Nobody needs to break into an archive to falsify the past. They only need to publish more convincing pictures of it than the archive has. A transparency in a dated sleeve is the stubborn answer.
The danger to the historical record is dilution, not forgery: a thousand plausible pictures of 1976 for every one that was taken there.
The usual fear about deepfakes and history is the perfect forgery: a fabricated photograph slipped into the record and believed. The more likely damage is duller and larger. It is dilution. When anyone can produce a plausible picture of a street in 1976, the genuine photographs of that street do not vanish; they are simply outnumbered, and a search for the decade returns a blend. Our argument is that archives answer this less by detecting fakes than by being fixed points: places where an image can be traced to an object that existed before the question was asked.
A note on scope. "Evidence" here means historical evidence, the kind a picture researcher, curator or biographer weighs. Whether a photograph is admissible in a court is a separate legal question, and this article does not address it.
Can AI generate convincing historical photographs?
Convincing to most people, yes. In the Microsoft AI for Good Lab's experiment, around 12,500 participants made about 287,000 judgements of real and generated images and were right 62% of the time. Period styling makes the job harder, since faded colour, grain and soft focus are exactly the cues viewers use to decide that something is old, and they are trivially easy to synthesise.
Convincing to a specialist is another matter, and here the research is instructive. In Synthetic History, a paper first posted in May 2025, Maria-Teresa De Rosa Palmini and Eva Cetinic assembled 30,000 images from three diffusion models, prompted with ordinary human activities across historical periods. They report "systematic inaccuracies": the models attach default visual styles to eras without being asked, introduce anachronisms such as modern objects in earlier settings, and produce demographic mixes that do not match plausible historical baselines. We look at the causes in why image models get the past wrong.
So the generated past is wrong in ways an expert can catch and a casual viewer cannot. That asymmetry is the problem. The record is not consulted only by experts.
Is faking photographs of the past a new problem?
No, and the history is a comfort of sorts. In 1917 two girls in Cottingley, Yorkshire, photographed paper cut-outs and persuaded a good many adults that they had photographed fairies; the admission did not come until the 1980s. Darkroom retouching removed inconvenient people from official photographs throughout the twentieth century. Captions have always lied more often than pictures.
What is new is cost and volume. A darkroom forgery took skill and hours and produced one image. A generator takes seconds and produces as many as you ask for, each different, each without a source. The old forgery was an object that could be examined. The new kind is a stream, and there is nothing to send to a laboratory.
What makes a film original hard to forge?
A transparency or negative is a manufactured object processed at a particular time, and both manufacture and processing leave marks that have nothing to do with the picture.
Kodachrome is the best documented case. The conservator Gawain Weaver publishes a timeline that dates the film's cardboard mounts by their design: a red border with plain lettering from 1939 to 1949; the words "Processed by Kodak" added between 1955 and 1959; a small edge-curl logo with slab lettering from 1965 to 1972; "Slide" replacing "Transparency" on many mounts from 1975 to 1980; rounded corners from about 1980. Many mounts of the late 1950s and 1960s carry a stamped processing date. A small red plus sign marks the changeover to the K-14 process, roughly 1974 to 1983. The film types themselves have dates: Kodachrome II was sold from 1961 to 1974, Kodachrome 25 from 1974 to 2001. The same timeline warns that designs varied by laboratory and country, and that its ranges are estimates to be cross-checked.
Other films and formats have their own tells, less thoroughly published: edge printing between the perforations, notch codes on sheet film, the look of a particular laboratory's sleeves. Then there is everything the photographer added. Handwriting on a mount. A job number. The contact sheet with a chinagraph ring around the chosen frame. And the roll itself, on which the frame in question sits between others, taken seconds apart, showing the same light.
The point is not that any one of these is unforgeable. It is that they are independent. A forger has to get the stock, the mount, the stamp, the handwriting, the sequence and the paper trail all right, and all consistent with a date. Each one is a separate chance to be caught.
Could someone forge a transparency?
Yes, and an honest account has to say so. Film recorders, which write digital images on to film, were ordinary equipment in the 1990s. A screen can be photographed on old stock. Old mounts can be reused.
Three things limit this. The material is finite: nobody is making Kodachrome, and the last roll was processed in January 2011. The forgery has to survive magnification, where the structure of a recorded or re-photographed image differs from a camera original. And it has to arrive with a history. A lone slide of uncertain origin is weak evidence however good it looks. A slide that has sat in a numbered sleeve, in a listed folder, in an inventory published years before anyone had a reason to dispute it, is strong evidence. Provenance was always the real test; the film is what gives provenance something to hold.
What can archives do to remain a fixed point?
Most of it is unglamorous record-keeping, done early.
| Measure | What it fixes | Limit |
|---|---|---|
| Publish an inventory | What the archive held, and when it said so | Describes folders, not every frame |
| Number each original | Which object a scan was made from | Only as good as the register |
| Hash each master scan | That a file is byte-for-byte the archive's scan | Says nothing about copies that were resized |
| Declare the source in metadata | That the file was digitised from film | Metadata can be stripped |
| Sign with Content Credentials | Who vouches for the file from the scan onward | Starts the clock at the scanner |
| Keep and house the originals | The possibility of looking again | Costs space and care |
| Label restorations | Which version is the record and which is an interpretation | Depends on discipline |
Museums and collecting institutions, many of them listed in our photography directory, have been doing much of this for a century under other names. A few notes on the less obvious rows. The hash in the third row is a SHA-256 value recorded when the scan is made and kept apart from the file; the image hash and fingerprint tool computes one in the browser. Collections Trust's guidance on numbering, part of the Spectrum standard for museums, asks for a unique number for every accessioned object, physically associated with it. That is the hinge between object and file, and we set out the options in giving every photograph a permanent identifier. For the metadata row, the IPTC's Digital Source Type vocabulary has had values for "digitised from a transparent negative" and "digitised from a transparent positive" since 2008, sitting in the same list as the newer value for media created with generative AI. One field, honestly filled, separates the two kinds of picture.
The last row matters more each year. AI upscaling and colourising tools produce attractive versions of old photographs by inventing detail. If those versions circulate without labels, the archive itself becomes a source of synthetic history. Our rule for that is in AI colourising and upscaling of archive photographs.
What archives cannot do is police the internet. Detection will help at the margins, as the checks that still work describes. The durable contribution is to make it easy for an honest researcher to reach the real thing.
Where does the Robert Harper archive stand?
The Robert Harper Photograph Library has published its complete inventory: 503 folders, searchable, with shot and frame counts. Some entries carry dates in their descriptions. "Jethro Tull, November 1974". "Stan Webb Band, March 1975". "Eric Clapton at The Royal Albert Hall 1993 (five sheets of slides)". "Street scenes: Chelsea Arts Club 1970s, New Bond Street 1976". A folder of 560 black-and-white landscape shots from 1975 is described as negatives and contact sheets, which is to say the sequences survive alongside the selected frames.
We should be exact about what this does and does not establish. These are the inventory's descriptions. They have not yet been verified frame by frame against the film, the counts have not been independently reconciled, and more than half of the archive is still on film and has yet to be scanned. The dating evidence described above, the mounts and edges and sleeves, has yet to be recorded. That work belongs to cataloguing, and it is one reason we are looking for partners to do it properly.
What can be said now is simple. The archive holds 52,684 unique analogue photographs by its inventory count, the work of one photographer over about thirty-five years, owned by a single rights holder whose position is set out under rights and stewardship. The list of what is in the boxes is public and dated. The analogue collection describes the formats. The aim, as scanning proceeds, is that every file can be traced to an object.
Museums, picture researchers and publishers who want to examine material at its source can ask through professional access.
Frequently asked questions
Can AI create fake historical photographs?
Yes. Current image generators produce convincing pictures in the visual manner of any decade, and people identify generated images only modestly better than chance. Research on generated depictions of the past has also found anachronisms and defaults to period styling. The pictures need not deceive a specialist to do harm; they only need to circulate more widely than the genuine record.
Why are film negatives and transparencies good evidence?
Because they are physical objects made at the time. The film stock, the edge markings, the mount, the sleeve, the handwriting and the adjacent frames each carry information about date and origin, and they have to agree with one another. Fabricating one convincing image is easy. Fabricating a consistent object with a consistent history is expensive.
How can you date an old slide?
Start with the mount and the film. Processing laboratories changed mount designs, logos and wording over time, and some stamped the month and year of processing. Film types were sold only in certain years. A conservator's published timeline for Kodachrome, for example, dates mount styles to ranges of a few years. Treat these as estimates and compare them with written records.
Can a film original be faked?
In principle. A digital image can be written to film with a film recorder, or a screen can be photographed. What is difficult is making the result agree with everything around it: a stock that was on sale at the claimed date, period mounts, a matching sequence of neighbouring frames, and a paper trail that predates the claim. Evidence from originals is cumulative.
What can archives do about deepfakes?
Fix their own position in advance. Publish an inventory, give each original a permanent number, record a cryptographic hash of every master scan, declare the source in the file's metadata, keep AI restoration out of master files, and label any enhanced version as such. An archive cannot stop false pictures being made; it can make its own holdings checkable.
Sources
- De Rosa Palmini and Cetinic, Synthetic History: Evaluating Visual Representations of the Past in Diffusion Models (arXiv 2505.17064)
- Roca et al., How good are humans at detecting AI-generated images? (Microsoft AI for Good Lab, arXiv 2507.18640)
- Gawain Weaver Art Conservation: Kodachrome timeline (film types, processes and mount designs by date)
- Collections Trust: Numbering (accession numbers under Spectrum)
- IPTC NewsCodes: Digital Source Type vocabulary
- Wikipedia: K-14 process (dates of the process and its end)
Published by Archive.London. Corrections are welcome under our editorial policy.

