Photo or AI? It's the Wrong Question
- 4 days ago
- 7 min read
Why classical montage vanished from public perception when AI arrived — and why that concerns photographers too
The same comment shows up under my work again and again: "AI, right?"
It isn't meant unkindly. It's an honest question. And it points to something that has been on my mind for a while: in public perception, there are now only two kinds of images. Photography or AI. Anything that couldn't plausibly have come straight out of a camera lands automatically in the second bucket.
In between sits a category that has existed for a hundred and seventy years and is quietly disappearing.
What changed on August 2
Since August 2, 2026, Article 50 of the EU AI Act requires that AI-generated or substantially AI-manipulated content capable of passing as real be clearly marked. Ordinary editing — exposure, color, noise reduction, retouching — does not trigger the obligation. For work that is recognizably artistic, the requirement is reduced to a discreet notice that doesn't interfere with the experience of the piece.
For my own work, that's a manageable matter. What interests me more is what the law does for everyone else.
For photographers, this is good news
If convincingly realistic AI content will carry a visible mark, that changes the meaning of the unmarked image. Until now, an image without a notice said nothing at all, because there were no notices. Now it says something — purely by what's absent.
In feeds increasingly filled with labeled material, the unlabeled image becomes the difference. Not a marketing promise, not a claim you have to invent for yourself: a structural advantage handed over by the legislator. Anyone who photographs should use it. It's the best upgrade commissioned photography has received in years.
There's also a reassurance that's getting lost in the noise: as a photographer, the label will almost never apply to you. Not even if you edit heavily — and we all edit heavily. Exposure, color, contrast, skin retouching, the distracting branch at the edge of the frame: that has been part of the craft for a hundred and fifty years, and the regulation explicitly leaves it alone, as long as the statement of the image doesn't change substantially.
Where that line runs is the interesting part. A retouched-out trash can shifts nothing. A replaced sky does — it changes what there is to see. If you drop one in at the press of an AI button, don't count on it still passing as standard editing. If you composite it from a second frame of your own, the obligation never applies in the first place — the same visible change, two entirely different legal situations.
Which means the way of working I'm about to describe is closer to yours than it first appears. Between an elaborately edited photograph and a montage there is no gulf, only a gradient.
But: the missing label says something different from what most people assume.
Photo or AI? What the label answers – and what it doesn't.
Photo or AI? Why That's the Wrong Question. The disclosure requirement doesn't ask whether an image is true. It asks how it came about. That's a fine distinction, and a decisive one. An AI-generated image of a completely ordinary street scene gets a label, even though it shows nothing unusual. A scene composited by hand that never existed and never could exist gets none — because no AI system produced it.
The criterion is the technique behind the image, not the truthfulness of what it shows.
How far that reaches is clearest in a small routine move. I remove a power line from a landscape shot. Do it with Generative Remove, and an AI system was involved. Do exactly the same thing with the Clone Stamp, and none was. The finished image is identical in both cases, the line gone in both cases, the viewer looking at something that wasn't there in both cases. Legally, these are two different acts.
I don't want to overstate it. The regulation contains an exception for AI that merely assists or doesn't substantially alter the input data, and its reach is genuinely contested. A single power line probably falls under it. But the exception attaches to the behavior of the tool, not to the visible result — and so the finding stands: it isn't what happens in the image that determines the obligation, but what it happens with.
This isn't a criticism of the law. There's a good case for building it this way: the problem isn't the individual deception, it's the industrial scalability of it. But then you also have to say what disclosure fails to accomplish as a result.
And so the missing label carries less than people would like to load onto it. It doesn't mean: this is how it was. It means: no model generated the content here. That's a statement about the route, not about the destination.
For commissioned photography it remains a strong argument, because there route and destination coincide: whoever photographs real spaces, real products, real people can promise both. For my work they come apart. My images carry no label and still show no reality. They don't claim to.
This is exactly where the two-way split stops working.
The law recognizes two categories. There are three.
The disclosure requirement draws a line between AI and not-AI. That's sensible, and overdue.
What it doesn't draw is a line between photography and montage. And that second distinction is the one currently being forgotten.
A missing label doesn't answer the question people ask under my images. Anyone who sees a spare, deliberately ordered picture and instinctively thinks "prompt" won't be convinced by the absence of a symbol. The suspicion doesn't come from missing information. It comes from an entire genre of image-making having dropped out of collective memory.
That isn't a complaint. It's a diagnosis — and it doesn't only concern me.
Montage is older than the debate
The idea that a photograph shows a scene that took place exactly that way in front of the camera is younger and shakier than it sounds.
In the 1850s, Gustave Le Gray made seascapes by combining two negatives: one for the water, one for the sky. He did it out of technical necessity — the emulsions of the day couldn't hold the contrast range between bright cloud cover and dark water in a single exposure. Sky and sea didn't necessarily come from the same session, and sometimes not even from the same place.
In 1857, Oscar Gustave Rejlander assembled The Two Ways of Life from around thirty separate negatives. Henry Peach Robinson composited Fading Away from five the following year. Both were exhibited, sold, discussed — and yes, attacked. The question of whether a composited picture still counts as photography is exactly as old as the medium.
Montage isn't a break with photography. It's a branch from the same root, grown out of the same technical limits. What has changed since are the tools. What hasn't changed is the principle: an image is assembled from several exposures, and someone decides what belongs in it.
What's new is that this tradition no longer gets thought of. Before 2022, nobody would have mistaken a composite for anything other than a composite.
The better criterion
If "real or AI" is the wrong question, what's the right one?
I'd put it this way: who decided what is in this image?
For me, the answer looks like this. Every element that carries meaning in my work is a photograph, a vector graphic, or a 3D render — created by me, or by a person whose work I licensed. The sky is a sky that was photographed, in a place where a person stood — or a vector a person drew by hand. Selected from an archive that has grown over years. The mask is handwork. The position of every element, its size, the question of whether it gets to stay at all — those are decisions I make and answer for. Sky replacement happens manually here, not at the press of a button or klick.
Two photographs, one decision.
“Just you & me”, © Kathrin Federer, 2024
And because transparency is only worth something if it names the uncomfortable part too: yes, I use Generative Fill. With an empty prompt, to close edges, extend surface, or clean up remnants after a transformation. That is generative AI, and I won't pretend otherwise. In my work it produces surface, not content. No element anyone would see and name comes out of a prompt.
That's a hierarchy, not a gray zone. A tool is not the same thing as content, and I find that distinction more durable than any absolute claim.
The opposite pole isn't photography. The opposite pole is the fully generated image — a sky nobody saw, above a place nobody stood in, in an arrangement nobody decided. Not worse, not better, but fundamentally different. There the image is rolled. Here it is decided.
Why this concerns photographers just as much
At this point you might think it's my problem and not yours. I'd argue the opposite.
The Microsoft AI for Good Lab evaluated roughly 287,000 image ratings from more than 12,500 participants in an online quiz. The question was simple: real or generated? Accuracy averaged 62 percent, barely above a coin toss. One detail stands out: real photographs were taken for AI as well — specifically shots with unusual light.
In plain terms: photograph exceptionally well, and you come under suspicion. That's the same mechanism that hits my work, just at a different point on the scale. It isn't the fabrication being doubted, it's the skill.
There's a second thing that rarely gets said. Anyone who blends exposures, who stacks, who stitches panoramas, who drops in a sky from a second frame of the same session is already working compositionally. It just isn't called that. The difference between a classical montage and an elaborately assembled landscape photograph is one of degree, not of kind.
I don't say that to blur boundaries, but to draw them in the right place. The line doesn't run between photography and montage. It runs between images a person took responsibility for and images where nobody did.
On this side of the line, we stand together.
Made, not rolled.
The disclosure requirement will help contain obvious deceptions. What it can't do is teach a public that something exists between the shutter and the prompt.
That's our job, not the legislator's. And it begins with naming your own way of working — fully, including the parts that need explaining.
So the next time someone writes "AI?" under one of my images, the answer isn't no. It's: this is a montage. They've been around since 1857.
This article is general information and does not replace legal advice in an individual case.
Sources
Regulation (EU) 2024/1689 (EU AI Act), in particular Article 50, applicable from August 2, 2026: eur-lex.europa.eu/eli/reg/2024/1689/oj
Roca, T. et al. (Microsoft AI for Good Lab): How good are humans at detecting AI-generated images? Learnings from an experiment (2025): arxiv.org/abs/2507.18640
Official EU icons for labelling AI-generated content: digital-strategy.ec.europa.eu/de/policies/eu-icons-labelling-ai-generated-content