Design culture · 19 September 2026 · 12 min read

Retro is winning because it looks made

Vinyl, compact cameras, the twenty-year cycle, what the feed selects for, and why a label saying a machine made it costs a piece its credit: reading notes on a hunch that old-looking work is outperforming new-looking work, and on why it would be.

On this page

What I would bet on today

Nostalgia is a measurable premium, and a small one. Across 31 studies, nostalgic messages persuaded a little more than plain ones, an effect of about .19 (meta-analysis, 2023). People feeling nostalgic will also pay more (Lasaleta, Sedikides & Vohs).
The cycle is on schedule, and this is the 2004 to 2013 turn. The people who grew up inside Windows XP and the iPod are the adults buying now, which is why the glossy, bubbly look is back on time (Dazed on Frutiger Aero).
The feed does flatten, and the best evidence is at the listener, not the culture. On Spotify, listening that came through recommendations was less diverse than listening people found themselves (Anderson et al., WWW 2020). The bigger claim, that algorithms flattened everything, is still an anecdote.
People are buying the imperfect thing on purpose. Vinyl passed a billion dollars in the US in 2025, its nineteenth year of growth; fixed-lens camera shipments went from 1.8 to 2.4 million, bought for texture against the phone's clean look (RIAA, PetaPixel).
A label saying a machine made it costs the work its credit. Identical pieces were liked less and rated less beautiful when attributed to AI, even when every piece was AI-made (Bellaiche et al., 2023). My inference is that retro reads as made by a hand, and that this is the whole edge.

The rest of this page is how I got there, including the part of the hunch I still cannot measure.

Where this started

For about two years the work that travels in the design feeds I follow has been old-looking. Risograph grain, bubbly 2000s type, glossy glass buttons, film borders, brutalist forms, the Windows XP hillside. The clean, dark, gradient-and-glow future that was the default of 2019 gets less of the room every month.

The question I wrote down was whether that is taste, a real shift in what people want to look at, or the algorithm, a property of the feeds that show it to me. I wanted to know what the research says about each half, because if it is the second I should stop treating the feed as a measure of anything.

Short answer: both, in a loop, and a third thing I did not expect closes the loop. The wrong turn was assuming the two were alternatives.

Is it taste

Nostalgia is a thing with a literature

The psychology is older than the feeds. Nostalgia, in the research sense, is a bittersweet feeling about a personal past that reliably raises social connectedness and self-continuity, and Constantine Sedikides and colleagues have spent twenty years measuring what it does to behaviour. The finding a designer should know is from Lasaleta, Sedikides and Vohs in 2014: across six experiments, people made to feel nostalgic were willing to pay more for products, gave away more money, valued money less, and drew smaller coins when asked to draw one. The mechanism the authors found was connection. Feeling connected to people makes money feel less important, and nostalgia is a cheap way to feel connected.

In advertising the effect has been counted. A meta-analysis of 31 studies and about 5,800 participants put the effect of a nostalgic message on persuasion at r = .19: positive, reliable, and small. It was stronger for pleasurable products than useful ones, and stronger when there was a link between the product and the viewer's own memories. That last condition, a past the viewer actually had, is the one the design trend mostly ignores.

The cycle, and who is turning it

Aesthetic revivals run on roughly a twenty-year cycle, which is not mysterious: it is the gap between being a child inside a look and being an adult with money and a feed. The look that is turning now is the one that ran from about 2004 to 2013, named Frutiger Aero by Sofi Lee for the Consumer Aesthetics Research Institute after its typeface and its Windows theme: glossy, translucent, blue and green, optimistic about technology and the planet at once. Its revival started on TikTok in the early 2020s, about eighteen years in, right on the cycle.

Pinterest's 2026 predictions read the same way from the other end: searches for 1970s childhood toys up 125% and nostalgia toys up 225%, both under a trend it calls Throwback Kid. Pinterest's numbers are search growth on Pinterest, which is a measure of Pinterest. I would read them as direction.

Is it the algorithm

The claim

The case that the feed flattens taste is Kyle Chayka's, first in a 2016 essay called Welcome to AirSpace, which noticed that coffee shops from Brooklyn to Osaka had independently arrived at the same reclaimed wood, the same Edison bulbs and the same white tile, and then at book length in Filterworld in 2024. The argument is that recommendation systems reward what is already familiar, so makers make what the feed already likes, so the feed likes it more, and the result is a culture that is smooth, generic and everywhere at once.

It is a persuasive book and its critics have a fair point. The Reason review asks the question the book never answers: compared to what? Three networks and one radio dial were not a diverse culture either, and the book proceeds by anecdote where it needs a baseline. I think Chayka is right about the coffee shops. I do not think the book proves the mechanism.

The evidence that exists

The measurement I trust is smaller and cleaner. Anderson and colleagues at Spotify looked at what people listened to through recommendations against what they found on their own. Listening that came through the algorithm was less diverse. The recommendations worked better for narrow listeners than broad ones, and when people became more diverse over time, they did it by moving away from the recommendations and finding things themselves. That is the flattening claim at the level of one person, with the data to show it.

What it does not show is that the feed prefers old things. It shows the feed prefers near things: whatever is close to what already worked. And here the two halves of the question stop being alternatives. If the feed rewards the recognisable, then the most recognisable thing a maker can post is something the audience has already seen, ideally as a child. Nostalgia is not the feed's taste. It is what a feed that rewards familiarity will select for, from a population whose most familiar images are twenty years old.

Never in human history have so many people experienced the same things, the same pieces of content disseminated instantly through the feeds.

Kyle Chayka, Filterworld

What people are buying

The feed is one measure and it is a bad one. The better measure is what people pay for when nobody is recommending, and there the turn is not a mood at all.

The retro turn in money, in 2025. Industry figures: the records are US, the cameras worldwide and Japanese.
	2025	Source
Vinyl	Past $1bn on 46.8 million records, up 9.3%; the nineteenth consecutive year of growth; three times the revenue of CDs	RIAA
Compact cameras	2.4 million fixed-lens cameras shipped worldwide, up from 1.8 million; in Japan sales up 127% and 73% of units sold	PetaPixel, on CIPA and BCN+R
Kodak	24% of the fixed-lens units sold in Japan, on cheap models like the PixPro FZ55	PetaPixel, on BCN+R
Fujifilm X100VI	Routinely sold out two years after launch	PetaPixel

The camera story is the useful one, because the buyers say why. The Yomiuri Shimbun asked them: the buyers are mostly women in their twenties and thirties, and what they want is the price and the “textured image quality”, photographs like a film camera's without editing them, against the phone's processed, ultra-clean look. Nobody is buying a cheap Kodak because it is better. They are buying it because the phone's picture is perfect and perfect has become the default, and a default carries no information about the person who chose it.

That is a taste finding and an algorithm finding at once. The phone camera is the most optimised image pipeline ever built. Its output is the AirSpace of photographs. The retro camera is the reclaimed wood.

The third thing

What I did not have in the question was generative AI, and it turns out to be the part that makes the rest make sense.

The creator label effect is now one of the most replicated findings in the psychology of AI art, and a 2026 review gathers it up. Bellaiche and colleagues showed people identical images labelled as human-made or AI-made; the human label won on liking, beauty, profundity and worth, and it won even though every image in the study was generated. Millet and colleagues found the same artefact rated less creative and less able to produce awe once a machine was credited with it, and Ragot and colleagues found it rated lower on liking, beauty, novelty and meaning. The effect does not need a visible difference. It only needs a belief about the hand.

The same thing happens to words. In the note on AI chat, identical replies were worth less to people once they were labelled as a machine's.

Now put that next to what generated imagery looks like by default, as of 2026: smooth, lit from everywhere, symmetrical, rendered, futuristic. The 2019 dark-gradient-glow aesthetic is, to my eye, close to what a model produces when you do not tell it otherwise. So I think the futuristic look has acquired a new meaning in the last three years. It reads as unauthored. And the retro look, with its grain, its misregistration, its hand-set type and its period-correct mistakes, reads as the opposite: as evidence that a person was there.

This is the step the title rests on, so I should say which half of it is evidence. The studies above all used a label. None of them shows that people infer a machine from a style with no label in sight. The nearest support I have found is on detection: Högemann, Betke and Thomas (Frontiers in Artificial Intelligence, 2025) asked 104 people to tell generated photographs from real ones and to say how they decided. The cue they named most was texture: smoothness, overly polished surfaces, lines that were “too perfect”. So people do read a polished surface as a sign of a machine. They were right 63.7% of the time overall and only 29% of the time against the hardest of the models tested, and the authors expect the visible cues to keep fading. What nobody has tested, that I can find, is the rest: whether a grainy, hand-set, period-looking piece earns the credit the human label earns. That half is my inference.

That is my answer to why old-looking work is outperforming new-looking work. It is not that people want the past. It is that the past is currently the cheapest available proof of a hand, and the label effect says a proof of a hand is worth real credit. Retro is winning because it looks made.

Where this leaves me

The question was taste or the algorithm. The answer is a loop: nostalgia is on a schedule, the feed rewards the recognisable, the most recognisable thing is the schedule's output, and generative AI has made the alternative look like nobody was there. None of the three needs the others to be true. The first two have evidence behind them. The third is half evidence and half my inference, and I think all three are true.

Use the personal past, not the historical one. The advertising evidence is that a look the audience actually lived through does the work, and a generic vintage does not. Know which childhood you are quoting.
Retro is a reference, not a costume. The camera buyers want the texture, not a replica. Take the grain and the hand; leave the period drag.
Show the hand. Whatever the style, the thing the label effect rewards is evidence that a person made a choice. A visible decision beats a perfect surface.
Do not read the feed as taste. It measures what is near what already worked. Read what people pay for instead.
Expect the cycle to turn again. The next childhood in line is the flat, bright, rounded one of 2014 to 2020. Its revival is due around 2034, and it will look like a rejection of whatever the machines are making then.

Sources

The label effect

Will people embrace AI art? Deconstructing psychological barriers in human appraisal of AI-labeled artworks — Zhang et al., Computers in Human Behavior Reports (2026). Read for the label-effect literature; the findings quoted are Bellaiche et al. (2023), Millet et al. (2023) and Ragot et al. (2020), read separately in abstract
What you see is not what you get anymore: a mixed-methods approach on human perception of AI-generated images — Högemann, Betke & Thomas, Frontiers in Artificial Intelligence (2025). 104 participants; read for the cues people named and the accuracy figures
Where does the bias begin? Cognitive processing differences in the aesthetic evaluation of AI-generated art — Frontiers in Psychology (2026)

Last updated 19 September 2026

The other questions

Back to the notes Discuss this note