AI and images

Why image models get the past wrong, and what dated photos correct

A fleet of yachts heels under white sails on choppy water.
Classic yachts under sail.Photograph by Robert Harper. Robert Harper Photograph Library, Archive.London. From the archive, shown for illustration.

Ask an image generator for an ordinary street in 1976 and it reaches for flares, a faded filter and props from the wrong decade. The cause is what the web kept, and the cure is duller than you would think.

The internet remembers the 1970s the way a fancy dress shop does: vividly, selectively, and with everything from the decade on at once.

Image models are bad at the past for a simple reason: they were taught it by the internet, and the internet's memory of any year before about 2000 is a highlights reel. Researchers have now measured the result. The angle we add is a practical one. The corrective for a model that thinks 1976 was a costume party is not more famous pictures of 1976. It is boring ones: a telephone on a white background, a calculator, a shop front, each with a date somebody can vouch for. Studio product photography, of all things, turns out to be a clock.

What do image models get wrong about the past?

The most systematic study so far is Synthetic History, by Maria-Teresa De Rosa Palmini and Eva Cetinic, first posted in May 2025 and revised in February 2026. They generated 30,000 images from three text-to-image diffusion models, using prompts about universal human activities set in different historical periods, and assessed the results on three counts.

Unprompted styling. Asked only for an activity in a period, the models applied a visual manner they associate with the era: the abstract speaks of "implicit stylistic associations". The past arrives pre-filtered, whether or not anyone asked.

Anachronism. Modern artefacts turn up in pre-modern settings.

Demographics. The mix of people generated does not match historically plausible baselines.

The authors' phrase for the overall picture is "systematic inaccuracies in historically themed generated imagery". Their benchmark concerns broad periods. Our interest is in a finer grain, the difference between 1976 and 1986, where the same mechanisms operate and the errors are easier to miss because everything looks roughly right.

Why does training data skew towards the present?

Consider how a decade gets into a training set. LAION-5B, the largest public example, was assembled from 5.85 billion images found on the web alongside text. For the years since camera phones, that means an ocean of ordinary life: lunches, living rooms, commutes. For the years before, it means only what somebody took the trouble to scan and post. Four filters apply.

  1. Survival and effort. Scanning film is slow. People scan what they think matters: family, celebrities, events, cars they loved.
  2. Fame. A few hundred much-reproduced photographs stand in for whole decades, each copied many times over.
  3. Re-creation. A great many images captioned with a decade were made long after it: period dramas, themed parties, fashion revivals, advertising pastiche. To a model reading captions, a 1970s party held in 2015 is as much "1970s" as a photograph from 1975, and it is usually the cleaner digital file.
  4. Captions. Very few online captions state a year. "Vintage", "retro" and "old school" are moods. Even a decade label is ten years wide, and ten years was a long time in telephones.

The outcome is a model with a strong sense of the decade as a brand and a weak sense of it as a sequence of particular years. It knows flares. It does not know which telephone.

Why do generated images of the 1970s look like a costume party?

Because, in the training data, many of them are. But there is a second effect that has nothing to do with content. Most genuine photographs of the period reach the web as scans of prints and slides that have aged: colour shifted, contrast flattened, as we explain in why colour slides fade. The model therefore learns that the past is orange and soft.

It was neither. A transparency shot in a studio in 1976 was, on the day it came back from the laboratory, as sharp and saturated as the film could make it, and medium-format film could make it very sharp indeed. The washed-out look is a property of fifty years in an album, not of the year. A generator that renders 1976 with a nostalgic haze is depicting our photographs of the time, not the time. We make the related case for film stocks in what a Kodachrome prompt gets wrong.

How do dated objects expose anachronisms?

Historians date photographs by what is in them, and the method works in reverse on generated images. An object cannot appear before it existed. A handful of consumer landmarks, with their sources, shows how tight the constraints are.

Object Earliest date What it rules out
Barclaycard, the first UK credit card Launched 29 June 1966; revolving credit added November 1967 A British credit card at a till before the mid 1960s
Kodachrome 64 slide film Sold from 1974 A slide on that stock from an earlier trip
Personal cassette player with headphones (Sony Walkman TPS-L2) On sale in Japan 1 July 1979 Anyone wearing one in a 1976 street scene
Handheld mobile telephone (Motorola DynaTAC 8000X) US regulatory approval 21 September 1983 A handheld mobile in any 1970s scene
Mobile call on a UK commercial network 1 January 1985, on Vodafone's network A mobile in use on a British street before 1985

So a request for "commuters in London, 1976" has hard edges. No personal stereos. No mobile telephones. Cars, buses and shop signs of the right generation. An image that breaks one of these is not a matter of taste; it is wrong, and checkable. The constraints run the other way too. A scene set in 1986 that contains nothing introduced after 1976 is just as false, only harder to notice, because absence does not catch the eye.

This is also why the errors matter beyond pedantry. A generated illustration in a documentary treatment, a museum interactive or a school resource teaches the anachronism to people who have no reason to doubt it, a risk we discuss in deepfakes and the historical record.

What kind of photographs correct period bias?

Not the famous ones; the model has those many times over. What is scarce is the following.

Ordinary objects, photographed plainly. A product shot is the cleanest possible training example of an object: isolated, well lit, unambiguous. Catalogues and magazine shopping pages were full of them, and they are precisely what nobody scanned.

Dates at the right resolution. A year, not a decade. Commercial product photography is usually made when the product is current, which is what makes it useful, though each picture's date has to be established from the job record and not assumed.

Consecutive coverage. One photographer or studio working through twenty or thirty years records the succession: the dial telephone, the push-button, the cordless, the brick-sized mobile. A model needs the sequence to learn that these did not coexist.

Scans made from originals. A new scan from a transparency, with documented colour handling, shows what the film recorded. It lets a model, or the people evaluating one, separate period from patina. Resolution helps as well: a 6 x 7 transparency holds enough detail to read the lettering on a keypad or a cassette door, and our scan resolution calculator shows what a given original yields at a given setting. Legible detail is what turns a picture of "an old telephone" into evidence of a particular one.

Honest labels. A description that says what is visible and what is known, and distinguishes the two, as set out in image captions for machine learning.

Dated, documented material of this kind is as useful for testing as for training. A set of photographs with known years is a ready-made check on whether a model can tell 1979 from 1989, which is the argument of evaluation sets for image models.

What does the Robert Harper archive record?

The Robert Harper Photograph Library spans about thirty-five years of professional work, from the 1960s and 1970s into the digital era, and a good deal of it is exactly the unglamorous category described above. The published inventory lists, among its 503 folders:

  • Telephones, mobiles, pagers and fax machines: 130 unique shots, 305 frames
  • Speakers, radios, Walkmans, stereo equipment, hi-fi systems and headphones: 153 unique shots, 661 frames
  • Computers, calculators, mouses, screens and displays: 77 unique shots, 279 frames
  • Cash, credit cards and related subjects: 75 unique shots, 180 frames
  • Street scenes, described as Chelsea Arts Club 1970s, New Bond Street 1976 and a 1960s fire: 101 shots

These sit within the corporate and technology collection, and the story of the objects themselves is told in vintage technology photographs of the 1970s and 1980s.

The limits need stating as clearly as the numbers. The inventory describes folders, not individual frames. It does not yet give a date, make or model for each photograph; those will come from examining the film and any job records during cataloguing. The counts are the inventory's and have not been independently reconciled, and more than half of the archive is still on film and has yet to be scanned. We do not claim that any of this material has improved any model, since it has not been used to try.

What can be said is that it is the right kind of material: everyday objects, photographed professionally at the time, by one photographer, with a single rights holder. Teams working on temporal accuracy who want to see what a dated selection could look like can start with AI and data licensing, which is available by negotiation and subject to rights review.

Frequently asked questions

Why do AI image generators get historical details wrong?

Because they learn from captioned images on the web, and the web's record of any period before digital cameras is small, selective and loosely dated. Famous events, fashion and later re-creations dominate, while ordinary objects and streets are scarce. The model fills the gaps with whatever is most strongly associated with the decade's name.

What is temporal or period bias in training data?

It is uneven coverage of time. Recent years are represented by billions of everyday photographs; earlier decades by a much smaller set that someone chose to scan and publish. Captions seldom state a date, so words such as vintage or 1970s attach to a mood. Models trained on this reproduce the mood and miss the specifics.

Can AI generate historically accurate images?

It can produce plausible ones, and with careful prompting and reference images a skilled user can get close on particular details. Research evaluating 30,000 generated images of historical scenes found systematic inaccuracies, including anachronistic objects and default period styles. For anything presented as history, generated images need labelling and expert checking.

How do you spot an anachronism in a photograph?

Look for objects with known introduction dates: vehicles, telephones, audio equipment, packaging, signage, typefaces. A scene cannot contain a product that had not been released. Then check the rendering: a photograph claimed to be from a given year should match the films and printing of that time, not a modern idea of how old pictures look.

What kind of data helps models represent the past better?

Photographs with reliable dates and plain descriptions, especially of ordinary things: products, interiors, streets, workplaces. Scans made from originals with documented colour handling help separate what a period looked like from how its photographs have aged. Coverage across consecutive years matters more than volume from one famous moment.

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