Sameness

Heilig-Geist-Spital, Nuremberg — the old almshouse spanning the Pegnitz on its double arches. 1995. Pentax MX, Kodak Ektachrome.

On why everything it makes starts to look the same — and what that costs.

During my working life I had the good luck to work with a few genuinely good coaches. One of them, years ago in Germany, was Jan de Zwarte. We’ve stayed in touch the way you do now — five thousand miles apart, a LinkedIn post here, a comment there — and a little while ago Jan wrote something that stopped me.

He’d asked an AI to transcribe some handwritten notes, and in the back-and-forth that followed, the model handed him a word he couldn’t put down: sameness. Not as an insult. As a diagnosis. He turned it over for a whole post — sameness as the quiet pull of these systems, the thing that flattens difference while everyone’s looking the other way.

I wrote back that it was brilliant, because it named something I’d been feeling for some time without a word for it.

Here’s the feeling he named. Like most of you, I guess, I can often tell when a piece of writing was made by a machine — not from any one mistake in the text itself, but because it just didn’t feel right. You sense it before you can justify it; call it a gut feeling. It’s easiest when you see the posts drowning in emoticons on LinkedIn, or the stock sentences that don’t quite land like a person talking. I assumed I was reacting to something overcooked — that bland corporate politeness these tools fall into. But that wasn’t it. What I was picking up on was sameness. Thank you again, Jan, for the word. It nailed it. And once you have it, you can’t stop seeing the thing.

So I went looking for what, exactly, I was sensing. This is what I found.

What sameness actually is

Sameness is not a side effect of these systems. It is what they are built to do. I don’t want to sound too technical or schoolmasterly about it, so in plain words:

A language model predicts the next word; an image model predicts the next pixel. In both cases the machine is reaching for the most likely continuation — and the most likely thing is, by definition, the least surprising one.

Does this make sense to you? You don’t believe it? Do that a few hundred times in a row and you don’t get a real voice. You get an average kind of text — very polished, very smooth. An average text, an average emotion, an average story. Sameness. Maximizing probability and regressing toward the mean are the same motion described in two vocabularies.

And it isn’t only that each output drifts to the middle. It compounds. Feed a model enough text and images that were themselves generated — and the internet is filling with exactly that — and the rare, the strange, the particular gets thinned out a little more with each generation, until the edges are gone and only the center remains. Researchers have a name for this: model collapse. I have a simpler one. Sameness is what the machine is for.

Here’s where I want to be careful, because there’s a lazy version of this argument and it’s wrong. The lazy version says: the internet is full of garbage, so garbage comes out. That’s comforting, because it implies better inputs would fix it. The unsettling truth is that sameness comes from the good material too — because the average of a thousand fine photographs is not a fine photograph. It’s the ghost of all of them, and the ghost belongs to no one. It’s flat, without emotion.

Elias at the lighthouse

I’m not the only one who went looking. A software engineer named Daniel May ran a small, clean test that says it better than any argument. He opened an AI model and typed the emptiest possible prompt: write a story in ten sentences. Then he did it again. Both times, word for word, the story began the same way — an old lighthouse keeper named Elias, polishing the brass, a storm coming in. Not similar stories. The same story, sentence for sentence.

Then he asked a different company’s model, built on entirely separate machinery, the same thing. Out came Elias again, at his lighthouse, with his fog. He ran a batch of eight models from labs that share nothing but the task — and again and again, the lighthouse keeper, sometimes a clockmaker, sometimes named Elias outright. He has a good phrase for it: the default basin. Ask for nothing in particular, and the machine rolls downhill to the same small set of safe, vaguely-literary, easy-to-like archetypes. The basin already has a name sitting in it, and the name keeps coming back.

I didn’t take May’s word for it. I opened ChatGPT and typed the same empty prompt — write a story in ten sentences — twice. Try it yourself; I’ll wait. Here’s what came back:

It didn’t even hand me the same story twice. It handed me the same world twice — the lonely keeper, the portentous clock, the twist saved for the last line like a held breath. Different words, identical furniture. I had wandered into the basin without being told it was there.

And it isn’t a fluke of one engineer’s afternoon, or mine. In May, two Cornell researchers sampled twenty thousand stories from four different large language models, all built from the prompt “tell me a story.” The name Elias turned up in more than a quarter of them. And over eighty-eight percent of the stories — eighty-eight — shared the same small handful of names, places and professions: Elias, the lighthouse, the keeper, the clockmaker. Their best guess at why is almost funny: told to avoid copyrighted characters and anything adult, the models back away into a tiny, safe corner of the imaginable — and then, because they learn from one another, they teach each other to stay there.

And now Elias has escaped the chat window. The Guardian ran a piece on him — “The curious case of Elias Thorne, and what he tells us about AI inbreeding.” The fictional keeper is turning up as the credited author of self-published books across whole genres on Amazon, in obviously machine-made YouTube videos, even in questionable health guides — one of them, a cancer handbook, ranking in the oncology categories where a frightened person might actually find it. That’s the moment sameness stops being a matter of taste. The danger was never one fake name. It’s that every surface that used to earn our trust can now be flooded with passable fakes faster than anyone can check them — the same average, copied outward, until the copies become the training data for the next round. Model collapse. AI inbreeding. Sameness, eating itself.

But May’s test had one honest limit, and he said so himself: an empty prompt invites an empty answer. Ask for nothing and you get the average of everything.

Isn’t that crazy?

So I wondered — what happens at the other end? What if you give it everything? All the details and context.

So I tested it on myself

As a film photographer I set the machine a harder problem than May’s. Not “hey, make a photograph.” I gave the machine everything. My Flickr, my name, my camera, my film stock, the light I work in, the kind of place I photograph. Everything particular. If the basin only swallows the empty prompt, a prompt this full should climb right out of it.

Guess what? It didn’t.

AI-generated — the experiment, not a photograph.

If the thing were actually intelligent, this is the moment it would push back — circle round and ask what I was really after, the way a person would. It didn’t. It just delivered. And what it delivered is competent. It would pass on a feed. It is also the exact statistical center of “moody black-and-white film landscape” — a wet path, a soft sky, a curve leading in. Not a place. The average of every place like it.

Now the next catchy thing. The prompt asked for two cameras at once: a 6×6 Pentacon frame and a 6×9 Ikonta — two different negatives, two different ways of standing in front of the world, two different ways of composing. A photographer would refuse, or choose, or better yet ask me what I’d been smoking and what on earth I meant. I even told it “no artificial enhancement” — the language of authenticity, typed into a machine that is nothing but enhancement. The machine noticed none of it. It simply averaged the impossibility into one smooth square.

AI-generated — same recipe, new subject.

It returned the same photograph — the same composition shifted onto a new noun. Same light, same frame, same tree, give or take. It had learned my name as a recipe and applied it to anything I named. But a way of seeing was never a recipe. Sameness. Average.

AI-generated — the control, my name removed.

Then I took myself out of it. I deleted my name and the Flickr link and asked for the same kind of scene. No difference — the unnamed frame is a sibling of the others. If anything it came back prettier, because now nothing held it back from relaxing all the way into the average. My name had been decorative the whole time. It was never painting me. It was painting the genre and letting me believe it was a portrait.

So here’s the bracket, and it’s my whole argument. May gave the machine nothing and got Elias. I gave it everything I am and got the genre mean. The basin catches the empty prompt and the overflowing one alike. There is no prompt rich enough to climb out, because climbing out is the one thing it cannot do — it can only fall toward the middle, because the middle is the only direction the math knows.

Then I asked it to shoot the rest of the roll

When you go out to shoot and you’ve loaded a roll, you finish the roll in most cases. So that was the next prompt: keep walking, shoot the rest of it, and give me the contact sheet. Same photographer, same walk, same light — the other ten frames.

AI-generated — ten frames, but not a roll.

What came back was ten variations on a theme — a tree, a path, some water, reshuffled. Pretty, each of them. But not one frame answered another. Nothing was rejected to make room for something better. There was no morning, no order, no reason this frame came before that one. It wasn’t a roll. It was a pile.

And that’s the difference the whole essay has been circling. A roll is a walk somebody took. The sequence is the meaning — it is the story being told, the same person carrying the same intention from frame to frame, deciding what each one has to say and where it falls. You can generate twelve pretty pictures in an afternoon. You cannot generate a roll, because a roll isn’t twelve pictures. It’s one person — present, composing, thinking about the story to tell — twelve times.

The same face on every answer

There’s a darker cousin to all this, and it follows from the same fact. Let me ask you straight: do you trust people who hallucinate? Then why do we trust a machine that does it constantly?

The easy answer is: you verify. You treat it like a brilliant assistant who now and then makes things up, and you check its work. Fair. But here’s what makes it worse than a person, not better. When someone is confabulating — making it up as they go — they usually give themselves away. A hesitation, a story too smooth, something in the eyes. You learn to read it. The machine gives you nothing to read. It delivers the invention in precisely the same confident, even voice it uses for the truth, because to the machine there is no difference between them — both are just the most likely next words. The fabrication and the fact wear the same face. That’s sameness again, in its most dangerous form: not boring this time, but indistinguishable. A human’s lie looks different from their truth. The machine’s doesn’t.

There’s a deeper version of this, and the encyclical names it better than I can. The machine treats every error as a defect to be smoothed away. A person doesn’t.

That’s the whole difference, really. A scratch on the negative, a frame I misjudged, a moment I almost missed — those aren’t failures to be optimized out. They’re part of the record of having been there. The machine has nothing to be wrong about, because it was never anywhere.

An older name for it

I didn’t expect to end up in a papal encyclical, but here we are.

This past May, Pope Leo XIV published his first, Magnifica Humanitas, on what artificial intelligence does to the human person. What struck me is that he reaches for the oldest story we have about exactly this. The Tower of Babel: a single people, a single language, a single technology, all bent on building something grand enough to make a name for themselves. We remember it as a story about pride. Leo reads it as a story about uniformity — a project, in his words, supported by “a uniformity that eliminated diversity and that chose homogenization over communion.” And the punishment fits with a terrible neatness: the reward for all that sameness was not unity. It was dispersion. Everyone talking, no one understood.

He puts it in an image I can’t shake. “In an ecosystem,” he writes, “balance is disrupted when one species expands at the expense of others.” When one faculty — intelligence, efficiency — is made the measure of everything, the rest withers: feeling, will, devotion, the slow work of relationship. A monoculture is poor precisely because it won. He has a name for the whole pattern — the “Babel syndrome”:

Against Babel he sets Nehemiah rebuilding Jerusalem — and the detail he lingers on is that the wall got rebuilt not by one voice but by many, each family assigned its own section to raise. A common language, he says, but not one of uniformity — one of communion.

That distinction is the whole thing. Sameness and communion both look like togetherness from a distance. Up close, one is a thousand voices saying the same word, and the other is a thousand voices saying their own.

Why I keep loading film

Writing by hand forces your thinking to diverge; the machine’s whole motion is to converge. The act of writing is the thinking. As Flaubert had it: the art of writing is the art of discovering what you believe.

Here’s how a piece like this actually starts: with a fountain pen, in a Leuchtturm notebook — the paper takes ink the way good paper should — the same pen I’ve written with for as long as I can remember. I write by hand first because it’s the same slowing down as the film. You can’t delete ink. You cross out, you keep going, you commit. It’s its own kind of digital detox — the thinking happens at the speed of the hand, not the cursor. Then I let an AI read my handwriting and sort the pages out. The seeing happened on paper; the machine just developed it. After that, I use the assistant to think out loud, to correct my grammar and my typos, to catch the places I’m being lazy or repeating myself. None of that is new to me, and none of it troubles me. In my professional life I worked with copywriters — and it’s the same when they’d debrief me.

Now, returning to where this started: what I believe is that we, as human beings, are unique, differentiated — and that, so to speak, is what makes the difference.

So that’s why I keep loading film. I don’t shoot to prove anything to a machine. I shoot that way because I’m tired of all the “perfection” — the empty images with no story and no emotion. And yes, I do it for myself, not for a living. I enjoy a real photographic essay. I enjoy all the Leuchtturm notebooks I’ve filled over decades. Because every frame is a place I actually stood and a decision I actually made, and the roll is the proof of it: twelve frames, no two the same, each one answering the last. That’s the one thing the basin can’t reach. It can average a thousand of my photographs into a smooth grey nothing. It cannot stand where I stood. It was never there.

The machine gave Jan a word for what it does to us: sameness. Here’s the one it can’t touch — presence. Human being makes the difference.

© 2026 Henry

I publish my photography under my middle name Henry — a small tribute to my father Heinrich, and his lifelong love of making photographs.