an archive's bulk this can be. For checking a handful of files by hand, the free image hash and fingerprint tool compares two images' perceptual hashes in the browser.
For a collection of controlled product photography, this first layer is not trivial. Getting "watch", "strap", "steel" and "white background" on to thousands of records in an afternoon is a real saving, provided those terms are then mapped to the archive's own list.
What does it get wrong?
Identity. General-purpose models are not a safe means of naming people. Many are designed to decline, and where one does offer a name it is a resemblance, not an identification. Naming a person in a record is a factual claim with consequences for rights and reputation, and it should rest on the photographer's notes or other documentary evidence.
Date. A model can guess a decade from clothes, cars or print quality. A guess is not a date. Film scans add a complication: the colour rendering of an older transparency can itself push a model towards "vintage" whatever the content.
Place. Famous landmarks are recognised. An unremarkable street, a studio or a hillside is not.
Specific objects. A system trained largely on recent web images may name a current product when shown an older one, or fall back on a vague term. Period technology is a good test. The archive's collection of telephones, mobiles, pagers and fax machines is the kind of subject where "mobile phone" is easy and the exact model is not.
Context. Client, publication, purpose, whether a frame was the selected one. None of it is visible.
Film-specific material. An uninverted negative, a contact sheet of twelve frames, a transparency photographed in its sleeve: each can produce nonsense if fed to a tool that expects a finished picture.
Fabrication. Caption models generate plausible text. Asked for a detailed description, they may supply a brand, a location or an action that is not there. The fluency of the sentence is no guide to its truth.
| Task | Machine alone | With human review |
|---|---|---|
| Generic subject and setting terms | Usually sound | Spot-check a sample |
| Colour, composition, genre | Usually sound | Spot-check a sample |
| Near-duplicate grouping | Good | Confirm before merging records |
| Reading printed text in the image | Often good | Verify anything used as a fact |
| Reading handwriting on sleeves | Unreliable | Transcribe by hand, or check every line |
| Brand and exact product | Hit and miss | Confirm from records |
| Named people | Not appropriate | Documentary evidence only |
| Date and place | Guesswork | Documentary evidence only |
| Caption sentences | Draft only | Rewrite to what is evidenced |
Should you use facial recognition on an archive?
Only after taking advice. Under the UK GDPR, biometric data processed for the purpose of uniquely identifying a living person is special category data, which requires a lawful basis and a further specific condition. Running face matching across an archive of portraits and street scenes to attach names may fall within that. There is also a plain accuracy problem, since a false match puts the wrong name on a picture offered for publication. This is general information, not legal advice.
What should you check before uploading unpublished scans?
Many tagging services run in a provider's cloud. For an archive whose appeal includes material that has not been published online, the terms on which images leave the building matter as much as the quality of the tags.
Ask, and get the answers in writing:
- Are images retained after processing, and for how long?
- Are they, or data derived from them, used to train or improve the provider's models?
- Where is processing carried out, and which sub-processors are involved?
- Can the service be run on the archive's own hardware instead?
The second question has a commercial edge. An owner who intends to license images as AI training data, by negotiation and subject to rights review, should not find that a tagging contract has granted similar use as a side effect. The principles behind that position are set out under rights and stewardship.
A workflow that keeps people in charge
- Pilot first. Take a few hundred images spread across every format and subject area. Have a cataloguer keyword them blind, run the tool, and score the tool's output term by term: correct, wrong, or too vague to be useful.
- Constrain the output. Map machine labels to the archive's controlled vocabulary. Discard labels with no mapping.
- Keep suggestions separate. Store machine terms in their own database field with the tool name, version, date and confidence. Promote them to the public record only after review. The embedded keyword list has no widely used flag for "machine-suggested", so that distinction has to live in the catalogue.
- Never automate names, dates or places. These come from sleeves, contact sheets and job records, as described in writing captions from contact sheets and sleeve notes.
- Audit continuously. Review a random sample of accepted terms each month and track the error rate by term type.
- Verify the export. Open delivered files in the EXIF and metadata viewer to see exactly which terms ended up embedded.
Used this way, automation does the repetitive description and people do the knowing. Used the other way round, it fills a catalogue with confident errors that are expensive to find later.
An open question for this archive
For the Robert Harper Photograph Library the question is live, not historical. Its 52,684 analogue images, by the archive's inventory count, largely await both scanning and metadata cataloguing, and the archive is actively seeking partners for that work. Vendors of tagging, search and cataloguing software who would like to propose a measured pilot on sample material are invited to read the digitisation partners page and get in touch.
Frequently asked questions
Can AI automatically keyword a photo archive?
It can supply a useful first layer: generic subject terms, settings, colours and composition, applied quickly to large numbers of images. It cannot supply names, dates, locations or commissioning context unless that information is visible or supplied from another source. Most archives use it to draft terms that a cataloguer then confirms against a controlled vocabulary.
How accurate is AI image tagging?
It varies by term type more than by product. Broad categories such as 'shoe', 'beach' or 'portrait' are usually right. Specific models, brands, periods and identities are frequently wrong or missing. Published accuracy figures are measured on benchmark photographs that rarely resemble film scans, so the only meaningful figure is one measured on a sample of your own material.
Is it legal to use facial recognition on an archive in the UK?
It needs care. Under the UK GDPR, biometric data processed to identify a living person uniquely is special category data and requires both a lawful basis and a separate condition for processing. Matching faces to names across an archive may fall within that. Take advice before doing it. This is general information, not legal advice.
Will a tagging service use my images to train its AI?
That depends entirely on the provider's terms, which differ between services and between pricing tiers. Ask for the position in writing: whether images are retained after processing, whether they are used for model training or improvement, where they are processed and who the sub-processors are. Locally run models avoid the question.
What is the difference between AI tagging and semantic image search?
Tagging writes words into a record. Semantic search converts each image into a numerical embedding and finds pictures that match a text query or resemble another picture, with no keywords involved. Semantic search is good for exploration and for finding similar frames; keywords and captions remain the durable, exportable record that travels with a licensed file.
Published by Archive.London. Corrections are welcome under our editorial policy.

