If AI Is Trained on Your Photographs, Is the Result Still Photography?

Photography has worked without a camera and without shooting. Generative images test what is left after that.

You spend twenty years building a photographic archive, then train a generative model on it. The model begins producing images that look unmistakably like your work. Are those images photography?

Two questions hide in that one. Are the images photographs? And can the practice that produced them still be photographic?

The camera does not settle the question either. Photography has worked without one for almost two centuries. In the 1840s, British botanist Anna Atkins placed algae and other plants directly on sensitized paper to make cyanotypes; her self-published book is often described as the first illustrated with photographs. Man Ray later made photograms he called rayographs as part of his work in Dada and Surrealism. László Moholy-Nagy approached the same process differently at the Bauhaus, treating light itself as photographic material rather than simply a means of recording an object.

Wolfgang Tillmans continued working without a camera decades later. A German photographer known for casual-looking pictures of friends and club culture, and the first photographer to win the Turner Prize, he made works such as Blushes and Freischwimmer directly on photographic paper in the darkroom. There is no negative and no object in the usual sense. These works still belong to photography because light physically changes a light-sensitive surface.

Photography learned long ago to do without a camera. It still needed something to happen with light.

Later the artist did not need to make the source image either. Some photographers began working with existing images as files: Thomas Ruff, a German photographer known for his large, passport-style portraits and for series built from images made by others, from newspaper pictures to astronomical photographs, enlarged compressed pictures, many of them found online, until their digital basis became visible. Others worked with subjects people photograph again and again. Penelope Umbrico, an American artist known for collecting found images from online platforms, from sunsets on Flickr to broken TV screens in eBay listings, cropped the suns from thousands of sunsets posted on Flickr and printed them as ordinary snapshots.

Here, the artist does not make the source photographs. But they are still photographs: each one records something that was once in front of a camera.

Anna Ridler, a British artist known for building her own datasets by hand and exhibiting them as artworks, brings the artist back behind the camera, but for a different purpose. Over three months, roughly one tulip season, she photographed around ten thousand tulips. She exhibited the photographs as a work, Myriad (Tulips), and used the same set to train a generative model for Mosaic Virus, a video of tulips that bloom and change on screen. Each photograph in Myriad is tied to a particular flower and a particular moment. The tulips in Mosaic Virus are not traces of particular flowers.

A generated image can look like a photograph, but no corresponding event stood before a camera. The photographs in a training set have a real origin. The model learns patterns across them, and the image it produces does not inherit that link.

A photographer might ask how this differs from heavy compositing. A composite still carries its source photographs into the final image. A generated image does not work that way. What the model does inherit is what the material carries: styles, habits of framing, standard ways of showing people and events, the biases of the archive.

The model's output is not a photograph in the sense a cyanotype or a Tillmans print is. Nothing in it was exposed to light from the scene it shows. It does not become a photograph because photographs went into the model.

Photographic material has never decided the medium on its own. In the 1960s, Gerhard Richter, a German painter known for blurred paintings copied from photographs as well as large abstract canvases, began painting from photographs, from newspaper pictures and family snapshots. The source is photographic. The blur and the flat light imitate photography. The result is a painting, because the image was made with paint.

Umbrico's work remains photographic at both ends: the source images are photographs, and the final objects are photographic prints. A generated image can also be printed on photographic paper. This introduces a photographic process at the final stage, but it does not restore the missing link to an event.

The claim here is therefore narrower. The generated image is not a photograph. The open question concerns the practice around it.

Ridler's work shows the difference. Myriad (Tulips) consists of photographs. Mosaic Virus does not. The work as a whole moves from one to the other. She photographed the flowers, organized the set, exhibited it, and trained the model on it. Most of the decisions in this work are photographic decisions. Its final image is not a photograph.

This is the question behind this text. Can a practice built on photographs remain photographic when its result is not a photograph?

The question is becoming practical. Most generated images come from systems trained on images, many of them photographs, that no user selected. Fred Ritchin, a former picture editor of The New York Times Magazine who has written about the digital manipulation of photographs since the 1980s and is the author of The Synthetic Eye, recently compared images he generated from the same prompts several years apart. In 2022, "The perfect family" gave him a strange, almost archival group portrait. In 2026, the same prompt gave him smiling parents, two children, and a dog on a white sofa. He finds such results "predictably banal" and links this partly to the mass of online photographs these systems learned from.

Photography has already expanded beyond the camera, and later beyond the need for the artist to make the source image. In both cases, either the photographic process or the photograph itself remained. With generative images, neither has to remain in the final result. Photography is present only as material.

The art historian Noam M. Elcott and the technologist Tim Trombley trained separate models using curated sets of roughly 1,000 public-domain photographs by Dorothea Lange and 1,400 by Gordon Parks from the Library of Congress's FSA/OWI collection. The collection includes Lange's Migrant Mother, made in 1936, as well as Parks's later work for the Farm Security Administration. Their models reproduce with striking precision the manner of Lange and Parks, who later became the first Black staff photographer at Life. Elcott and Trombley also take a position on whether photography as material is enough. In AI imaging, they write, "there is no photography, only photographic style."

That answer concerns the generated image. The tulips show what it leaves out. Ridler's photographs are more than the source of a style. She made them and exhibited them as part of the work.

Can photography remain the medium when neither a camera nor a photographic process produces the image, but photography itself becomes the material?