AI and images

Can a machine date a photograph? Clothes, cars, film stock and clues

Houses, a church and winding roads sit among fields and trees.
Village from the air.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

Of the 503 folders in one studio archive, seven have a year written on them. Dating the rest is the archivist's daily puzzle, and the machines that offer to help are reading only two of a photograph's four clocks.

A photograph has four clocks: what is in it, what it is made of, how it was styled and what the sleeve says. Machines read two.

Of the 503 folders in the Robert Harper inventory, seven have a year or a decade written into their description. Jethro Tull, November 1974. Stan Webb Band, March 1975. New Bond Street, 1976. Eric Clapton at the Royal Albert Hall, 1993. The other 496 say what is inside, not when. That ratio is normal for a working studio's files, and it is why dating is the archivist's daily puzzle. Software now offers to solve it from the pixels. Our angle is that a photograph keeps time on four separate clocks, that machines can read two of them, and that the only way to know how well is to test them against pictures whose dates are already known.

Can AI tell when a photograph was taken?

To a degree, and the best evidence comes from two studies that are honest about their conditions.

In 2012 Frank Palermo, James Hays and Alexei Efros presented Dating Historical Color Images at the European Conference on Computer Vision. Their task was to sort historical colour photographs by decade. Instead of looking at what was in the pictures, they designed features to capture how colour imaging processes themselves changed over time, and reported significantly greater accuracy than untrained humans on the same images. The machine, in other words, was dating the film as much as the scene.

In 2015 Shiry Ginosar and colleagues assembled 168,055 portraits from American high school yearbooks, 37,921 of them frontal-facing, and trained a classifier to predict the year. The median error was 4 years for women and 6 years for men. The authors also showed which parts of the image the network relied on, and found they corresponded to the fashions of each era.

Those are good results, and they define the ceiling. Yearbook portraits are the easiest material imaginable: one country, one genre, one pose, subjects of the same age wearing their best current clothes, and a printed year on every volume. Even there, half the estimates for men were more than six years out.

General-purpose models are a different case. A vision-language model in the CLIP family, trained on hundreds of millions of captioned web images, can be asked which decade a picture looks like and will answer. It has seen many captions containing dates. But nobody has checked those captions, and a chat model built on such a system will produce a fluent paragraph of reasoning whether or not the reasoning is sound.

What are the four clocks in a photograph?

Clock Examples Typical precision Read best by How it goes wrong
The subject Clothes, hair, cars, telephones, packaging, shop fronts A decade; a few years with luck Machine and human Old things persist; fashions return
The material Film stock, format, mount, edge markings, print paper An earliest date; sometimes a year Human with the object in hand Lost when only a scan survives
The style Lighting, lens, pose, colour palette, grain A decade Machine and human Deliberate revival and pastiche
The paperwork Sleeve notes, job sheets, processing stamps, the published page A month or a day Human Missing, or wrong

The table carries the main point. The two clocks a model can read from a scan, subject and style, are the two least precise. The two most precise are either physical or written, and they are exactly what gets separated from the image when a transparency is scanned and the file is renamed.

How do clothes, cars and telephones date a picture?

By setting limits. Each datable object gives the earliest moment the photograph could have been taken, and the collection of objects narrows the window.

Vehicle registration plates are the classic British example, because the state wrote the year on the car. A suffix letter denoting the year first appeared in February 1963, became compulsory on 1 January 1965, and from 1967 changed each 1 August. When the alphabet ran out the letter moved to the front: prefix registrations began on 1 August 1983. The current format, in which two digits change every March and September, started on 1 September 2001. The plates themselves changed as well. White front and yellow rear reflective plates became a legal requirement for most newly registered vehicles on 1 January 1973, so a street full of black plates with silver characters is probably earlier, and a mixture suggests the mid-1970s.

The logic has a firm rule and a soft one. The firm rule: a photograph cannot be older than the newest thing in it. The soft one: streets are full of old cars, and people wear coats for years, so the typical object in a scene predates the photograph. A picture's date lies at or after its newest element, not at the average.

Technology dates well for the same reason and is easier to read than clothing. A telephone with a dial, a first-generation mobile handset, a particular model of personal stereo: each has a launch date that can be looked up. This is one reason product photographs, which seem undatable, are often the most datable of all once someone identifies the product. A fragrance bottle on a white background contains no clothing and no car, yet the fragrance has a launch year and the packaging design has a lifespan. That is research, though, and no vision model does it reliably.

Can film stock date a photograph?

Often better than the image can, provided the film is in front of you. Manufacturers printed the stock name and codes along the edge of the film, outside the picture area. Mounted slides frequently carry a processing date stamped on the mount by the laboratory. Emulsions and chemical processes have introduction dates, so a transparency on a stock that reached the market in 1990 cannot have been exposed in 1985. Formats and mounts changed too, and Polaroid materials have their own chronology.

A scan throws most of this away. The edge markings are cropped off, the mount stays in the box, and the colour characteristics that Palermo and colleagues exploited are deliberately corrected by a careful operator. A restored scan of a 1975 transparency may look, to a model, like a photograph from no particular year. There is a practical lesson here for digitisation: photograph the mount and the sleeve, scan the full frame including the rebate where the standard allows, and record the stock in the metadata. Our guide to captioning archive photographs from contact sheets covers what to transcribe.

Born-digital files have a clock of their own. A camera writes the capture time into the file, and the EXIF viewer will show it. It is as accurate as the camera's clock was on the day, which is to say usually right and occasionally set to the wrong year.

How do dating models fail?

In ways an archivist would recognise, and one folder in this archive makes the point better than any benchmark. The inventory records the London to Brighton Veteran Car Rally, November 1974. The event is for cars built before 1905. The veteran cars at such an event are some seventy years older than any photograph taken of them in 1974. A model that dates pictures by their cars will be wrong by a lifetime, confidently, and a human would not be, because a human looks at everything else in the frame.

The failure modes generalise:

  • Old objects in new pictures. Vintage cars, antiques, period costume, heritage railways, film sets.
  • Persistence. The newest object in a scene sets the date, but models tend towards the average object.
  • Pictures with no clock. Studio still life on a plain background: food, jewellery, a shoe. There is nothing to read except style.
  • Shortcuts. Fading, colour casts, scratches and monochrome are associated with age in web images. A model can learn that faded means old, which fails on a well-kept original and on a restored scan alike. Black and white is not a date.
  • Training bias. A model taught on American yearbooks knows American teenagers. London street scenes, Scottish landscapes and Italian interiors are another distribution.
  • Fluency. Conversational models explain their answers persuasively. The explanation is generated, not observed, and can cite details that are not in the picture.

None of this makes the tools useless. A decade estimate is a good first sort for a box of unlabelled scans, and outliers within a folder are worth a second look. It does mean a machine estimate belongs in a catalogue as a flagged suggestion, in the same way as machine keywords; we discuss the companion technique of finding similar frames in searching an archive by picture.

Why do dated collections matter for testing?

Because a dating model can only be scored against pictures whose dates are known from something other than a guess. The yearbook study worked because every portrait came with a printed year. Most real collections are not like that, and the ones that are tend to be narrow.

A useful test set has properties that are hard to find together:

  1. Dates from records. A sleeve, a job sheet, a magazine issue. Not a cataloguer's estimate, which would only test the model against another guess.
  2. Many genres. Portraits, streets, landscapes, concerts, still life, technology.
  3. A long span. Several decades, with enough pictures in each.
  4. Hard cases on purpose. Old objects, plain backgrounds, revival styles.
  5. Unseen by the model. Pictures that have not been published online are less likely to be in any training set.

The Robert Harper Photograph Library is a candidate on some of these counts and not others, and we would rather be exact. Its seven dated folder descriptions run from a 1960s fire and 1970s London street scenes to black-and-white landscapes of 1975, the Art Blakey Big Band in Edinburgh in 1987 and Eric Clapton in 1993. The material spans the 1960s and 1970s to the digital era, across 24 categories, and the great majority of the archive has never been digitised or published online. Against that, frame-level dates mostly remain to be established from the sleeves and paperwork as cataloguing proceeds, and the inventory counts have not yet been independently reconciled. The London collection, the published inventory and the dataset card show what is recorded today.

Where the archive fits

Dating is where digitisation, cataloguing and machine learning meet: the scanner operator preserves the evidence, the cataloguer records it, and only then is there anything to test a model against. Researchers and data teams who want dated, human-made photographs for evaluation, on terms negotiated directly with the single rights holder and subject to rights review, can start with AI training data licensing.

Frequently asked questions

Can AI tell when a photograph was taken?

It can estimate, usually to within a decade and sometimes closer. Research models trained on dated images use clothing, hairstyles and colour characteristics. In a 2015 study of American yearbook portraits the median error was 4 years for women and 6 for men, under near-ideal conditions. Outside such conditions errors grow, and the estimate is a suggestion to check, never a date to record.

How do you date an old photograph yourself?

Work from the hardest evidence to the softest. Start with anything written: sleeves, mounts, job sheets, processing stamps. Then the material: film stock, format, print paper. Then datable objects in the picture such as vehicle registration plates, telephones, packaging and shop fronts. Use clothing and hairstyles last, because people wear old clothes and fashions return.

How can UK number plates date a photograph?

A plate shows the earliest the photograph can be. Year suffix letters started with A in February 1963, became compulsory in 1965 and changed each August from 1967. Prefix letters began on 1 August 1983, and the current format, with a two-digit age identifier changing every March and September, on 1 September 2001. Reflective white and yellow plates were required for most new vehicles from 1973.

Why do AI models get photograph dates wrong?

Because they infer from appearance. An old object in a new photograph, a period costume, a vintage car rally or a deliberately retro style all point to the wrong year. Models also learn shortcuts, such as treating fading or black and white as signs of age, and they are trained mostly on certain kinds of picture, so unfamiliar genres are dated poorly.

What is needed to test a photo dating model?

A set of photographs with dates known from independent evidence, across many genres, places and decades, kept separate from the model's training data. The dates must come from records such as sleeves, job sheets or publication, not from someone's estimate, and the set should include deliberately hard cases such as still life and photographs of old objects.

Sources

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