A few days ago I scrolled past four or five AI-generated landing pages on X: purple gradient base, glassmorphism cards, one big headline, three testimonials underneath. All made fast — one prompt, probably. The problem is they looked identical. I genuinely couldn't tell whether different people had made them at all.
Generating something "not ugly" costs nothing in 2026. The consequence is bigger than most people think: the scarce asset moved.
1. The model can only give you the average
A large language model is, at its core, predicting the "mean" of its training data. Ask it for "a nice product page" and it hands you the greatest common divisor of every product page on the internet: gradients, rounded corners, glassmorphism, a centered headline. None of that is ugly — it's not ugly. And between "not ugly" and "good" sits the most expensive thing in this era.
It used to be that making things was expensive: knowing how to code, slice layouts, deploy — that itself was the moat. Now the price of implementation has been driven through the floor; it's one prompt away. The bottleneck moved from "can't build it" to "can't tell which one is good."
Judgment can't be outsourced. The model can hand you ten options overnight, but the second where you go "this is the one" can only happen inside your own head.
2. Taste isn't a gift, it's a diet
I grew up soaked in Macs, and the software I've used for years is the "every pixel has a reason" kind — Things, Raycast, Surge. After enough exposure, you build a reference frame in your hands: whether a spacing is right, whether an animation flows, one glance and you know.
So there's nothing mystical about taste. It's a massive amount of input plus a bit of fastidiousness. It's slow — the model can binge every landing page humanity has ever shipped in one night; my eye took twenty years to get here. But that's exactly why it can't be mass-produced.
This very blog is the example. I spent an entire evening tuning the typography — line height, letter spacing, how the TOC indents — one revision at a time. The model can produce a "not ugly" version in a second. It cannot produce the "right" version, because "right" lives in me, not in the training set.
3. Where slop comes from
I'm more and more convinced that AI slop is not the model's fault — it's the human's. A person without judgment, given free generation, produces "not ugly" at scale. The output belongs to the model, the taste belongs to the person, and without the second half the first one amplifies mediocrity ten-thousandfold.
A brutal self-test: when the model hands you something, can you say what to delete and why? The model adds everything — adding is its instinct. The human does the subtracting, and subtracting is the taste.
4. Which assets are appreciating
"Knows framework X" — depreciating; the model knows it better than you. "Can build system Y" — depreciating; anyone can prompt their way to one.
"Knows what good is" — appreciating. It decides whether your agent produces work or slop, and whether your name is still worth anything in ten years.
Generation keeps getting cheaper. Taste doesn't.