What is a photograph?

Not long ago, I had a discussion with a friend who said: “those AI-generated things aren’t photographs.” I replied that whether they are or not depends on how strict—or how broad—a definition of photography we choose.

We didn’t reach an agreement. But we’re still friends.

I’m not trying to lay down the law here, because I don’t have the ultimate answer. What I can do is explain my own point of view. My friend argued that, for her, to consider an image photographic it had to be made with a “tangible” tool, something physical. I, on the other hand, am less restrictive: I understand a photograph as an image created by the human hand with the help of a tool that records it. The notion of “recording” is key to distinguishing photography from drawing or painting: in a photographic image, the tool itself plays an active role in the process.

And for me, an AI algorithm is just another tool. No, I’m not saying that an AI-generated image is the same as a camera photograph. It isn’t, and it doesn’t need to be. Just as a film photo isn’t the same as a digital one, nor is color equivalent to black and white. Different techniques and tools yield different results.

“But an AI image isn’t real.”

Neither is this:

Oscar Rejlander. Study of the Head of St. John the Baptist on a Charger (c.1855). Source: Wikimedia Commons. Public domain image.

As far as we know, Rejlander didn’t actually behead anyone to make this picture. No photomontage or staging—like those that made Jeff Wall or Gregory Crewdson famous—are real either.

“That’s different. Those are artistic licenses. AI is pure deception.”

Enter Stalin:

Stalin, Molotov and Nikolai Yezhov. Unknown author (1937). Source: Picryl. Public domain images.

Just one of the hundreds of examples of deceptive manipulation—in this case, for political reasons, though motives for falsifying photographs have varied widely throughout history.

“No, that’s not what I mean. AI doesn’t create anything. It just copies what already exists.”

In 1981, Sherrie Levine re-photographed Walker Evans prints and published them as After Walker Evans. This is called appropriationism, an artistic movement born in the 1970s that consists of using an existing artwork to create a new one. It’s a concept we probably owe, like so much in contemporary art, to our dear Marcel Duchamp. Levine’s work is the textbook case, but there are plenty of others: Douglas Levere reworking Berenice Abbott’s New York Changing; Carrie Mae Weems with J.T. Zealy’s daguerreotypes of enslaved Black people; Mishka Henner with Robert Frank’s The Americans… shall I go on?

“But in all those cases, there’s a camera! Someone re-photographed those images with a camera. With AI, there’s no such thing.”

As if cameraless photography had been invented yesterday:

László Moholy-Nagy. Photogram (1926). Source: Wikimedia Commons. Public domain image.

A photogram is a cameraless photograph, made by placing objects directly on a photosensitive surface and exposing it to light. The early-20th-century avant-garde loved them, and there are contemporary artists still working in this way, like Michael Flomen, Adam Fuss, or Garry Fabian-Miller. And photograms aren’t the only cameraless technique—there are cyanotypes, scanography, chemigrams, luminograms…

“Damn, at least in those there’s light. That’s the key: in AI there’s no light. So it’s not photography!”

Now we’re getting somewhere: indeed, in an AI-generated image there isn’t, at least directly, any impression of light on a chemical or digital surface, which is the foundation of the classical definition of photography.

But AI, in reality, is just a supercharged pattern-finder that can recognize structures and recombine them into new ones. It doesn’t know what it’s doing, nor is it conscious of doing it (thankfully—we’ve all seen Terminator). So if you ask it to paint a blue sky, it looks in its gigantic neural networks for blue skies, but it won’t be able to generate anything unless you describe it in terms that allow it to find a base pattern. You can ask it for an alien—something no one has ever seen—but you still have to describe it using existing elements it can draw upon to build the new image.

Even if my explanation of how AI works is a bit crude, the key is its essentially imitative (or recombinative, if you prefer) nature. For the blue sky, it draws from previous photographs of blue skies, which it modifies and combines. And in those photographs, light was very much present… So, couldn’t we say that in an AI-generated image there is light, albeit “second-hand”?

“Alicia, you’re stretching it too far… and besides, it’s not just that there’s no light, there isn’t even an author!”

Let’s break it down: indeed, that last point about light is, in my view, the key factor for not considering an AI-generated image as photography. As for authorship… I’ll just give one counterexample: The Pillar. Stephen Gill mounted a surveillance camera on a pole, one of those that trigger automatically with movement. With it, he captured wonderful images of birds perching next to the camera and published one of the most important photobooks of 2019. Who took those pictures? The animals?

“No, of course not, those photos belong to Gill, because he had the idea. To do something like that you need ideas, talent… But anyone can type an AI prompt. Where’s the merit in that?”

I think I’ve heard that lack-of-merit argument before:

“Don’t think you became an artist the instant you received a gift Kodak on Xmas morning.”

— Alfred Stieglitz, from Twelve Random Dont’s, Photographic Topics 7 (1909). Source: The Phoblographer.

“There are four simple words on the matter, which must be whispered: Color photography is vulgar.”

— Walker Evans, at a MoMA exhibition (1956). Source: Greg Fallis.

And there are artists who regularly use AI in their work, for example, my beloved Joan Fontcuberta. I’ll gladly duel anyone who claims what he does isn’t photography.

I think the “AI yes or no” debate is just an updated version of the “snapshot yes or no” that entertained the Pictorialists in the early 20th century, or “color yes or no” in the 1950s, or even “digital yes or no” from the 1990s onward. Same story again. Every new technology has always sparked the same discussion about whether using it was legitimate. It’s always been that way, and it always will be.

As I said at the outset—and I doubt anyone has the definitive answer—we can’t just dismiss AI with a blanket “those aren’t photographs.” This technology is here to stay and, in some contexts, it has clear advantages over camera-based photography. For example, an online fashion store could photograph garments on hangers and, using existing photos of models, generate AI images of those garments being worn, saving expensive, complex photo shoots. Does it pose a threat to certain jobs? Yes—just as refrigerators did to the doorstep ice trade—and at the same time it creates new specialties. As usual, legislation lags behind. Urgent debates are opening up around authorship, the legitimate use of training data, bias, and the duty to disclose when an image is synthetic. In the end, though, it’s just a tool, like a hammer. Can you really be “against” a hammer?

Notes:

• The opening illustration was generated with Sora AI.

• This article contains no affiliate links.