A polished screen is not a solved problem
7 May 2026·5 min·Essay
I keep seeing the same misunderstanding show up in slightly nicer clothes.
A new tool appears. It can generate an interface from a prompt. It can turn text into screens. It can move from idea to code before the designer has even finished explaining what they meant. The demo looks clean. The motion is smooth. The landing page says something about collapsing the distance between imagination and product.
Fine. Impressive. Also, not the point.
A polished screen is not a solved problem.
That sentence sounds defensive if you read it the wrong way, like designers trying to protect a professional border. I do not think that is the interesting version. The better version is more basic: design is not the production of a thing. Design is the work of understanding what kind of thing should exist, under what conditions, for which people, with which constraints, and with what tradeoffs hidden inside it.
The interface is evidence. It is not the whole act.
Karri Saarinen wrote this clearly in Linear's essay, and the part that stuck with me is Christopher Alexander's idea of fit. A good form is not good because it looks good in isolation. It is good because it fits its context. The context is not a mood board. It is the ugly complete environment around the problem: user habits, technical constraints, edge cases, business pressure, existing workflows, social friction, timing, permissions, bad data, weird exceptions, and the thousand tiny forces nobody puts in the prompt because nobody has noticed them yet.
This is why AI-generated product work can feel so strange. The output arrives fast, but the context did not get less complicated. The model can skip to a plausible form before anyone has sat with the problem long enough to deserve a form.
That is where the brittleness comes from.
You can feel it in products that look finished in a screenshot and start falling apart the moment you use them. Everything has a card. Everything has a gradient. Everything has a confident empty state. But the flow does not know why it exists. The hierarchy is not making a choice. The copy sounds right without committing to meaning. The interaction handles the happy path and then quietly panics when reality walks in wearing shoes.
This is not a visual polish problem. It is a fit problem.
The dangerous part of AI tools is not that they make bad work. Bad work existed before AI. We had entire eras of dashboard sludge and SaaS confetti without needing a model to blame. The dangerous part is that AI makes unresolved work look resolved earlier.
That is new.
Before, if the thinking was weak, the output often looked weak too. The wireframe was awkward. The deck was unfinished. The prototype had friction. The mess was visible, so people knew they were still inside the work.
Now the mess can get dressed up very quickly. The thing looks done before the team has earned clarity. That changes the social dynamics of design. A polished mockup creates pressure. People react to what is on the screen. Stakeholders start choosing between options. Engineers start estimating. Someone says "this is close" because the surface looks close.
But close to what?
That is the question AI does not automatically answer.
I still like working visually because the slowness is part of the thinking. Moving a block five times is annoying, but the annoyance teaches you something. You notice the relationship between two pieces of information. You realize the button is asking the user to make a decision too early. You see that the empty state is not empty, it is actually a moment of anxiety. You find out that the page is trying to do two jobs because the product strategy is avoiding a decision.
Those discoveries do not happen after the design process. They are the design process.
Writing works the same way. A paragraph is not just a container for a thought you already had. Half the time, the thought becomes real because the sentence fails three times and forces you to admit what you actually mean. The friction is not waste. The friction is where the shape appears.
This is why "AI as designer" feels off to me. AI can produce options. It can prototype. It can make a wall of variations. It can help you see directions you might not have drawn yourself. That is useful. I want that. I use that.
But option generation is not the same as judgment.
The designer's job is not to make more possible forms. The world already has too many possible forms. The job is to reduce possibility into a thing that fits. That reduction is not mechanical. It involves taste, yes, but also responsibility. You have to decide what the product is refusing to do. You have to decide what pain is acceptable. You have to decide which user deserves priority when two real needs collide. You have to know when the beautiful version is lying.
Most AI tools are still much better at abundance than refusal.
They give you more screens. More variants. More copy. More ideas. More animations. More paths. They are engines of more.
Design is often an act of less.
Not minimalism as an aesthetic. I mean less as discipline. Fewer promises. Fewer concepts. Fewer decisions pushed onto the user. Fewer places for the product to pretend it is smarter than it is. The hard part is not filling the canvas. The hard part is knowing what should survive.
That is why the future of design with AI should not be framed as "prompt to product." That phrase is cute, but it is spiritually lazy. It makes design sound like translation: I have an idea, the machine produces the artifact. The real work is not translation. It is investigation.
The better question is: can AI help us understand context faster without replacing the part where a human has to care?
Can it surface edge cases we missed?
Can it simulate how different users might interpret the same flow?
Can it help map constraints, contradictions, dependencies, and failure modes before we fall in love with a screen?
Can it act less like a vending machine for polished output and more like a thinking partner that keeps pulling the problem back into the room?
That version is interesting to me.
Because I do not think designers need to protect the old process just because it is old. A lot of design process is theater. A lot of workshops are expensive ways to produce obvious sticky notes. A lot of critique is taste politics with better vocabulary. If AI kills those parts, good. They had it coming.
But if we replace fake process with fake output, we did not progress. We just made the costume render faster.
The real standard should be fit.
Does the thing understand the context it lives in? Does it resolve the forces around the problem, or does it hide them behind polish? Does it make the user's next action clearer? Does it make the system more honest? Does it know what it is not doing?
If the answer is no, then the output is just output.
Maybe beautiful output. Maybe impressive output. Maybe investor-demo output.
Still not design.
Design begins when the form starts answering to the world around it.