Here's an experiment you can describe in one line. Ask 25 language models for a metaphor about time, 50 times each, with the randomness turned up. You'd expect 1,250 attempts to cover some ground. They cover two ideas: time is a river, time is a weaver.

I went looking for results like that one because of something smaller. This autumn, friend after friend showed me the apartment they're about to build. Different cities, different budgets, different lives. The same room. Pale oak slats on one wall. A boucle sofa the colour of oat milk. An arch where a door used to be, linen curtains, one olive tree in a pot, a pendant lamp like a paper moon. By the third render I stopped looking at the picture and asked to see the prompt. Each of them had asked ChatGPT to show them their future home. It showed all of them the same one.

None of them knew. How would they? Each had seen exactly one render, their own, and it looked like taste. I was the only person in the story who had seen the set. From the inside, a default feels like a decision.

I build with these systems every day and I'm not going to stop. What follows is what I found when I went from the feeling to the measurements: what the models do alone, why the popular explanation is wrong, what happens when people use them, and where the money goes.

25 models, two metaphors

The full study threw 26,000 open-ended prompts at more than seventy models. A single model mostly repeated itself, and different models mostly repeated each other: cross-model similarity ran from 71 to 82 percent, and a typical cluster of fifty look-alike answers held the work of about eight different models. You can change your vendor. You'll get the same river.

None of this is new, only better measured. In January 2023 two researchers in Germany asked ChatGPT for a joke 1,008 times. More than 90 percent of what came back was the same 25 jokes, and one of them, about a scarecrow who won an award for being outstanding in his field, turned up 140 times. (Go and ask the chatbot in your other tab for a joke. I'll wait. If a scarecrow, a tomato or an atom is involved, you've just replicated a paper.)

The obvious reply is that people are no better. Ask a hundred humans about time and you'll hear about plenty of rivers. So a second team ran the comparison. Twenty-two models and 102 people took the same standard creativity tests. On the alternate-uses task, the one where you list things to do with an everyday object, the average distance between two humans' answers was 0.738. Between two models' answers it was 0.459. That gap is about two standard deviations, and a system prompt begging the models to "be very creative" closed less than half of it.

The detail I can't shake is that the models scored slightly higher than the people on individual originality. Any one answer was good, often better than yours or mine would have been. It was the set that was a monoculture.

You'd think there'd be a dial for this. Turn the temperature up, ask for something weird. But the Hivemind runs were already sampled hot. "Be very creative" bought back less than half the gap. A third team, comparing model-written book reviews with human ones, tried both tricks and barely moved their numbers. Temperature adds noise around the favourite answer. It doesn't change which answer is the favourite. To see why, you have to look at how these systems are finished, which is a different matter from how they're sampled.

The mechanism: an election, not an average

The picture everyone reaches for, and the one I used myself, is the bell curve. The model learns the distribution of human work and hands you its middle. Everyone uses the same curve, so everyone gets the same middle. It's tidy, it's wrong, and the way it's wrong is the whole story.

Try it on my friends' living room. If you really averaged every living room on the internet you'd get a brown smear. The mean of a red sofa and a blue sofa is mud, and the mean of a thousand floor plans is fog. What ChatGPT drew was nothing like that. It was crisp and specific, a room you could order from a catalogue by Friday. An average is a blur of everything, and this was one particular thing. It wasn't averaged. It was picked.

The picking happens late. Pre-training teaches a model the whole messy distribution of what people have made, every cluster, with the empty space between them. Then comes post-training, the stage where people rate outputs and the model is tuned toward the ratings. That stage is an election, and it's held under a brutal rule. A 2025 paper from Meta's FAIR lab wrote the rule down: for the standard preference objective, putting all of the probability on the single highest-rated answer is an optimal solution. Winner takes all. And when two answers tie, moving everything onto one of them is still optimal, because the loss never looks at what was thrown away. (Real systems add a regularizer that stops the collapse from being total. It's a brake. The engine still pulls one way.)

So the oat-milk sofa is the candidate that won, under a system that gave the winner every seat. Averaging had nothing to do with it. The runner-up rooms, the ones with green velvet and terrible wallpaper, are probably still in the weights somewhere. They just never get called on.

The part that made me sit up was the fix. The same authors changed the objective to reward answers for being good and rare, and got about 75 percent more variety in stories and up to 46 percent more in invented personas, while winning head-to-head quality comparisons about as often as before. Which means the sameness is a default, the kind that stays put until somebody has a reason to change it.

Other groups have reached the same place from other directions. Take five models and compare each with its "aligned" chat version on a test of conceptual variety: alignment narrows every one of them, and none gets near the human baseline. Change nothing but the chat template, the role markers wrapped around your prompt, and variety drops anyway. And there's a finding that explains why switching vendors doesn't help, under the grand name of the Platonic Representation Hypothesis (ICML 2024): as vision and language models get stronger, their internal representations converge on one another. Tolstoy would have recognised it. All strong models are alike; each weak model is weak in its own way.

The model is not averaging us. It is agreeing with itself.

So the honest version of my thesis is narrower than "AI averages culture," and I think worse: a handful of systems, finished by similar methods toward similar ratings, settle on the same winners, and then everybody draws their first draft from those winners. Tyler Cowen called his 2013 book Average Is Over. He was writing about the labour market. For the work itself I'd put it differently. The average isn't over. It's free, and nobody pays for what's free.

What happens when people use it

The study that got to me came out of Wharton in 2025, in Nature Human Behaviour. Five experiments on brainstorming with ChatGPT. In one, people had to invent a toy using a brick and a fan. With the chatbot's help the individual ideas got better, as they nearly always do. But 94 percent of them overlapped with somebody else's, and nine participants, working independently, gave their toy the same name: "Build-a-Breeze Castle."

Think about those nine people for a second. Each of them, as far as they knew, had named a toy. Each had seen one answer. Only the researchers saw nine.

Across the five experiments, the pool of ideas was narrower with ChatGPT in 37 of 45 comparisons. The most varied ideas in the whole study came from people using nothing but their own heads and a web search.

One lab could be a fluke. The meta-analysis is harder to wave away: 28 studies, 8,214 participants. People working with generative AI do somewhat better work than people working alone (effect size +0.27). The variety of ideas across those same people falls off a cliff (effect size −0.86, and in social science you rarely see a number that size with a minus sign in front of it). If you keep one sentence from this essay, I'd like it to be the next one.

The tool makes each person better and the group narrower.

That is also why it's so hard to notice. Your own work really did improve, and you have the evidence in front of you. The cost turns up somewhere you can't see from your desk. It turns up in the set.

You can hear it happening to the language. A 2025 paper in Science Advances went through 15.1 million biomedical abstracts looking for words whose frequency suddenly broke from its own trend. At least 13.5 percent of the abstracts published in 2024 carried a language model's fingerprints, and in some sub-fields 40 percent. "Delves" appeared 28 times more often than its history predicted. The authors counted 454 such excess words in 2024, against 190 at the height of COVID. By that measure a chatbot bent scientific English about twice as hard as a pandemic did.

Then it got out of the text box. Researchers at the Max Planck Institute analysed 824,634 podcast episodes, the unscripted kind, and found the same vocabulary climbing in human speech. ChatGPT's favourite words were rising 25 to 50 percent a year, and "delve" beat a synthetic control at p = 0.01. Nobody pasted those sentences. People said them, with their own mouths, because a model had said them first. Listen for it in your next meeting.

The pull has a direction, too. In a controlled study of 118 writers in India and the United States, the same autocomplete suggestions nudged the Indian writers toward Western conventions, in what they said and in how they said it, and gave them smaller productivity gains for their trouble. Autocomplete has an accent, and it probably isn't yours.

Now count the people in the room. About forty percent of American adults had used generative AI by late 2024. OpenAI claimed 800 million weekly users in October 2025. Ahrefs found machine-written text in 74 percent of newly published English web pages, though only 2.5 percent were machine-written from top to bottom. Radio, television and the feed all distributed culture after someone had made it. This is the first medium that hands out the first draft, to most of the people making things, at the same moment, from a handful of sources that agree with one another.

Your landing page, in indigo-500

The web is the easiest place to watch this, because the web can be crawled.

It was converging long before the models showed up. A CHI paper from 2021 measured how visually different website layouts were from one another and found the distance had fallen 43 percent between 2010 and 2019. Blame the frameworks. Bootstrap sits under three out of four sites whose CSS framework can be identified, and Tailwind is downloaded 176 million times a week. The interaction designer Boris Müller had the diagnosis in 2018, in one line: templates are content-agnostic, and that's the problem.

Then the generators arrived and the sameness grew a face. You know this page. Inter, or the system font. A hero section, centred, under a small pill-shaped badge that says something like "Now in beta." A gradient running from indigo to purple. Three feature cards in a row, icon on top. Rounded corners, a shadow at ten percent opacity, shadcn components straight out of the box. You may have shipped this page. I'm in no position to judge.

In August 2025 Adam Wathan, who created Tailwind, posted a half-serious apology. Five years earlier he had picked bg-indigo-500 as the default button colour in Tailwind UI. The default was copied into so many tutorials that the models learned it as what a button looks like, and now it's the complexion of the machine-made web: one person's offhand choice, laundered through a training set into the taste of a generation of startups. When a parody site of vibe-coded landing pages reached the Hacker News front page, the top comment was two words long. Mode collapse.

In April 2026 the joke became a measurement. A developer named Adrian Krebs pointed a headless browser at 1,590 recent Show HN launch pages and ran sixteen deterministic checks for exactly these tells. Twenty-two percent of the pages had four or more. Another 32 percent had two or three. (He puts his false-positive rate at five to ten percent.) So more than half of the startups launching on Hacker News that spring wore at least two fingerprints of machine-made design, and they wore them on the one page where a company tries hardest to look like nobody else.

The benchmark followed in October. Fifteen frontier models, eighty design briefs, 2,400 generated interfaces. Every model scored above 90 percent on appropriateness, so these are not bad pages. But on how distinct different models' answers to the same brief were, the models scored 0.59 where human designers scored 0.76. On creative range, meaning how differently one designer handles different briefs, the models managed 0.58 against the humans' 0.90. Fifteen different AIs, asked for fifteen different apps, produced work more alike than human designers given the same brief. I should say that the authors build UI-generation tooling, so hold the decimals loosely. The direction matches everything else here.

Each page is fine. That's the trouble. The set is one page.

It was already happening offline

Kyle Chayka gave the look a name in 2016: AirSpace, the chain of near-identical cafés, co-working lofts and Airbnbs, all subway tile and reclaimed wood and Edison bulbs, in cities that have nothing else in common. His evidence was anecdote. The market's isn't. In 2017 Zillow went through 32,000 listing photos and found that a "greige" exterior sold for about $1,500 more than expected. In 2023 it surveyed 4,700 buyers and found that a dark graphite kitchen added roughly $2,500 while a white one cost you $600. Six years apart the winning colour was grey both times, pale and then dark. People were being paid to converge.

Nobody chose greige. Greige is what's left when everyone stops choosing.

Scale handles the rest. IKEA has sold more than 140 million Billy bookcases, which is more bookcases than America has households. In American cities one paragraph of the 2000 International Building Code, section 510.2, allowed five storeys of timber frame on a concrete podium, and the "five-over-one" became the skyline of every boomtown: 450,000 apartments finished in 2023 alone, nearly all of them to rent. When computer-vision researchers built a system in 2012 to tell cities apart from single Street View frames, it learned Paris, London and Barcelona easily. It couldn't find anything distinctive to learn about American cities.

Cars are the cleanest record of taste we have, because every sale is logged and every colour is something a person paid for. In more than 22 million US vehicle sales, the four non-colours (white, black, grey, silver) went from 47 percent of model-year 1996 cars to 80 percent of model-year 2025 cars. In 1996 one new car in 28 was grey. Now it's one in four. Worldwide, the paint maker Axalta has had the same four at around 80 percent year after year, and its 2025 report puts them at 81. These are a supplier's figures, with the caveats that implies, but you can check them in any car park.

The numbers in one place

  • 90% of 1,008 · Requests for a joke answered with one of the same 25 jokes (Jentzsch & Kersting 2023)
  • 0.738 vs 0.459 · Mean distance between ideas, 102 humans vs 22 models (Wenger & Kenett 2025)
  • 2.8% vs 100% · Book-review sets at least 95% positive, humans vs GPT-4 (ICLR 2025)
  • +75% · Story variety recovered by a diversity-aware training objective (Meta FAIR 2025)
  • 37 of 45 · Comparisons in which ChatGPT narrowed the pool of ideas (Nat. Hum. Behav. 2025)
  • g = −0.86 · Variety of ideas across people working with AI, 28 studies, 8,214 people (Meta-analysis 2025)
  • ≥ 13.5% · 2024 biomedical abstracts with a language-model fingerprint (Science Advances 2025)
  • −43% · Visual distance between website layouts, 2010 to 2019 (CHI 2021)
  • 22% of 1,590 · Show HN launch pages with four or more machine-design tells (Krebs, April 2026)
  • 0.59 vs 0.76 · Distinctiveness of generated interfaces, 15 models vs human designers (2,400 UIs, 2026)
  • 47% → 80% · US cars sold in white, black, grey or silver, model year 1996 vs 2025 (22M sales)
  • +29% / −11.5% · Personalized podcast picks: streams vs each listener's variety (Spotify RCT 2020)
  • 10% → 50%+ · Fully AI-generated share of Deezer's daily uploads, Jan 2025 vs June 2026 (Deezer)
  • inverted U · Airbnb photo uniqueness vs occupancy, 481,747 photos (J. Consumer Res. 2025)

The case against all of this

An argument that can't say what would refute it is just a mood, so here is the best case against mine. It's better than I expected.

Fashion is the one domain someone has measured across more than 150 years of images, and it shows variety rising since the 1980s, with a crowd of niche looks replacing the single dominant one. Names went the same way: in 1955 American parents gave nearly a third of their sons a top-ten name, and by 2007 it was fewer than one in ten. National music charts on Spotify and iTunes drifted apart after 2017 when you'd have bet on them merging. People who take up streaming end up listening more widely than before. The genre map built from Spotify's own data counts more than 6,000 genres.

It gets closer to home. The collapse in human-AI work turns out to depend on the interface. When the model writes for you, variety falls. When it asks questions or offers options and the human stays in charge, quality goes up and variety holds; that's from a CHI 2026 study of more than a thousand people. Pay people for originality and the variety comes back, in two preregistered trials. The training collapse has a known fix. And the oldest result in this whole field, the MusicLab experiment of 2006, found that when its 14,341 participants could see each other's choices, hits became more unequal and less predictable. People herd, but where the herd ends up is close to a coin toss.

I accept all of it, and it narrows what I'm claiming. The free tier of everything becomes the same. The free tier is most of the volume. And the price of being visibly, legibly different from it goes up. Read that way, the counterexamples stop being exceptions. The niche label, the unusual name, the playlist somebody actually made: that's what the premium looks like when people pay it.

Where the rent goes

Back to the ladder. The Tail Premium says that when a technology compresses the middle of a distribution, the returns run to the ends. Everything above describes the compression, with one detail the first essay didn't have: the middle is crowded and it is also correlated. A million variations on one winner are one variation. So competence is no longer the scarce input, and "creativity" isn't quite it either, since the models do fine on individual originality. What's scarce is independence of source: an idea that didn't come from the same place as everyone else's.

The market already prices this, and the price has a shape. A 2025 paper in the Journal of Consumer Research scored 481,747 photos from New York Airbnb listings for visual uniqueness and matched them against occupancy. The curve was an inverted U. Listings that looked like every other listing were penalized. So were the truly strange ones. The best performers, by a margin worth thousands of dollars a year to a host, sat a measurable distance from the pack in a direction a guest could still read. That's the Tail Premium drawn as a curve: the rent goes to the room that is different on purpose and still makes sense, and the weirdest room does about as badly as the blandest. The grey car, seen this way, is a survey. It shows you where the margin went.

It also changes what a buyer can learn by looking. When artefacts converge, inspecting the artefact tells you less each year, and the question "who stands behind this, and what did they risk?" has to carry more of the weight. That's Human Signal. The number I expect to move is the Originality Premium, the price gap between work of verified origin and generated work in the same market, and it's the one I intend to measure.

What do you do about it on a Monday? I'm not going to tell you to stop using the tools. I won't, and you won't either. Using them costs you nothing. Accepting the first draft is the expensive part. So decide before you prompt: write down what you think, however badly, and only then ask. When you do ask for options, ask for twenty and bin the first five, because the first five are everybody's. And treat a model's defaults the way a decent cook treats a stock cube. Fine in an emergency. A confession on the menu.

Three bets

The rule of this series is that an essay with nothing staked is an opinion in a nice jacket. So, three bets. They go on piszczek.pl/predictions with a commit hash before any of the windows open, each with a threshold, a deadline, a source that isn't me, and the outcome that proves me wrong. I've put a confidence on each, because I'd rather be calibrated than brave. If I say 80 percent and turn out right every time, I was sandbagging.

  1. The flood doesn't recede. In Deezer's first disclosure of 2027, fully AI-generated tracks are still at least 45 percent of daily deliveries. They were 44 percent in April 2026 and passed half in June. Wrong below 40. Confidence: 80 percent.
  2. The car park stays grey. In Axalta's 2027 global colour report, white, black, grey and silver together still hold at least 78 percent, against 81 today. Wrong below 75. Confidence: 85 percent. This one is a bet that the convergence is structural, in a market with no model in the loop.
  3. The era gets named after the models. By the end of 2030 the name of an AI model or model family is in ordinary use as the name of a style: in a major dictionary, in a stand-alone Wikipedia article about that style, or in the title of a museum exhibition. Any one of the three counts. Wrong if none exists on 31 December 2030. Confidence: 65 percent.

The third is the one I care about, so let me say it properly. Every era ends up named after whatever shaped its surfaces. We named most of a century after a queen and a decade after a school of design. We named a palette after a film stock and a face after an app. I think the 2020s get named after model families, because that's where most people making things drew their first draft. And I don't only mean pictures. A way of furnishing a room, a register of prose, a kind of chorus, a kind of face. People already say a poster looks "very Midjourney." The bet is that the habit leaves the trade and reaches things no model ever touched, until one day somebody describes a real kitchen, in a real house, with the name of a model, and everyone at the table knows exactly what they mean.

We named a century after a queen and a decade after a school. We'll name this decade after a model family.

If these come in wrong, the essay was wrong, and I'll say so here in the same font size.

One ugly lamp

I don't know yet which of my friends will actually build the room in the render. Probably all of them, and I should be honest about why: it's a lovely room. That's how it won. It will photograph well, and nobody who visits will ever dislike it, which is a strange thing to be able to say about a place where a person lives.

So I'm not going to talk anyone out of the oak slats. If one of them asks, I'll say this, and I'll say it to you. Somewhere in the new place, keep one object that would lose every vote. Everybody owns one. The lamp from a grandmother's flat, in a colour that has no name. The chair that matches nothing. The painting a cousin made in a bad year. A model would never draw it, because no panel of raters would ever have agreed on it, and that's why it's worth more than the rest of the room put together. It's the only thing in there that couldn't have happened to anyone else.

Ask a machine what time is and you get a river. Ask it what home is and you get my friends' apartment. Tonight somebody, somewhere, is typing a warm, modern living room into a box, and the box is about to show it to them. They'll love it. They'll only ever see the one.

Scarcity never dies. It climbs. This time it climbed out of the product and into a smaller, odder question: whether the idea was yours before it was everyone's.

Keep the lamp.

So: signed.

Michał Piszczek