Why emphasize human review in a world where AI will do everything?

The model writing your CSS this week is the worst it will ever be.

By 2027 it will outperform most front-end engineers on most tasks.

By 2029 the line between “a senior front-end” and “a model with the right context window” will be hard to parse. The models are good, real good. None of that is going to slow down.

So why put a human in the loop?

How we think about thinking.

By the early 20th century, philosophy had essentially split into two ways of thinking about thinking.

The analytic tradition, mostly Anglo-American, treats the mind as a system of rules, inputs, and predicate logic. Most of what we call “AI” in 2026 is the empirical apex of that tradition. AI can predict the next token. It is not deterministic, like an IF/THEN statement, but probabilistic, and well enough. Eventually you have something that behaves like reasoning.

But is reasoning the path to UI design? Sometimes. More often, design is feeling.

OK, so — the other philosophy tradition is called “continental.” Mostly French and German, and a bit more like sitting at a café and ruminating on what it subjectively feels like to be alive, to be human, to be anything at all.

This type of philosophy is much more popular at dinner parties. Continental philosophy is less interested in what the mind does and more interested in what it is like to be one. The phenomenology of subjective experience. The fact that there is something it is like to be you, sitting there, reading this sentence.

It feels like something to see design that has been done well and serves its purpose. Probabilistic rules that line up CSS and Tailwind according to a corpus of really, really good CSS and Tailwind capture most of what is needed to not have broken design.

AI is eating the entire analytic stack of UI design and will likely eliminate “analytic-style” work entirely.

But the LLM still does not know what it is like to feel like it is seeing something that should be changed. There is no evidence that current technology will solve this part in the next few years.

Nobody knows how to build that. The hard problem of consciousness is hard because it is hard. Even GPT-7 in 2029, even an omniscient model that has eaten every CSS spec back to 1996, that can create photorealistic stock photography from scratch, and writes shadcn in its sleep, probably does not know what it feels like to look at a page and know that something is just off.

Taste is a human problem.

Taste is a constantly shifting target. It is a feeling that comes from a body, and from the bodies of your users.

This feels off. This feels right. This makes me trust the company. This makes me close the tab.

You can train a model on a billion screenshots and a billion tags and it will not have the involuntary flinch of a person who has stared at one too many SaaS landing pages. That flinch is what we sell.

What that means for uiopsy.

We started uiopsy because the part of UI work that is about producing pixels is now solved or close to it. The part that is about reading them is not solved, and possibly never will be.

The agent does the part agents are great at. We do the part that requires being a person.

Chances are, you will ship a page through Claude Code or Cursor or whatever you have got next year, and that is a fine and increasingly excellent thing.

We provide the small, specific, embodied thing the agent does not have. The final 2% of human taste.