I keep seeing the same page. A tiny label in capital letters sits above the headline. Under the headline sit four boxes marked 01, 02, 03 and 04. Each image has a caption, and under the caption sits a description of the image. I close the tab and open a different company’s site. I find the same page again.
Slide decks look the same way. Every slide has its little eyebrow line above the title. Blog posts open with a hook, announce what they will cover, and close on three takeaways. I read the headline. I skim the first paragraph. By the third paragraph I am gone. You probably are too. I think we leave for the same reason. Nothing on the screen made us feel anything.
I have started calling this the gentrification of experience. A neighborhood gentrifies one renovation at a time. No single owner plans the result. Years later the corner bodega, the dive bar and the odd little bookshop are gone, and every block looks like every other block. Our screens are going through the same process, and generative AI just turned up the speed.
Nobody leads the flock
My first picture of this was a flock of birds. Thousands of starlings twist across the sky in waves and shapes. I imagined a kind of middle bird at the center, the Pareto-optimal average that everyone else follows.
That picture is wrong, and the way it is wrong taught me something. In 1987 Craig Reynolds built a simulated flock he called boids. His birds had no leader. Each one followed three local rules. Avoid crowding your neighbors. Match their heading. Stay close to them. A convincing flock appeared anyway. Later field work on real starlings by Ballerini and colleagues found that each bird tracks about six or seven of its nearest neighbors. Each bird ignores the rest of the flock.
So nobody gives the orders. The middle of the flock emerges because every bird copies the few birds closest to it. Design works this way now. A designer opens the last ten landing pages she admired. A developer installs the framework everyone else uses. A founder asks for the look of the startup down the street. Each choice is local and sensible. Added together, they pull everything toward one shape.
Researchers have measured this. Sam Goree and his colleagues at Indiana University ran computer vision over years of archived websites for a CHI 2021 paper. They found that designs became significantly more similar after 2007. For page layouts, the average distance between sites dropped by more than 30 percent. The use of a few shared code libraries was a strong predictor of that similarity. Their own project page is a hand-built Neocities site with loud colors, which made me smile.
Here is the uncomfortable part. The middle bird is us too. A 2025 paper on mode collapse in language models traced part of the problem back to the people who rate model answers. Those raters favor text that feels familiar, fluent and predictable. Models are tuned on those ratings. So the models learn our taste for the familiar and hand it back to us at scale.
Hexagons nobody planned
Bees build perfect hexagons. They never sit down and decide on hexagons. Mathematician Thomas Hales proved in 1999 that the hexagon is the most efficient way to divide a flat surface into equal cells. A 2013 study by Karihaloo and colleagues reported that bees build roughly round cells. The warm wax then flows, and the round cells settle into rounded hexagons. Physics finishes the job.
I find that a comforting and a scary idea at the same time. Convergence is what efficient systems do when nobody steps in. Blogs always followed a formula. Magazines had house styles. Sites in the late 1990s leaned on the same HTML table layouts. The pull toward one shape has always existed. The difference today is how fast the wax flows.
The old statistics of the middle
Francis Galton noticed something odd in 1886. Very tall parents tended to have children who were tall, but less tall than the parents. Very short parents had children closer to average too. He called it regression towards mediocrity. Today we call it regression toward the mean.
Large language models are built to produce the most likely next word. The most likely word is the typical word. A 2025 perspective paper on variance reduction in generative AI describes this as regression toward the mean in model output. The spread of answers shrinks compared with the spread in the real world, unless the prompt feeds the model rich, specific detail.
The effect compounds. Shumailov and colleagues showed in Nature in 2024 that models trained on model-generated content lose the tails of the original distribution. The strange, rare and surprising material disappears first. Writers feel this effect as well. In a study published in Science Advances, Anil Doshi and Oliver Hauser gave some writers story ideas from an LLM. Those stories were rated more creative one at a time. As a group, they were more similar to each other than stories written without AI help.
Noam Chomsky made the deeper version of this argument in a 2023 New York Times essay. His view is that these systems predict without understanding. They describe what is likely. They cannot explain why, and they cannot judge what should be.
I would add one thing. The best creative choices often look irrational. A designer picks a typeface because it feels like her grandfather’s workshop. A writer breaks a rule because the sentence sounds better broken. That choice comes from feeling, and the data offers no reason for it. A model has no grandfather’s workshop. Left alone, it will choose the middle every time, because the middle is the safest prediction.
AirSpace
Kyle Chayka saw the physical version of this a decade ago. In a 2016 essay for The Verge called Welcome to AirSpace, he described coffee shops from Odessa to Seoul with the same raw wood tables, exposed brick and hanging Edison bulbs. Nobody mandated the look. Thousands of owners made the same independent decisions while looking at the same apps. Chayka used the word gentrification for it.
Just as a gentrifying neighborhood starts to look less diverse as buildings are renovated and storefronts replaced, so economically similar urban areas around the world might increasingly resemble each other and become interchangeable.
Kyle Chayka, Welcome to AirSpace, The Verge, 2016
Branding went the same direction. Critics gave it the name blanding. Logos lost their serifs and their quirks. Illustrations became the flat purple people with long limbs that everyone recognizes from Corporate Memphis.
.png){kind=link}
Interfaces followed. Bootstrap gave developers the same grid.
Material Design gave apps the same cards. The SaaS home page settled into a headline, two buttons and three feature boxes.
Now generated pages add the numbered blocks and the eyebrow labels on top.
The body knows
Here is my test for whether content works. Watch the reader’s face.
When something lands, the body reacts. Someone nods along because you said exactly what they have felt and never put into words. Someone smiles. An eyebrow goes up. That small physical response tells you the piece reached a person. Without it, people leave.
The data on leaving is grim. A 2016 study by Gabielkov and colleagues found that 59 percent of the links shared on Twitter in their dataset were never clicked. People shared the headline and never read the story. Jonah Berger and Katherine Milkman found that content evoking high-arousal emotions such as awe, anger or anxiety spread further than content that left people flat.
I don’t think short attention spans are the whole story. People will read for an hour if something holds them. They quit by the third paragraph when the third paragraph looks like a thousand other third paragraphs. Readers sense the template. Sameness is the signal to leave.
Same bottles, different bars
Walk down any busy street at night. Most bars stock the same bottles from the same distributors. The vodka is the same vodka. Yet one bar is a speakeasy behind an unmarked door. One has a magician working the room on Thursdays. One is a dark EDM club with a line around the block. Each one serves a specific crowd, and each crowd would feel out of place in the other two.
Homes work the same way. Some people want a castle. Some want a Frank Lloyd Wright house that grows out of the hillside. Some want a small apartment in the middle of the city. Some want a farmhouse with nobody around for a mile. They all use the same plumbing, the same wiring and the same building codes. Nobody confuses the castle for the farmhouse.
This is how I think about heterogeneity on top of shared patterns. Keep the plumbing. Shared conventions help people. A menu that sits where readers expect it, buttons that look like buttons and pages that load fast on a phone all belong in the common stock. Frameworks and models are fine tools for that layer. Readers should never have to relearn how a door works.
Then build the room for one crowd. Decide who the speakeasy is for before you pick the furniture. Make the choices a model would never make, because a model has no idea who is walking through your door. Use the strange detail from your own life. Cut the eyebrow label and the numbered boxes, unless your audience truly needs them. Write the sentence the way you would say it to one person across the table. When you do use AI, feed it the specifics only you know, and ask it for a range of options instead of one safe answer. The research on mode collapse suggests that asking for a spread recovers some of the variety that one direct request loses.
When the artisanal comes back
Some of it never left. Brutalist websites, the IndieWeb, Neocities and the small web have been quietly building odd, personal, hand-made corners the whole time. Creative Review has written about designers bringing serifs back and leaving Corporate Memphis behind.
My guess is that the bigger return starts now, for a simple reason. When anyone can produce the average page in thirty seconds, the average page stops meaning anything. It no longer signals effort, care or taste. The handmade thing becomes the signal again. The same thing happened to bread, to coffee and to beer. Mass production made the default cheap, and then people went looking for the bakery with a name on the door.
I want to go deeper than one post, so I’m turning this into a research project. I plan to collect screenshots of the most popular interfaces from each wave, from GeoCities through Web 2.0 gloss, iOS 6 leather, iOS 7 flat, Bootstrap, Material and today’s generated templates. Then I want to measure how fast they converged, and test how readers respond to work built for them compared with work built from the template.
For now, I’ll leave you with the bars. Tomorrow night every one of them will pour from the same bottles. Most people will still walk past the bars that look like everywhere else. They will go looking for the unmarked door, the magician at the corner table and the room where the bass hits them in the chest. They want the place that feels like it was made for them. I want to build more of those places, on the web and on the page. If you nodded at any point while reading this, you already know why.