All Roads Lead to the Cathedral: How AI “Flattens” Creativity (and How to Defend Against It)

2026-04-102026-04-09

Artificial intelligence can make beautiful things in 10 seconds. And that's precisely why it can be dangerous: it's easy to mistake aesthetics for originality, and speed for progress. When AI starts feeding AI, and recommendation algorithms begin feeding us—the world becomes more and more level, more predictable, and... less creative.

There are ways to counteract this, but first you must understand the mechanism: AI naturally converges on attractive, safe clichés. And if I let it take the lead, I’ll start converging too.

An Experiment Ending in Kitsch: 100 Iterations and Suddenly a “5-Złoty Warehouse”

Researchers conducted an experiment that seems simple, but reveals a lot about us (and about AI) [source].

They built a loop: text → image → text → image. First, a model generated an image from a prompt, then another model described the image, and this description became the next prompt.

There were 100 iterations of this loop—not just once, but at scale: 700 trajectories (with 100 trajectories for each of 7 temperatures: 0.1, 0.3, 0.5, 0.7, 0.9, 1.1, 1.3). The starting points were not random: they generated 100 diverse prompts.

Intuitively, since starts differ and there's randomness, the finals should spread out. The opposite happens.

After 100 steps, the systems converge to nearly the same motifs, directly labeled as “visual elevator music”—an aesthetic that is safe, commercial, predictable. Cluster analysis after collecting results across temperatures produced just 12 dominant motifs.

These “attractors” repeat like a boomerang: - gothic cathedrals, - palace interiors and “pompous interior design”, - lighthouses (often “stormy lighthouses”), - urban night scenes, - formal interiors, - pastoral villages and “pastoral scenes.”

This is not the chaos from the early days of generators, when AI would “mess up faces.” This is something worse: smooth mediocrity.

In the same experiment, there were cases where the initial prompt was about a person (e.g., a man packing for a trip) and after several dozen iterations the person disappeared from the description. What remained was a motif that is statistically stable.

This resembles what we see daily: “office” images from LinkedIn, the same faces, the same interiors, the same gloss. Even texts are stamped with an overuse of emojis—as if someone spent time matching 28 icons to a paragraph.

The mechanism is simple: models are shaped by the data that dominates the Internet. And the Internet rarely rewards difficult, niche, and weird things. It rewards what's common, recognizable, and safe.

The Child Draws a Ferrari, AI Draws a Golf: Where Creativity Lives

Here’s a great metaphor: if I ask AI for a “car”, and it sees the world through a street full of gray compacts, it will draw a statistical Golf/Mazda/Opel. If I ask a child—especially a boy with Hot Wheels—I’ll get a red Ferrari, a dream car.

That’s the crux: creativity lives in the outliers, in the tails of the distribution.

A very practical test: AI by default reduces variance, humans can increase it.

Where It Comes From: AI Is Shaped by the Image of the Internet (and the Internet Has Its Obsessions)

AI is trained on incomprehensibly large datasets, but that does not mean it absorbs the world like a human does.

A human gathers experience “lazily and on-demand” — when needed: to pass a quiz, survive, communicate, do something real in their environment. AI is fed wholesale: “here, have hundreds of thousands of examples, learn from them.”

If you trained a human the way AI is trained, it would be torture: 100 years of viewing the same office images from LinkedIn every minute. That exact problem transfers to the model: even if a dataset “is supposed to be diverse,” someone has to manually (or algorithmically) ensure that — and the algorithm itself has biases, since it’s based on prior data.

As a result, AI knows extremely well what dominates the Internet (cats, vacations, offices, stock aesthetics), but struggles more with what is rare, niche, or truly new.

When AI Feeds on AI: Nature and “AI Model Collapse”

The most alarming topic isn’t about images, but data itself.

Nature described a phenomenon explicitly termed AI model collapse. The idea seemed tempting: since many domains lack data (e.g., medicine, law — especially due to legal constraints in Europe), maybe it’s possible to “fill in” synthetically, generating data with AI and training further models on these synthetics. [source].

The problem: such a model begins to forget edge cases. These aren’t just generated less often—they can statistically disappear from the distribution.

For graphics, this means fewer “weird” ideas. In medicine, it can be catastrophic: a disease occurs once in 100,000 or once in a million cases. If the model “loses it,” it may distort the clinical picture and push a doctor towards a wrong decision.

A human, moving through the world, absorbs a lot of seemingly irrelevant entropy: accidental street scenes, unusual events, details. AI trained on synthetics loses this entropy systematically.

Arno Stern and “La Trace de la Mémoire Organique”: Why People Do Not Start From Zero

Arno Stern—a researcher and practitioner of children's development through drawing, founder of a Paris studio, and author of “La Trace de la Mémoire Organique”.

Stern noticed something striking: children around the world—even in places without television, the internet, or exposure to mass culture—draw “a house” in a similar way: four walls, a pitched roof, often a chimney, even where there’s never been snow and chimneys are not part of their experience.

This leads to an important difference: - humans have certain innate or deeply rooted concepts and are immersed in an environment, - AI “learns from absolute zero” and is bound by how data is presented.

ARC 3 and the “Intelligence Gap”: Where Humans Still Outclass Models

In all this confusion, it’s easy to mistake “doing tasks” for intelligence. A cold shower is provided by the ARC 3 benchmark. [source].

It’s a set of games/tasks where you do not know the rules at the start. You have to discover them through observation as you go. This tests the essence of intelligence: adaptation to the unknown.

Results: - in ARC 3 Grok: 0%, - best models: around 0.6–0.7%.

The ARC chart shows a yawning gap between humans and models—the “intelligence gap”. It’s a reminder that humans are not “task machines.” Humans act contextually, iteratively, in an environment where rules often emerge as you go.

Recommendation Algorithms: Creativity Needs Discomfort, but the Feed Removes It

Artificial intelligence is not just generating texts and images. It’s also recommendation systems on YouTube, Instagram, and other platforms.

Their goal is simple: keep you on the platform. Not: help you grow, broaden your perspective, or increase your creativity. Hence mechanisms like:

you spend 93% of your time on one narrative and 7% on another,
the system assumes you “prefer” the first one,
you get served even more of it.

The side effect is polarization and the breakdown of shared debate. Worse yet, even a “valuable feed” changes little, because there’s the further problem of memory and attention: viewing 100–200 posts a day doesn’t mean those contents stick. A good test: after two days, try to list what you saw—being able to recall 1% is already a great result.

Creativity needs otherness, which by definition is slightly uncomfortable. Platforms smooth out that discomfort, since discomfort shortens session length.

Conclusions

Generative AI has attractors: with longer loops and weaker context signals, it sinks into stereotypes (in the study: 12 topics, e.g., cathedrals and offices).
Synthetic data can hobble a model by removing edge cases.
Humans win by adapting in unknown environments (ARC 3).
Algorithms feed bubbles, as they optimize for platform time, not development—and creativity declines without diversity.
The key competency is becoming metacognition: awareness of one’s own thought processes and reactions to “pretty, quick” AI responses.

How to Apply This (Starting Now)

Below are steps you can take without major revolutions—but that consistently build “anti-average” habits.

Add Entropy
Once a week, visit a public library and choose a book from a random shelf (not from the “new releases” as in a bookstore).
For visual creativity: look at images (painting, photography, posters), not just “AI art”.
For linguistic creativity: read poetry from different cultures (Japanese, Polish, Ukrainian, British).
Design Discomfort Instead of Removing It
Deliberately consume content outside your bubble: politically, professionally, aesthetically.
Set a rule: spend 10–20% of your time on “not mine” topics, even if it's slightly irritating.
Work With AI as a Partner, Not as a Shortcut
Before accepting a result, ask yourself 3 questions:
- Does this realize my vision, or just look nice?
- What has been smoothed out/lost? (people, details, edge cases)
- Which outlier do I deliberately want to recover?
Enforce variance in your prompts: “give me 5 versions, including 2 extremely unconventional ones.”
Do a Quick “Memory Test” on Feed Content
After 48 hours, write down what you remember from Instagram/YouTube.
If almost nothing—that’s a sign the feed is entertainment, not fuel for thinking. Then either reduce the dose or change how you use it.
Build a Micro–Thinking Week
Once a week, do 2 hours offline and phone-free: just a notebook/paper and one topic.
Once a quarter, spend half a day “offline” with books or a long text (no link hopping).
If You Work with Data (Business/Medicine/Law): Don’t Trust Synthetics Blindly
Treat synthetic data as an addition, not a foundation.
Monitor metrics on rare classes and edge cases—because those disappear first.

This is a real defense against a world that increasingly tries to convince you that “average” is the same as “meaning.” Creativity doesn’t need more “how-to”s. What it needs is diverse input, discomfort, and conscious control over what shapes your thinking.

Share on Share on Share on Share on