What if all the world's design were done by AI?

Experiments say AI makes each design a little better and all designs a little more alike. Run that for a few model generations and the maths of model collapse shows where the world's visual culture would end up.

A wooden Jacquard loom with a long chain of punched cards that control the woven pattern, at the National Museum of Scotland
Photo: Stephencdickson, CC BY-SA 4.0, via Wikimedia Commons

If all the world's design were done by AI, the first effect would be good: more design, made faster, and on average slightly better, which is what controlled experiments find. The second effect would be sameness. The same experiments show AI-assisted work becoming measurably more similar, a study of 4 million artworks found average novelty falling after people adopted AI, and the mathematics of model collapse shows that models trained on their own output steadily lose rare styles. The world would end up with competent, interchangeable design, and human taste would become the scarce input.

Key takeaways
  • Each piece gets a little better: AI ideas raised judged novelty by up to 8.1% and usefulness by up to 9.0%.
  • The group gets more alike: AI-assisted stories were about 9 to 11% (of range) more similar.
  • On one platform of 4 million+ artworks, output rose about 25% while average novelty fell.
  • Model collapse: models trained on their own output lose rare styles first.
  • Fresh human-made data and human selection are what keep variety alive.

This is a tier 1 what-if: extreme, but already partly happening. Adobe said in April 2025 that its Firefly models had generated more than 22 billion assets in under two years. So the question is not whether AI can design. It is what happens to the whole system when AI-made design becomes most of what people see, and most of what the next AI learns from.

I'll take it in three steps:

  1. the individual designer;
  2. the platform;
  3. the feedback loop.
22 billionassets generated with Adobe Firefly models by April 2025 (Adobe)
+8.1%novelty of stories written with five AI ideas (Doshi and Hauser, 2024)
4 million+artworks analysed for the effect of AI adoption (Zhou and Lee, 2024)
40–50%share of repeated prompts on a big image platform (De Rosa Palmini and Cetinic, 2025)

Does AI make each piece of design better?

Yes, modestly, and mostly for the people who start weaker. In the cleanest experiment so far, AI ideas raised judged novelty by up to 8.1% and usefulness by up to 9.0%, with the biggest gains for the least creative participants.

Generative AI enhances individual creativity but reduces the collective diversity of novel contentAnil R. Doshi and Oliver P. Hauser · Science Advances · 2024

293 writers wrote short stories in three conditions: alone, with one AI-generated starting idea, or with up to five AI ideas to choose from. 600 evaluators made 3,519 ratings. Compared with writing alone, five AI ideas raised novelty by 8.1% and usefulness by 9.0%. For writers who scored low on a divergent-thinking test, the gains were 10.7% and 11.5%. But stories in the AI conditions were more similar to each other (by about 9 to 11% of the human-only range) and about 5% more similar to the AI's own ideas. The limit: short stories, one model generation, and online participants, not professional designers.

Read the paper (free)
What AI ideas did to stories, compared with writing alone
See the data as a table
One AI idea (%)Up to five AI ideas (%)
Novelty5.48.1
Usefulness3.79.0
Similarity to other stories in the group10.78.9

Source: Doshi and Hauser, Science Advances 2024 · Novelty and usefulness are percentage increases over the human-only group. Similarity is the increase expressed as a share of the human-only range.

Read the chart as a trade. Each writer gains; the group loses spread. That is the social dilemma the authors name in their title, and it maps directly onto design: every brand that uses the same tools for its logo, deck or packaging gets a slightly better result and a slightly more generic one.

I covered the visual version of this, where corporate marks all drift toward the same flat sans-serif, in why logos look dated.

The story

In July 2024, soon after Figma launched a generative feature called Make Designs, Andy Allen, founder of NotBoring Software, asked it for a "not boring weather app". What came back, again and again, looked like a near-copy of Apple's built-in Weather app. Figma's chief executive Dylan Field temporarily disabled the feature. He denied that it had been trained on Figma content, community files or app designs; it ran on off-the-shelf language models paired with design systems Figma had commissioned, and he said more quality checks could have caught the problem. Allen accepted the explanation as plausible. The episode is the whole of this post in miniature: ask a machine for something unusual and it hands back the most common answer it knows.

The Register: Figma pulls AI design feature after it copies Apple's Weather app (3 July 2024)
"Like the fast food of creativity."Andy Allen on generative AI, to The Register (July 2024)
The glass-walled workshop wing of the Bauhaus building in Dessau, Germany
The Bauhaus in Dessau, 1926. Modernism already showed how one strong idea about design can spread worldwide and make cities look alike. AI could do the same, faster and without a manifesto.Photo: JensKunstfreund, CC BY-SA 4.0, via Wikimedia Commons

What happens when a whole platform switches to AI?

When a whole platform adopts AI, output jumps, the best work gets a little more novel, and the average work gets less novel. That is what a study of more than 4 million artworks found.

Generative artificial intelligence, human creativity, and artEric Zhou and Dokyun Lee · PNAS Nexus · 2024

The authors tracked over 4 million artworks by more than 50,000 users on a large art-sharing site, comparing about 5,800 artists who adopted text-to-image tools with those who hadn't yet. Adoption raised output by about 25% overall, with roughly a 50% jump in the month of adoption. Favourites per view rose from about 2% to about 3% within six months. But average content novelty and average visual novelty both declined over time, while the peak novelty of the best ideas rose slightly. The authors read this as artists converging on a preferred style while exploring ideas through the tool. The limit: one platform, observational data, and novelty measured by image embeddings rather than human judges.

Read the paper (free)

A 2025 study of more than six million prompts on the CivitAI platform found the same squeeze from the user side:

  • 40 to 50% of submissions were repeated prompts;
  • prompt wording became more uniform as communities grew;
  • more similar wording produced measurably less diverse images.

So the sameness is not only inside the model. It comes from people copying what already works, a loop design culture has always had, now running at machine speed.

LevelWhat the evidence showsSourceDirection
One person's workNovelty +5.4 to +8.1%, usefulness +3.7 to +9.0%Doshi and Hauser 2024Better
Group of people using the same toolOutputs about 9 to 11% (of range) more similar to each otherDoshi and Hauser 2024Less varied
Whole platformOutput +25%, favourites per view 2% to 3%, average novelty downZhou and Lee 2024More, liked more, less novel on average
Prompt culture40–50% repeated prompts; lexical sameness predicts visual samenessDe Rosa Palmini and Cetinic 2025Less varied
Model trained on its own outputTails of the distribution vanish over generationsShumailov et al. 2024Collapse toward the mean

What happens when AI learns from AI-made design?

When generative models are trained on data produced by earlier models, they progressively forget the rare parts of the original distribution, an effect called model collapse. In a world where all design is AI-made, every new model would be trained mostly on AI output, so this is the mechanism that decides where the experiment ends.

The Curse of Recursion: Training on Generated Data Makes Models ForgetIlia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, Ross Anderson · arXiv 2023, published in Nature 2024 as "AI models collapse when trained on recursively generated data"

The authors trained Gaussian mixture models, variational autoencoders and a language model on data generated by the previous generation of the same model, and repeated the cycle. In every case the tails of the original distribution, the rare and unusual examples, disappeared first, and the models drifted toward a narrow, over-confident version of the data. In the language-model test (OPT-125m), training only on generated text worsened perplexity on the original test set, the paper reports, from about 20 to 28 points; runs that kept a random 10% of the original data each generation degraded less. They argue this is a general property of learned generative models and that original human-made data becomes more valuable as synthetic data floods the web. The limit: the experiments are small and simplified.

Read the paper (free)
The loop that decides the endingWhat people seeTraining dataNext AI modelAI-made design(share grows)human-made design enters hereEach lap, rare styles that were not sampled cannot come back out.
If AI makes most of what people see, most of the next model's training data is AI output. Only fresh human work entering the loop adds back what earlier rounds dropped.

You can see why with a toy model. The working below is my own simplification of the mechanism, not a calculation from the paper.

  1. Assume the world's design styles form a distribution, and each model generation is trained on \( M \) examples sampled from the previous generation's output. A rare style, say a regional hand-painted lettering tradition, has share \( p \).
  2. Chance a rare style is missed in one generation. If it never appears in the \( M \) samples, the next model cannot learn it:
    \[ P(\text{missed}) = (1-p)^M \approx e^{-pM} \]
    With \( p = 0.001 \) and \( M = 1{,}000 \), \( P \approx e^{-1} = 0.37 \). One time in three, the style is gone after a single round.
  3. Chance it survives \( n \) generations (treating each round as independent, which is optimistic, because survivors also lose share):
    \[ P(\text{survives } n) \approx \left(1 - e^{-pM}\right)^n = 0.63^{10} \approx 0.01 \]
    After ten rounds, a 1-in-1,000 style survives about 1% of the time.

    Calculated in this post with these example inputs (a style with a 1-in-1,000 share, 1,000 training examples per generation, 10 generations), as a toy model of the mechanism in Shumailov et al.; it is not a figure from that paper.

  4. Spread shrinks too. Fitting a Gaussian to \( M \) samples by maximum likelihood gives an expected variance of \( \frac{M-1}{M}\sigma^2 \), so after \( n \) refits:
    \[ \mathbb{E}[\sigma_n^2] = \sigma_0^2 \left(\frac{M-1}{M}\right)^n \]
    With \( M = 100 \) and \( n = 100 \), \( 0.99^{100} \approx 0.37 \): about a third of the original variety remains, on average.

The real numbers are far larger (billions of samples, not a thousand), but the shares of rare styles are also far smaller, so the product \( pM \) is what matters. Anything that appears less than about once per training sample is at risk every round.

Model collapse: the tails go firstrare stylesrare stylescommon stylesgeneration 0 (dashed)generation 5generation 10Schematic, after Shumailov et al.: the spread shrinks and unusual examples vanish.
Each refit is trained on samples of the last, so the rare edges are missed and never relearned. Expected variance falls by a factor of (M−1)/M per generation in the Gaussian toy model.

Toy model: does a rare design style survive recursive training?

Stacks of punched pattern cards used on an Indian hand loom for weaving silk saris
Punched pattern cards for silk sari weaving in India. Rare regional designs like these appear in few training images, which is exactly where model collapse bites first.Photo: McKay Savage from London, UK, CC BY 2.0, via Wikimedia Commons

Which designs would disappear first?

The first designs to fade would be the rare, local and hand-made ones: regional scripts, folk motifs, unusual layouts, and anything that appears in few training images. Common corporate patterns would survive and multiply, because they sit in the fat middle of the distribution.

In India that list is easy to imagine:

  • the hand-lettered signboards of small-town shops;
  • the dense colour logic of truck art;
  • the many less-digitised Indian scripts beyond the big two or three, whose type design I explain in how Devanagari fonts work.

A 2024 study of Stable Diffusion, by AlDahoul, Rahwan and Zaki, already found the model homogenising how it drew groups of people, for example showing nearly all Middle Eastern men as bearded and in traditional dress. Visual styles face the same pressure, because the model reproduces what is common and smooths away what is rare.

Would anyone notice?

Most people would notice only slowly, because each individual design would look fine, often better than before. The loss is collective: the range of the whole visual culture narrows, and you see it only when you compare the whole set with what it used to be.

The Wundt curve

This is where an old idea from psychology helps. Wilhelm Wundt proposed in 1874 that pleasure peaks at a moderate level of stimulation, the "Wundt curve", and Daniel Berlyne extended it to art in 1971: too little novelty or complexity bores, too much unsettles.

The evidence for Berlyne's inverted U is mixed; Marin and colleagues (2016) found it for beauty ratings of paintings once familiarity was controlled, but not for every kind of stimulus.

AI tools are trained to produce what people rate well, which tends to sit at the comfortable middle of that curve. A world of only AI design would be a world tuned to the middle of the curve, everywhere, at once. Pleasant, efficient, and slowly boring. Here I am applying a theory, not citing a measurement of it.

The painted back of an Indian truck with bright decorations and the words Blow Horn
Indian truck art breaks every rule of the comfortable middle: loud, dense, local. A design system tuned to average ratings would rarely produce it.Photo: Justin Morgan from Richmond, Virginia, USA, CC BY-SA 2.0, via Wikimedia Commons

What would keep design alive in an all-AI world?

Three things keep variety alive: fresh human-made data in every training round, people who choose deliberately against the average, and tools that reward exploration over repetition. All three are already visible in the research.

  • Keep human work in the training mix. Shumailov and colleagues found that preserving original data slows collapse, which makes human-made, well-labelled design archives more valuable, not less.
  • Treat taste as the job. Doshi and Hauser's gains came from people choosing among AI ideas. The designer's role shifts from making to selecting and directing, which is the subject of what changes when you design with AI and the practical AI design workflow guide.
  • Prompt against the average. The CivitAI study suggests repeated wording drives repeated images. Writing unusual, specific prompts and feeding in your own references widens the output.

The same logic shows up in marketing, where cheap AI creative floods feeds and trust drops; I looked at that in the real cost of AI slop ads. On my own work, the principle is the one on the applied AI page: AI should cut repetitive production and raise a team's effective capability.

Verdict: possible, with a slow loss of variety

An all-AI design world is technically possible today, and on day one it would look better on average: experiments find up to 8.1% more novelty and 9.0% more usefulness per piece. The cost is cumulative. Outputs grow about 9 to 11% (of range) more alike, platform-wide novelty falls, and models trained on their own output lose rare styles: in the toy model, a 1-in-1,000 style has a 37% chance of vanishing each generation. Keeping fresh human work and human selection in the loop is the change that makes it sustainable.

The 8.1%, 9.0% and 9 to 11% figures come from Doshi and Hauser (2024). The 37% is calculated in this post with these example inputs (a 1-in-1,000 style, 1,000 training examples per generation): e^-1 = 0.37, a toy model of the mechanism in Shumailov et al., not a forecast.

What if next?

Questions people ask

Will AI replace graphic designers?

The evidence so far shows AI changing the job more than removing it: in experiments, AI raises the average quality of output, especially for less experienced people, while the selection and direction of ideas still comes from humans. The skill that gains value is taste, meaning choosing what not to make.

Why does AI-generated design all look the same?

Because models reproduce the most common patterns in their training data, users repeat prompts that already worked (40 to 50% repeats on one big platform), and people pick outputs that sit at the comfortable middle of what audiences like. Each of these pulls toward the average.

What is model collapse?

Model collapse is the loss of rare features when a generative model is trained on data made by earlier models. Shumailov and colleagues showed it in Gaussian mixtures, autoencoders and a language model: the tails of the distribution vanish first, and outputs narrow over generations.

Is AI-assisted art liked more or less by audiences?

On one large art platform, favourites per view for adopters rose from about 2% to about 3% within six months of adopting AI, so audiences liked the work more on average, even as its average novelty fell.

Sources

  1. Doshi and Hauser: Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances (2024) (source for the novelty, usefulness and similarity figures)
  2. Zhou and Lee: Generative artificial intelligence, human creativity, and art, PNAS Nexus (2024) (source for the 4 million artworks, +25% output, favourites 2% to 3%, novelty trends)
  3. Shumailov et al.: The Curse of Recursion, arXiv; published in Nature 2024 (source for model collapse and the vanishing tails)
  4. De Rosa Palmini and Cetinic: Exploring Language Patterns of Prompts in Text-to-Image Generation, arXiv (2025) (source for the six million prompts and 40–50% repeats)
  5. AlDahoul, Rahwan and Zaki: AI-generated faces influence gender stereotypes and racial homogenization, arXiv (2024) (source for the homogenisation example)
  6. Adobe newsroom: Adobe revolutionizes AI-assisted creativity with Firefly (2025) (source for the 22 billion assets)
  7. Marin, Lampatz, Wandl and Leder: Berlyne revisited, Frontiers in Human Neuroscience (2016) (source for the Wundt curve, Berlyne's theory and the mixed evidence)
  8. The Register: Figma's Make Design feature pulled after copying Apple's Weather app (2024) (source for the Figma story and Andy Allen's comments)
AI designgenerative AImodel collapsehomogenizationcreativity researchAdobe Fireflydesign tastevisual diversity
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Artblog.

Keep reading in Artblog

From the other blogs